October 2026 Summaries
147 posts from Qovery
Filter
Month:
Year:
Post Summaries
Back to Blog
Vercel’s Daemons and Drives, introduced in October 2026, address two major limitations of serverless AI-agent workloads by enabling long-running container processes and persistent storage for Sandbox sessions, making Vercel more suitable for tasks such as queue workers, coding agents, and sustained tool loops. However, the comparison argues that Kubernetes or other infrastructure in a company’s own cloud account remains advantageous when workloads require strict data residency or compliance controls, GPU or high-memory hardware, direct private-network access, or cloud-provider discounts and cost control at scale. It contrasts Vercel with serverless container platforms, specialist sandbox providers, and self-hosted Kubernetes, presenting convenience, operational burden, portability, hardware support, and network access as the main tradeoffs. The author suggests that small teams with lightweight, bursty, Next.js-centered workloads may find Vercel sufficient, while organizations operating data-sensitive, GPU-intensive, or high-volume agent systems may prefer a hybrid arrangement with Vercel for frontend and orchestration work and Kubernetes in their own cloud for backend agents. The piece also promotes Qovery as a platform intended to provide Vercel-like deployment workflows, preview environments, and managed Kubernetes operations while retaining customer ownership of cloud infrastructure and billing.
Oct 31, 2026
2,674 words in the original blog post.
Migrating a Firebase backend to Kubernetes in a company-controlled cloud account is presented as a phased process driven by cost predictability, data residency, and customer ownership requirements, with the recommended order being an API layer, Cloud Functions, Firestore, and Firebase Authentication last. The approach uses an API layer to stop clients from directly accessing Firestore, converts HTTP functions into containerized routes, event triggers into queue consumers, and scheduled functions into Kubernetes CronJobs, while retaining Firebase Cloud Messaging where appropriate. Firestore requires the most redesign because no direct replacement reproduces its client SDK, security rules, real-time listeners, or offline persistence; PostgreSQL with JSONB or MongoDB are suggested alternatives, supported by dual writes, backfills, shadow reads, and custom WebSocket or SSE real-time features. Firebase Auth users can generally be moved without password resets by exporting Firebase’s modified scrypt hashes to compatible providers such as Ory or SuperTokens, or through just-in-time migration for providers such as Cognito or Zitadel, while supporting both token issuers during transition. The migration replaces Firebase’s variable per-operation pricing with a fixed Kubernetes infrastructure baseline plus database, networking, and operational staffing costs, making residency and control as important as direct cost savings. The piece also promotes Qovery as a platform for preserving Firebase-like deployment workflows, preview environments, and managed Kubernetes operations in an organization’s own AWS, GCP, Azure, or Scaleway account.
Oct 31, 2026
3,287 words in the original blog post.
Kubernetes Kueue is presented as a Kubernetes-native admission controller for sharing scarce GPU capacity across teams by queueing workloads until their full CPU, memory, and GPU requirements can be reserved, then allowing the standard kube-scheduler to place their pods. It addresses idle-but-reserved GPU pools, fragmented distributed-training jobs, and the lack of borrowing in standard ResourceQuota through ResourceFlavors for hardware types, ClusterQueues for team quotas, LocalQueues for namespace entry points, and cohorts that enable controlled borrowing, lending, and preemption. Teams receive guaranteed nominal GPU allocations while being able to use unused capacity from others, with owners able to reclaim their quota when needed; priorities, fair-sharing weights, checkpointing, topology-aware scheduling, and spot-capacity flavors can further shape behavior. Kueue does not partition physical GPUs, provision nodes, replace Kubernetes scheduling, or operate cluster infrastructure, so technologies such as MIG, MPS, time-slicing, autoscalers, GPU operators, and platform tooling remain separate concerns. The guide recommends measuring real GPU utilization with DCGM alongside Kueue queue depth and admission wait times before purchasing additional hardware, arguing that low utilization combined with long waits indicates a scheduling and governance issue, whereas sustained high utilization and long waits after tuning quotas may justify more capacity.
Oct 31, 2026
4,950 words in the original blog post.
ISO/IEC 42001:2023 certifies an organization’s AI management system rather than any individual model, requiring documented governance, roles, risk and impact assessments, lifecycle controls, internal audits, management reviews, and continual improvement within a clearly defined scope. For self-hosted AI, auditors focus heavily on infrastructure evidence demonstrating where models and data run, who has access, how each deployment is approved and linked to source commits, what is logged and retained, and whether production configurations match approved records. Organizations must address all 38 Annex A controls in a Statement of Applicability, with particular attention to AI-specific risk assessments and impact assessments covering effects on individuals, groups, and society. Certification is conducted by an accredited body through Stage 1 and Stage 2 audits, followed by annual surveillance over a three-year certificate period, making continuously generated records more reliable than manually collected screenshots. Existing ISO 27001 processes can reduce preparation effort, while managed AI providers can be treated as suppliers, though self-hosting gives organizations more direct control over data residency, logs, and operational evidence. The article presents Qovery as a platform that can support the infrastructure side of this evidence by deploying AI workloads in an organization’s own cloud account with access controls and commit-linked deployment records, while emphasizing that it does not provide certification or replace required governance documentation.
Oct 31, 2026
4,292 words in the original blog post.
Kubernetes multi-tenancy ranges from soft isolation through shared namespaces to fully isolated clusters, with each model balancing security, flexibility, cost, and operational effort differently. Namespaces, enhanced with RBAC, quotas, NetworkPolicies, and Pod Security Admission, suit trusted internal teams and preview environments but share the API server, cluster-scoped resources, nodes, and kernel, making them unsuitable for untrusted customer workloads. Capsule and the now-retired Hierarchical Namespace Controller improve namespace self-service and policy management without changing those underlying limitations. vCluster gives tenants separate virtual control planes, CRDs, RBAC roots, and Kubernetes versions while normally retaining shared worker nodes, whereas Kamaji hosts real tenant control planes on a management cluster and is better suited to fleets with dedicated worker nodes. For untrusted code, dedicated node pools combined with gVisor or Kata Containers can strengthen runtime isolation, while one cluster per tenant provides the strongest separation of control plane, nodes, networks, and failure domains but incurs roughly $73 monthly managed-control-plane fees per cluster before node, observability, ingress, and upgrade costs. The recommended approach is often hybrid: use hardened namespaces for internal workloads, vClusters for tenants requiring administrative autonomy or custom operators, and dedicated clusters for regulated, region-specific, or contractually isolated customers, with the chosen platform tooling responsible for automating the resulting lifecycle and operations.
Oct 31, 2026
4,204 words in the original blog post.
Following several reported developer-platform security incidents in October 2026, including breaches involving Railway and Northflank, the piece argues that shared PaaS models concentrate risk because customer builds, applications, management systems, and secrets may reside on common infrastructure, making stored environment variables and valuable AI keys attractive targets. It presents Qovery as an alternative that retains PaaS-style features such as git-based deployments, preview environments, databases, autoscaling, and logs while deploying applications and data into customers’ own AWS, GCP, Azure, Scaleway, or Kubernetes environments. Qovery’s control plane remains responsible for deployment management, meaning it can still access deployment paths and may store secrets entered through its console, while its default image builds use shared but isolated, disposable build nodes unless customers choose to build in their own clusters. The company recommends customer-managed secret stores and short-lived cloud access roles to reduce exposure, cites SOC 2 Type II certification, annual penetration tests, and support for regulatory requirements, and is developing a dedicated customer-hosted control plane for organizations requiring complete isolation.
Oct 31, 2026
3,201 words in the original blog post.
Northflank disclosed a major security incident on October 30, 2026, advising all users to assume environment variables may have been decrypted and to rotate all secrets after reports of unauthorized AI API key use; as of early October 31, the root cause, affected customers, and impact on BYOC deployments had not been publicly confirmed. The recommended response is to prioritize rotating AI, payment, cloud, and database credentials, review provider billing and usage logs, inspect newly enabled audit logs for unauthorized changes, and restrict replacement keys with quotas and regional or service limits. The incident is presented alongside Wiz research identifying nine AI developer-platform incidents during October, with attackers frequently targeting stored environment variables and AI credentials, though no single vulnerability has linked the cases. The author argues that shared deployment platforms concentrate valuable customer secrets and contrasts Qovery’s model, in which applications run in customers’ own cloud accounts and secret-manager references can keep values outside Qovery’s control plane, while acknowledging that secrets entered directly into Qovery’s console remain encrypted but could be exposed in a sufficiently deep breach.
Oct 30, 2026
3,444 words in the original blog post.
Italy’s ACN cloud-qualification regime requires both the underlying cloud infrastructure and a SaaS product itself to appear in the ACN catalogue before Italian public administrations can procure the service. Public data is classified as ordinary, critical, or strategic, requiring minimum qualification levels 1, 2, and 3 respectively; infrastructure must generally be located in the EU, while strategic workloads face additional controls over extra-EU access. As of 30 October 2026, AWS, Microsoft Azure, Google Cloud, Oracle, and several Italian or European providers hold level-2 listings suitable for ordinary and critical workloads, while Aruba has level-3 infrastructure services and the state-backed Polo Strategico Nazionale holds level 4 for strategic needs. Qualification is specific to a named service, level, and region rather than a provider broadly, and certifications such as GDPR compliance, ISO 27001, or SOC 2 do not substitute for ACN registration. Vendors must also document data residency, backup locations, encryption, access controls, deployment histories, continuity objectives, subprocessors, and exit plans, with overlooked cross-region backups, foreign observability tools, CI runners, and support access posing common audit risks. The article recommends maintaining one application codebase while parameterizing cloud accounts, regions, and Kubernetes clusters, and presents bring-your-own-cloud or bring-your-own-Kubernetes deployments as a way to retain contractual, regional, and audit control; it notes that Qovery can support these deployments but is not itself ACN-qualified.
Oct 30, 2026
2,906 words in the original blog post.
Following Railway’s reported October 21, 2026 security incident, which allegedly involved unauthorized access to some internal systems and potential exposure of environment variables, the post advises customers to separate immediate credential rotation from a planned infrastructure migration. It recommends rotating database credentials, signing keys, API and OAuth secrets, webhook tokens, Git and registry credentials, and cloud IAM keys using a create-deploy-verify-revoke sequence, while storing replacements in a cloud-controlled secret manager. The proposed migration approach moves existing containerized Railway workloads to Kubernetes or managed cloud services without rewriting application code, focusing instead on configuration, databases, private networking, persistent storage, egress allowlists, and DNS/TLS. It outlines a roughly 12-day parallel migration involving Dockerfiles, staging validation, PostgreSQL logical replication, DNS cutover with short TTLs, and a 72-hour rollback period. The post compares managed PaaS, self-managed cloud infrastructure, managed Kubernetes, and bring-your-own-cloud internal developer platforms, arguing that account ownership improves control over secrets, billing, and recovery options, while acknowledging the operational responsibilities and potentially comparable costs of operating infrastructure outside a PaaS. It also promotes Qovery as a platform intended to provide Railway-like deployment workflows in customer-owned AWS, GCP, Azure, Scaleway, or Kubernetes environments.
Oct 30, 2026
5,357 words in the original blog post.
A replay of a Qovery and TPC discussion examines how AI coding agents are increasing infrastructure demands by producing code, pull requests, builds, tests, and deployments at machine speed. It argues that ongoing “Day Two” operations—including maintenance, upgrades, cost management, and compliance—consume roughly 80% of infrastructure effort, while static shared staging environments become impractical when many agents work in parallel. The proposed approach is to create fast, isolated ephemeral environments for each branch or pull request, though this requires automated configuration of networking, variables, certificates, logging, caching, and multiple application services. The session also contends that agents can shorten infrastructure migrations by automating discovery, mapping, validation, and testing, while shifting DevOps, SRE, and platform engineers from direct operators to supervisors of automated systems. A deployment demonstration shows an agent provisioning an application and test environments through Qovery, with human confirmation required before infrastructure changes. To limit risks, the discussion recommends least-privilege policy tokens, policy-as-code testing, and audit logs that record both permitted and denied agent actions, while noting that future agent-specific identities could further distinguish automated activity from human activity.
Oct 30, 2026
2,932 words in the original blog post.
TISAX assesses an automotive supplier’s information security practices rather than certifying its chosen cloud provider, so AWS, Azure, Google Cloud, and other providers’ own TISAX, ISO 27001, or C5 attestations cover only their portion of shared responsibility. Suppliers must obtain and share their own TISAX labels through the ENX exchange, with assessment objectives determining whether AL2 or the more comprehensive, on-site AL3 assessment is required; prototype data, strictly confidential information, special data, and very high availability generally require AL3. Cloud-based evidence must demonstrate data classification, documented processing locations, least-privilege IAM and access reviews, encryption and key custody, separation of prototype, production, and non-production environments, retained logs, tested restores, patching, change management, incident processes, and supplier due diligence, including GDPR processor obligations where personal data is involved. The article argues that providers remain responsible for physical infrastructure while customers retain responsibility for configurations, data, workload security, and audit evidence, including managed Kubernetes operations. It also contrasts SaaS platforms, bring-your-own-cloud platforms, and self-managed Kubernetes, contending that retaining cloud accounts, keys, data, logs, and regional choices in the supplier’s name can simplify control and evidence collection, while emphasizing that no platform itself makes an organization TISAX compliant.
Oct 30, 2026
3,770 words in the original blog post.
DeepSeek V4.1-Flash-1017 is an MIT-licensed, open-weights multimodal mixture-of-experts model with a 1 million-token context window, allowing commercial self-hosting, fine-tuning, redistribution, and retention of prompts and data within a customer’s cloud environment. Although it activates only 16 billion parameters per token during decoding, its approximately 511 GB mixed-precision checkpoint requires GPU sizing based on roughly 763 billion total parameters; with memory headroom, native deployment needs about 614 GB of VRAM, while offloading its Engram memory tables to host RAM permits a replica on four H200 GPUs. The recommended production configuration is an eight-H200 node, which can run either the full model or two four-GPU replicas, while KV-cache requirements remain relatively low even at long contexts due to the model’s cache-efficient architecture. vLLM and SGLang are identified as the primary serving options, with vLLM offering released support and SGLang potentially benefiting workloads with shared long prompts. Major operational concerns include cloud GPU quotas, interconnect bandwidth, model-loading time, host RAM, driver compatibility, persistent weight storage, and autoscaling based on queue depth. The analysis concludes that self-hosting is generally motivated by compliance, data residency, latency control, and customization rather than cost, since an AWS eight-H200 node costing about $63 per hour would need to sustain roughly 14,650 output tokens per second continuously to match DeepSeek’s hosted API output-token pricing; hybrid deployments that self-host sensitive steady traffic and burst to APIs are presented as a common alternative.
Oct 30, 2026
3,669 words in the original blog post.
Multi-cloud Kubernetes observability across AWS, GCP, Azure, and other providers is presented as a three-layer approach rather than a single product: telemetry through OpenTelemetry and Prometheus-compatible systems, workload-level cost allocation through OpenCost or Kubecost combined with cloud billing exports, and a deployment/control layer that records what changed and enables actions such as rollback or environment management. The approach relies on consistently applying five workload labels—application name, version, environment, team, and cost center—and propagating equivalent cloud tags so metrics, costs, billing data, and deployment events can be grouped consistently across clusters. The discussion argues that teams often already have monitoring but lack deployment context and complete shared-cost attribution, while cloud cost consoles alone cannot allocate shared Kubernetes resources by namespace or team. It compares self-hosted, managed observability, and platform-control options in terms of engineering effort, capabilities, and pricing, positioning Qovery as a deployment and control layer that complements rather than replaces fleet-wide telemetry or cost tools. A proposed 30-day rollout starts with label enforcement and billing tags, then consolidates telemetry, deploys cost allocation tooling, and finally integrates deployment events and operational controls.
Oct 29, 2026
4,456 words in the original blog post.
Switzerland’s revised Federal Act on Data Protection, effective since September 2023, does not prohibit Swiss organizations from using US-owned cloud providers or require data to be hosted in Switzerland, but it requires lawful international-transfer mechanisms, risk-appropriate security controls, and documented accountability. Hosting in a Zurich cloud region establishes data residency but does not eliminate potential exposure to US legal orders under the CLOUD Act, which can apply where a US provider controls the data regardless of its physical location. The discussion distinguishes residency, legal sovereignty, and technical access control, arguing that encryption-key custody, tightly controlled human access, auditability, and complete subprocessor oversight are more meaningful safeguards than provider nationality alone. Transfers to US providers may rely on Swiss-US Data Privacy Framework certification or Swiss-adapted standard contractual clauses with transfer impact assessments, while banks, healthcare organizations, and other regulated sectors may face stricter secrecy and outsourcing obligations. It recommends a tiered infrastructure model in which ordinary data can use hyperscalers with customer-held or external keys, while secrecy-bound information may warrant Swiss-operated or self-hosted environments, supported by evidence such as processing records, DPIAs, access-control matrices, key policies, audit logs, vendor contracts, and breach-response procedures.
Oct 29, 2026
4,598 words in the original blog post.
Pending GPU quota increases can delay training because approvals depend on regional physical capacity, account history, and provider-specific quota rules, so teams are advised to first try another region, an older GPU generation, reservation-style capacity, or Spot instances with checkpointing before bursting to another cloud. The proposed same-day cross-cloud approach relies on a portable container image with pinned CUDA, cuDNN, and NCCL versions, S3-compatible storage for data and checkpoints, and Kubernetes with the NVIDIA GPU Operator, while cloud-specific settings such as endpoints, credentials, and GPU resource requests remain configurable. The outlined process is to verify available capacity, create or access a GPU Kubernetes cluster, ensure the image is accessible, copy data once, redeploy the job with limited configuration changes, and resume from the latest checkpoint. It distinguishes GPU-specialist providers, cross-cloud schedulers such as SkyPilot and Ray, managed training services, and deployment platforms such as Qovery, noting that Qovery helps transfer deployment configuration and environments between clusters but does not provide GPU capacity or distributed-training scheduling. The discussion also emphasizes that data egress, duplicate storage, and idle GPU nodes can be more consequential than hourly GPU prices, and recommends maintaining pre-approved backup capacity, documented data-transfer procedures, and periodic recovery drills to prevent future quota-related disruptions.
Oct 29, 2026
4,837 words in the original blog post.
Amazon EKS clusters running Kubernetes 1.34 are scheduled to leave standard support on December 2, 2026, after which AWS automatically enrolls them in Extended Support by default and raises the control-plane charge from $0.10 to $0.60 per hour, or roughly $438 monthly per cluster. The post advises organizations to inventory clusters across every AWS account and enabled region, rather than relying on a single console or account, and to cross-check findings through EKS insights, Cost Explorer, and AWS Health notifications. Because EKS upgrades require sequential minor-version hops, it recommends a pre-upgrade review of deprecated APIs, node versions, managed add-ons, third-party controller compatibility, PodDisruptionBudgets, available subnet IP addresses, backups, and nonproduction testing before upgrading the control plane, nodes, and add-ons in each step. It notes that teams may instead set the support policy to STANDARD before the deadline, causing AWS to auto-upgrade rather than charge for Extended Support, while emphasizing that version ownership, recurring lifecycle planning, and fleet-wide monitoring are needed to prevent unnoticed costs. The comparison with GKE and AKS characterizes paid extended support as an industry-wide pattern, although their enrollment models differ, and the post also promotes Qovery as a platform for managing Kubernetes upgrades across cloud environments.
Oct 29, 2026
4,729 words in the original blog post.
Qovery has introduced a preview Bring Your Own Kubernetes (BYOK) approach for existing EKS, GKE, and Kapsule clusters through its Qovery Operator, which is installed once using Helm and lets users manage Qovery component deployments and updates from the Qovery Console. The operator addresses limitations of the prior CLI-based installation process by enabling console-managed updates, configurable platform profiles, visibility into installed components, and preflight checks for issues such as connectivity, permissions, registry access, conflicting Helm releases, and storage configuration. Qovery emphasizes that customers retain control of their clusters because it does not store kubeconfig files, the operator communicates outbound only, and deployment jobs run inside the cluster, although these jobs use cluster-admin permissions. The service installs components including logging tools, Envoy Gateway, cert-manager, ExternalDNS, and Qovery agents, while supporting coexistence with existing StorageClasses but not yet adopting existing cert-manager, ExternalDNS, or Envoy Gateway installations. ArgoCD users can continue deploying through GitOps while viewing application status, revisions, and logs in Qovery. The preview is limited to Enterprise-plan organizations, requires an eligible cluster and cluster-admin Helm access, uses Qovery-provided domains, lacks certain monitoring features, and currently requires manual component removal.
Oct 28, 2026
1,531 words in the original blog post.
Self-hosting Google’s Gemma models can become cheaper than hosted APIs only when GPU utilization is consistently high, while hosted services remain more economical for low, spiky, or unpredictable traffic because they charge per token and scale to zero. The guide compares seven deployment approaches—Runpod, Vast.ai, Lambda and other GPU-native clouds, Modal, Hugging Face Inference Endpoints, direct hyperscaler instances, and Qovery-managed deployments in a customer’s own cloud account—according to price, operational effort, scaling, compliance, and data-residency needs. It recommends smaller Gemma variants or quantized 12B models for 24GB GPUs, while larger 27B-class models generally require 48GB to 80GB GPUs depending on precision and context requirements, with KV cache becoming a major memory factor at long contexts and high concurrency. For production serving, vLLM and SGLang are presented as preferable to Ollama and llama.cpp because they support efficient concurrent batching, while direct cloud deployments provide maximum control but require teams to manage drivers, Kubernetes, scaling, monitoring, and upgrades. The article also notes that self-hosting may be justified independently of cost for privacy, compliance, fine-tuning, custom adapters, and predictable latency, and positions managed platforms as options for reducing infrastructure work.
Oct 28, 2026
3,458 words in the original blog post.
Moving a containerized GPU inference workload from specialized providers such as Lambda Labs or CoreWeave to AWS, GCP, Azure, or Scaleway is usually an infrastructure change rather than a code rewrite, provided the image, CUDA stack, model artifacts, serving runtime, and API contract remain fixed. The main migration concerns are securing GPU quota and capacity early, selecting equivalent GPU hardware based on model and VRAM needs, ensuring compatible host drivers and Kubernetes NVIDIA device plugins, moving model weights to object storage, configuring startup probes for lengthy model loads, and preserving autoscaling and private networking behavior. The recommended process is to request quota first, validate the container and GPU node pool, deploy the same image digest, benchmark latency, throughput, utilization, and cost alongside the existing environment, then shift traffic gradually with a rollback period. Although specialized GPU clouds often offer lower raw hourly prices and faster access to new hardware, operating in a company’s own cloud account can improve total economics through committed-use discounts, reduced data-transfer costs, higher utilization, compliance control, and VPC-private data access. The piece also presents Qovery as a platform for managing Kubernetes-based deployments in customer-owned cloud accounts while retaining control of cloud billing and infrastructure choices.
Oct 28, 2026
5,496 words in the original blog post.
Seven EU-incorporated providers—Scaleway, OVHcloud, IONOS Cloud, StackIT, Open Telekom Cloud, UpCloud, and Nebius—offer GPU compute for AI training or inference in EU data centres, although their capacity, GPU types, managed AI services, corporate structures, and prices differ. The discussion argues that genuine EU AI residency requires more than storing data in an EU region: training and inference execution, management control planes, telemetry, support access, encryption keys, and third-party API calls must also be assessed, since remote support and US-controlled operators may create cross-border transfer concerns. It distinguishes EU operators from Hetzner’s dedicated-server-only offering, Swiss-operated Exoscale, and Norway-based Nscale, while noting that US hyperscalers’ European regions provide geographic EU compute but may require sovereign or partner-operated options such as AWS European Sovereign Cloud, Delos Cloud, T-Systems Sovereign Cloud, or S3NS for stricter requirements. Recommended verification includes obtaining a sub-processor list, contractual EU region pinning, key-management details, policies on non-EU support access, and a written commitment that customer data will not train models, supported by technical controls and audit logs. Published pricing cited for late October 2026 suggests several EU providers offer lower H100 and egress costs than AWS or Azure, though hyperscalers may offset this with more mature managed services; a hybrid approach can use EU GPU specialists for training and a primary cloud’s EU region for production inference.
Oct 28, 2026
4,256 words in the original blog post.
Spain’s National Security Framework (ENS), established by Royal Decree 311/2022, applies to public-sector information systems and private suppliers providing services or solutions to Spanish public bodies, requiring security policies, category-based controls, and documented conformity evidence. Systems are classified as BÁSICA, MEDIA, or ALTA according to the highest assessed impact across confidentiality, integrity, availability, authenticity, and traceability; BÁSICA systems may use self-assessed declarations, while MEDIA and ALTA require audits and certification from accredited bodies, generally renewed every two years. Annex II sets out 73 organisational, operational, and protection measures covering areas such as access controls, separation of duties, logging, change management, supplier oversight, encryption, backups, continuity planning, and cloud-service governance. ENS does not impose a universal Spanish or EU data-residency rule, but it requires organisations to assess data jurisdiction, encryption, key management, and cloud-provider conformity, while individual tenders or GDPR may add residency restrictions. The public body remains responsible for proving its own system’s conformity, so cloud platforms must provide usable audit evidence; the source argues that bring-your-own-cloud arrangements can simplify this by keeping accounts, logs, keys, backups, and regional choices under the customer’s control, though no platform alone makes a system ENS-compliant.
Oct 28, 2026
3,747 words in the original blog post.
Moving from Jenkins-based deployments to GitOps can preserve existing overprivileged cluster access unless Kubernetes multi-tenancy is deliberately configured across several layers. Effective isolation requires per-team, per-environment namespaces supplemented by ResourceQuotas, LimitRanges, default-deny NetworkPolicies enforced by the cluster CNI, and Pod Security Admission labels, while Kubernetes RBAC restricts direct access between tenants. In Argo CD, tenant boundaries depend on AppProjects that limit approved repositories and deployment destinations, combined with project-scoped RBAC grants and an empty default policy so users cannot view or sync other teams’ applications; Git repository write access must also be treated as deployment access. Admission controls such as Kyverno, OPA Gatekeeper, or Kubernetes ValidatingAdmissionPolicy provide a separate safeguard by enforcing manifest requirements including allowed registries, resource limits, labels, and restrictions on privileged workloads. The recommended migration sequence is to establish namespace protections, Argo CD projects and RBAC, Git controls, and audited-then-enforced policies before moving teams from Jenkins and revoking shared cluster-admin credentials, with recurring isolation tests to detect drift. Organizations can operate this stack themselves, use managed Argo CD offerings, adopt alternatives such as Flux, Capsule, or vCluster for different isolation needs, or use an internal developer platform such as Qovery to reduce environment provisioning and RBAC administration while retaining cluster-wide security controls.
Oct 28, 2026
3,969 words in the original blog post.
Cloud governance across GCP and Azure is complicated by different IAM models, resource hierarchies, benchmarks, and finding formats, while the main operational constraint for small teams is often triaging large volumes of alerts rather than purchasing licenses. Microsoft Defender for Cloud and Google Security Command Center provide the lowest-effort continuous posture monitoring in their respective clouds, although their cross-cloud coverage is limited and Security Command Center’s Azure Enterprise support is scheduled to end in May 2027. Open-source tools such as Prowler, Checkov, Trivy, and OPA/Gatekeeper can assess configurations or prevent unsafe infrastructure changes, but generally do not repair live resources; commercial CNAPP platforms including Wiz, Aqua Security, Prisma Cloud, and CrowdStrike offer more extensive workflow automation and remediation at enterprise pricing. The recommended baseline for teams without dedicated platform engineers combines each cloud’s native posture tool, Prowler or Checkov in CI to block high-severity issues before deployment, and a designated person to review critical findings weekly. Because automatic remediation can disrupt intentional production changes, prevention through standardized infrastructure templates, policy controls, and IaC scanning is presented as safer and more durable than fixing drift after deployment. Qovery is positioned not as a CSPM or compliance tool, but as an internal developer platform that can reduce future misconfigurations by provisioning environments consistently from reviewed templates.
Oct 27, 2026
3,733 words in the original blog post.
SecNumCloud is an ANSSI qualification granted to specific cloud services within defined technical, geographic, and time-limited scopes, rather than to entire providers, so buyers should verify that contracted services appear on ANSSI’s current list. The comparison recommends 3DS Outscale for AWS-oriented teams seeking API compatibility and relatively low migration effort, Cloud Temple for VMware-based or legacy applications needing managed operations and qualified OpenShift, and OVHcloud for public pricing and broad infrastructure options, while noting that its standard Public Cloud and Managed Kubernetes Service are not within its qualified perimeter. Other relevant providers include Oodrive for secure collaboration SaaS, S3NS for a Google Cloud-based qualified offering including GKE Autopilot, and Orange Business for designated SecNum services, whereas Bleu and Scaleway were not yet qualified at the time described. SecNumCloud adds EU ownership, operational, and legal-protection criteria beyond ISO 27001 and HDS, making it particularly important for sensitive French public-sector and regulated workloads. The discussion also emphasizes that qualified cloud providers generally do not supply a full internal developer platform, leaving teams responsible for deployment workflows, access control, environments, cluster operations, and CI/CD, while tools such as Qovery can provide self-service deployment capabilities atop customer-owned infrastructure but are not themselves SecNumCloud-qualified.
Oct 27, 2026
3,623 words in the original blog post.
NVIDIA AIStore versions through 4.9 are described as vulnerable to CVE-2026-65162, a CVSS 10.0 flaw in its legacy ETL PodSpec initialization API that could allow an unauthenticated remote caller to submit Kubernetes pod settings enabling privileged access, host filesystem mounts, namespace sharing, or service-account changes. The article advises teams to quickly assess exposure by identifying vulnerable AIStore images, checking API reachability, reviewing the AIStore service account’s pod-creation permissions, confirming enforcing admission controls, and searching for suspicious ETL or host-access pods. It recommends upgrading first to AIStore 5.0, then 5.1, which removes the PodSpec initialization path, while immediately reducing risk on unpatched deployments through default-deny NetworkPolicies, targeted admission policies, RBAC cleanup, credential rotation where compromise is suspected, and audit logging. More broadly, it argues that any platform accepting user-supplied PodSpecs can become a privilege-escalation path and that managed Kubernetes services do not remove customers’ responsibility to patch workloads, restrict pod creation, and enforce security controls.
Oct 27, 2026
4,520 words in the original blog post.
Microsoft’s sovereign cloud portfolio offers differing levels of exposure to the US CLOUD Act: Sovereign Public Cloud remains highly exposed because Microsoft operates it, Sovereign Private Cloud reduces exposure by placing Azure Local infrastructure under customer or European-partner operation, and National Partner Clouds such as France’s Bleu and Germany’s Delos Cloud offer the lowest exposure because EU-owned companies operate and contract for the services. The discussion emphasizes that corporate control, contractual entities, operational responsibility, and encryption-key custody matter more than the physical location of servers, since the CLOUD Act can require US providers to disclose data under their possession, custody, or control worldwide. Customer-controlled keys stored in external hardware security modules can limit accessible data to ciphertext, although this protection weakens if a provider can recover keys. Sovereign Private Cloud is presented as providing data and operational control but not technological independence, because Microsoft still controls the software, updates, licensing, and product roadmap. Comparisons with AWS, Google, Oracle, and European providers highlight a trade-off between hyperscalers’ broader service portfolios and European providers’ stronger jurisdictional separation and certifications. Organizations are advised to classify workloads by sensitivity, scrutinize contracts and support terms, use portable technologies such as Kubernetes, Terraform or OpenTofu, PostgreSQL, and S3-compatible storage, and regularly test redeployment to alternative providers to preserve a practical exit strategy.
Oct 26, 2026
4,325 words in the original blog post.
Startups evaluating alternatives to Red Hat OpenShift often seek to avoid its per-core or per-worker-vCPU subscription costs, operational footprint, and proprietary primitives while retaining governance features such as environment-level RBAC, audit trails, policy controls, and self-service deployments. The comparison ranks Qovery, Google GKE Autopilot, Amazon EKS Auto Mode, SUSE Rancher Prime with RKE2, Linode Kubernetes Engine, DigitalOcean Kubernetes, Talos Linux with Sidero Omni, and Cluster API, while identifying OKD, Plural.sh, Portainer, and Northflank as runner-ups. It distinguishes between managed Kubernetes services, which generally provide low-cost operated control planes but leave governance and platform engineering to customers; platform layers, which add deployment workflows and access controls; and self-hosted distributions or tools suited to air-gapped, on-premises, or sovereign environments but requiring more Kubernetes expertise. The author presents Qovery, the author’s own product, as the closest option for startups seeking OpenShift-like governance in their own cloud account with minimal platform staffing, while recommending Rancher, OKD, or Talos for highly controlled infrastructure and LKE or DOKS for teams prioritizing low infrastructure cost. It emphasizes that engineer time, upgrades, security patching, policy maintenance, and idle cloud resources may outweigh licensing differences, and explains that migration usually involves translating OpenShift-specific components such as Routes, DeploymentConfigs, SCCs, BuildConfigs, ImageStreams, and projects into upstream Kubernetes equivalents, with operator dependencies posing the largest challenge.
Oct 26, 2026
5,978 words in the original blog post.
ArgoCD records deployment synchronization details such as commit revisions, targets, timing, and manual sync initiators, but it does not independently establish who approved a deployment, particularly for automated syncs. A defensible multi-cluster audit trail therefore correlates Git pull-request reviews and protected-branch controls, ArgoCD sync events, Kubernetes API audit logs, and cloud IAM activity using the commit SHA, while exporting evidence to durable, tamper-resistant storage because ArgoCD history and Kubernetes events have limited retention. The recommended security model uses SSO-backed ArgoCD access, disables local administrator accounts, separates sync permissions from override and deletion rights, and restricts each AppProject’s repositories, destinations, and permitted resource types. Kubernetes auditing must be explicitly enabled on most platforms, including EKS, GKE Data Access logging, and AKS, and should avoid recording sensitive Secret contents while retaining write operations and RBAC changes. Organizations can build this correlation themselves with a SIEM and notification pipelines, use managed ArgoCD services, add promotion tools such as Kargo, or adopt developer platforms such as Qovery, which provides per-environment access controls and deployment history but does not replace cluster API auditing or function as a SIEM.
Oct 26, 2026
3,959 words in the original blog post.
A Qovery webinar replay argues that cloud-migration delays are driven more by organizational decision-making than by technical execution, estimating that technical work accounts for roughly 20% of a migration while planning, risk concerns, staffing, trust, and perceived value consume the rest. It identifies five common blockers: migrations feeling irreversible, limited internal capacity, fear of unfamiliar technologies, benefits that are not immediately visible, and uncertainty about vendors or service providers. Two anonymized customer examples illustrate this pattern, with decisions taking seven to twelve weeks while implementation took between an afternoon for staging and five days to a week for production. The session recommends reducing hesitation by keeping the existing stack available as a fallback, testing a target environment cheaply through a pilot or staging deployment, migrating in waves with databases moved last, and evaluating the cost of postponement. It also presents Qovery’s AI-assisted platform as a tool for automating discovery, planning, infrastructure setup, and deployment, while maintaining human approval for critical steps such as cutovers and database migrations.
Oct 26, 2026
2,649 words in the original blog post.
AWS Transform is presented as an agentic AI migration and modernization service for VMware and other server estates, IBM z/OS and Fujitsu mainframes, and .NET applications, delivering migrated workloads, infrastructure artifacts, landing-zone components, transformed code, and migration plans rather than a finished internal developer platform. After cutover, organizations remain responsible for operating accounts, networks, IAM, runtimes such as EKS, ECS, or EC2, CI/CD, monitoring, backups, patching, cost controls, and Kubernetes upgrades, while developers may still depend on tickets or infrastructure expertise to create environments. The text distinguishes migration tooling from developer platforms, defining the latter as self-service systems with golden paths, service catalogs, guardrails, preview environments, and simplified deployment workflows. It outlines three approaches to filling this post-migration gap: assembling AWS-native services, building and maintaining an open-source stack around tools such as Backstage, Terraform or OpenTofu, Crossplane, and Argo CD, or adopting a commercial platform product. It argues that the appropriate choice depends on organizational scale, platform-engineering capacity, governance requirements, and tolerance for third-party services, while noting that AWS Proton has reached end of support and that Backstage and Terraform alone do not constitute complete platforms. The recommended sequence is to migrate and stabilize workloads first, then introduce self-service for suitable containerized and greenfield applications, measuring delivery performance, ticket reduction, and non-production cloud spending.
Oct 25, 2026
4,018 words in the original blog post.
Meta’s Muse Spark Open is presented as an 816-billion-parameter mixture-of-experts model with roughly 42 billion active parameters per token, multimodal inputs, a one-million-token context window, and BF16 and NVFP4 weight options, though its large total parameter count makes GPU memory rather than compute the primary deployment constraint. The guide estimates that the model requires about 530 GB of weights in Meta’s NVFP4 format plus runtime overhead, making a single 8xH200, 8xB200, or comparable high-memory GPU node suitable for quantized serving, while BF16 deployment generally requires multiple nodes connected by high-bandwidth InfiniBand or EFA networking. It recommends Kubernetes components including the NVIDIA GPU Operator, vLLM or SGLang for inference, LeaderWorkerSet for multi-node replicas, shared storage for model weights, and DCGM-based monitoring, with long startup probes and careful NCCL configuration to avoid loading failures or performance degradation. The analysis argues that hosted APIs will usually be less expensive and operationally simpler unless organizations have sustained high GPU utilization, strict data-residency requirements, or specialized customization needs such as private LoRA adapters, routing, or modified serving behavior. It also describes Qovery as a bring-your-own-cloud platform for deploying and managing this Kubernetes infrastructure rather than a provider of GPUs or model hosting.
Oct 25, 2026
3,950 words in the original blog post.
Nvidia Nemotron 3.5 Bolt is an open-weights, agent-focused model family designed for single-device deployment, with the flagship 74B-A9B model compressed to a roughly 24GB 2-bit checkpoint despite having 74 billion total parameters and 9 billion active parameters per token, while the smaller E4B targets lightweight voice and edge use cases. The guide recommends an L40S 48GB GPU with llama.cpp for lower-cost single-stream deployments or an RTX PRO 6000 Blackwell 96GB GPU with vLLM for production batching, noting that 24GB L4 and A10G cards are suitable only for E4B. vLLM is presented as the preferred 74B serving option because of NVIDIA’s official recipe, continuous batching, OpenAI-compatible APIs, and tool-calling support, whereas NVIDIA NIM currently supports only E4B. Although self-hosting can provide data control, predictable fixed spending, freedom from rate limits, and use of existing cloud commitments, it is generally more expensive per token than hosted APIs unless a GPU sustains billions of monthly tokens; an always-on RTX PRO 6000 instance may require roughly 10.3 billion blended tokens per month to match cited API pricing. The central operational recommendation is to control idle GPU time through scheduled shutdowns, scale-to-zero policies for non-production environments, context and concurrency tuning, persistent model caches, monitoring, and consolidating work onto well-utilized GPUs, while choosing between raw VMs, Kubernetes, managed GPU services, or cloud-management platforms based on operational capacity.
Oct 25, 2026
4,544 words in the original blog post.
For small teams without dedicated MLOps staff, the comparison recommends Google Cloud Vertex AI as the simplest greenfield option because it consolidates training, registries, pipelines, and endpoints under a relatively unified service and IAM model, while Amazon SageMaker AI is presented as the better fit for organizations whose data, networking, compliance requirements, and cloud commitments already reside on AWS. The central decision factor is data gravity, since transferring large datasets and rebuilding identity and networking across clouds can outweigh modest GPU price differences. Managed Kubernetes options such as GKE Autopilot and EKS with Karpenter offer portability and lower costs for sustained inference workloads but require teams to manage upgrades, GPU configuration, autoscaling, monitoring, and on-call operations. Qovery is positioned as a middle path for deploying containerized inference workloads in a team’s own cloud account while reducing Kubernetes operational burden, though it does not replace managed training, feature stores, model registries, or AutoML. Serverless GPU providers including Modal, Replicate, and Hugging Face Inference Endpoints are described as fastest for prototypes and irregular traffic because they scale to zero, but can become more expensive than committed cloud capacity as utilization rises. Across all options, the analysis emphasizes that idle GPU time is usually a larger cost concern than AWS-versus-GCP pricing, and advises teams to choose based on data location, need for managed ML capabilities, traffic pattern, compliance requirements, and responsibility for production incidents.
Oct 25, 2026
4,380 words in the original blog post.
SecNumCloud and BSI C5 can be addressed through one shared Kubernetes platform and delivery process, but they require separate hosting footprints because SecNumCloud is an ANSSI qualification held by specific cloud providers and services, with EU sovereignty and extraterritorial-law requirements, while BSI C5 is an ISAE 3000 attestation that can cover a company’s own cloud service on many providers. For French public-sector workloads within SecNumCloud scope, organizations should use a dedicated cluster in an ANSSI-qualified environment, while German and other enterprise workloads can run in a standard EU cloud cluster; shared GitOps workflows, manifests, RBAC models, policies, image-signing processes, and audit tooling can operate both. Data, backups, secrets, logs containing personal information, ingress, and identity boundaries should remain separated by footprint, and organizations must carefully document control-plane access, staff jurisdiction, vendor support, and service dependencies. Most security evidence for SecNumCloud, C5, ISO 27001, and SOC 2 can be reused through automated records for access control, changes, logging, encryption, patching, backups, and incident response, though SecNumCloud’s provider ownership and jurisdiction requirements remain unique. The approach reduces costs compared with operating two independent stacks, but qualified providers may have narrower managed-service options and require greater operational effort. Qovery is presented as a tool for deploying consistently to customer-owned sovereign and conventional clusters, but it neither grants SecNumCloud qualification nor provides a C5 attestation.
Oct 25, 2026
4,884 words in the original blog post.
The comparison ranks seven container platforms by how quickly their limitations may require a move to Kubernetes, arguing that the decisive factor is usually unsupported workload or infrastructure requirements rather than raw compute capacity. AWS App Runner is placed first because it supports only HTTP-driven containers without native jobs, cron, or sidecars and has been closed to new customers since April 2026; Google Cloud Run follows, offering jobs, sidecars, GPUs, and longer timeouts but lacking support for stateful services, custom operators, and cluster-level networking controls; Azure Container Apps ranks last among the major managed offerings because its Kubernetes-based foundation includes KEDA, Dapr, jobs, and sidecar capabilities, though it does not expose the Kubernetes API. ECS Fargate, Fly.io, Render, and managed Kubernetes options are also assessed according to workload coverage, networking, account ownership, portability, and cost. The analysis notes that serverless per-second pricing can favor intermittent workloads, while several continuously running services may become less expensive on a right-sized Kubernetes node pool, although savings depend on utilization and commitments. It presents Cloud Run as relatively easier to migrate from because of Knative-compatible concepts, while App Runner and Azure Container Apps require more configuration rewriting, and it promotes running Kubernetes in a company-owned cloud account through Qovery as a way to avoid a later platform migration.
Oct 24, 2026
4,515 words in the original blog post.
Sovereign EU cloud regions differ from standard AWS or Azure EU regions mainly by strengthening operational and legal control over who runs systems, accesses data, manages encryption keys, handles telemetry and responds to foreign government requests, rather than simply storing data in Europe. GDPR, DORA and NIS2 generally do not require sovereign hosting, while national schemes, sector-specific rules and public procurement requirements such as France’s SecNumCloud and HDS, Germany’s C5-related public-sector requirements, Spain’s ENS and Italy’s ACN qualification may do so. The article recommends evaluating providers through nine areas: operating entity, access controls, key ownership, data flows for metadata and support, certification scope, service availability, portability, foreign-request procedures and total cost. It argues that missing managed services and feature gaps often create greater challenges than the roughly 15% estimated price premium for AWS’s European Sovereign Cloud, because organizations may need to self-host services and maintain a second deployment environment. Most organizations are likely to adopt a split estate, placing regulated workloads in sovereign or EU-only environments while retaining standard regions for other applications, and can reduce lock-in by using portable technologies such as Kubernetes, OCI images, open service protocols and regularly tested exit plans.
Oct 24, 2026
4,097 words in the original blog post.
Spot GPU instances can be reclaimed with little warning—typically two minutes on AWS and about 30 seconds on Azure or Google Cloud—so training workloads must be designed to resume rather than restart. The recommended approach is to save complete, atomic checkpoints to durable regional object storage on a time-based schedule, including model, optimizer, scheduler, scaler, random-number-generator, and dataloader states, while retaining recent valid versions. Training applications should catch SIGTERM, set a shutdown flag, finish a safe point, write an emergency checkpoint within the available grace period, and exit cleanly; Kubernetes Jobs can then recreate failed pods and restore from the latest checkpoint. For distributed workloads, sharded and asynchronous checkpointing, elastic frameworks such as torchrun or Ray, gang scheduling, diversified spot pools across instance types and availability zones, and on-demand or reserved fallback capacity can reduce interruption impact. The discussion also emphasizes measuring actual checkpoint duration and local interruption rates, since GPU spot volatility differs by provider, region, and hardware type, and presents Qovery as a platform for operating Kubernetes clusters in a customer’s cloud account while external tools manage training orchestration and recovery.
Oct 24, 2026
4,033 words in the original blog post.
Kubernetes 1.37 became available on Amazon EKS on October 1, 2026, introducing stable Dynamic Resource Allocation device taints and tolerations, which allow operators to isolate a failed accelerator such as one GPU without draining an entire node, and graduating the metrics.k8s.io/v1 API without changing its fields from v1beta1. Clusters must continue serving v1beta1 because the Horizontal Pod Autoscaler still depends on it, while custom and external metrics APIs used by tools such as Prometheus Adapter and KEDA remain beta and warrant compatibility testing. Pod-level CPU, memory, and hugepage resource budgets remain a beta feature enabled by default since EKS 1.34, offering potential efficiency gains for sidecar-heavy Linux workloads but requiring staged validation with quotas, limits, and autoscaling tools. The recommended upgrade sequence is non-production clusters, stateless production workloads, GPU clusters using compatible DRA drivers, and finally clusters with KEDA, older operators, or sensitive StatefulSets, although clusters on EKS 1.34 or older should upgrade before standard support ends to avoid the substantially higher extended-support charge. Pre-upgrade checks should include deprecated APIs, add-on and Karpenter versions, webhooks, CRDs, disruption budgets, SELinux shared-volume behavior, and node-version skew; EKS permits a one-minor-version control-plane rollback within seven days, but add-ons and nodes require separate recovery planning.
Oct 24, 2026
4,017 words in the original blog post.
Qovery has strengthened AWS account access security by sending each Qovery organization’s ID as the sts:ExternalId when assuming IAM roles and adding a trust-policy condition that permits role assumption only for that organization. Previously, this isolation was enforced solely through Qovery’s application logic and database constraints; AWS now independently enforces it within the customer’s account, providing defense in depth. Existing Qovery IAM roles have already been updated without service interruption, but customers using the qovery-role-creation CloudFormation stack should update it to the latest template to prevent future stack changes from removing the new condition. Qovery also recommends that customers still using static AWS credentials migrate to STS role assumption, which uses short-lived credentials and benefits from the strengthened trust relationship.
Oct 23, 2026
891 words in the original blog post.
A 2026 comparison of cloud accelerators argues that Google Cloud TPU Trillium v6e and AWS H100 GPUs have nearly identical on-demand costs for a comparable 70B Llama-class serving workload, at roughly $0.94 and $0.91 per million output tokens respectively, while long-term commitments can reduce Trillium to about $0.42 per million tokens. AWS L40S instances are presented as a particularly economical option for 8B models, reaching about $0.32 per million tokens with a far smaller hourly billing unit, while CUDA compatibility, mature tooling, and flexibility favor AWS GPUs for many teams. TPU economics are strongest for high-utilization, XLA-native workloads with committed capacity, but porting effort and limited support for custom CUDA-dependent models can add engineering costs. AWS Inferentia2, Trainium2, Azure H100, Scaleway, and neutral GPU clouds offer alternatives shaped primarily by model compatibility, regional requirements, capacity, and cloud commitments rather than headline prices. The comparison emphasizes that utilization, batching, quantization, autoscaling, spot capacity, and reserved-use discounts usually affect total cost more than accelerator selection, noting that idle hardware can multiply per-token costs. It also positions Qovery as a Kubernetes platform for operating workloads within customers’ own cloud accounts so that they retain discounts, data residency, and infrastructure control.
Oct 23, 2026
3,112 words in the original blog post.
Migrating production workloads from AWS to Google Cloud without rewriting application code requires selecting a workload-specific approach—rehosting virtual machines, replatforming containerized applications, or reprovisioning infrastructure—and typically combines separate compute and data migration tools. Google Migration Center and Migrate to Virtual Machines are positioned as the default for EC2 and server lift-and-shift moves, while GKE and Qovery support container portability, with Qovery targeting teams that want managed Kubernetes operations across their own cloud accounts. Terraform or OpenTofu can recreate infrastructure but do not move workloads or data, while Database Migration Service and Storage Transfer Service address databases and S3 data separately; Carbonite Migrate and Cloudsfer serve narrower legacy-server and file-transfer use cases. The discussion emphasizes that AWS-specific integrations such as S3 SDK calls, Secrets Manager, IAM roles, Lambda, DynamoDB, SQS/SNS, and Kinesis may require configuration changes or full rewrites, despite a nominally “no rewrite” migration. It also notes that CloudEndure is retired and its AWS successor cannot migrate workloads out of AWS, and recommends a staged cutover involving inventory, landing-zone setup, data replication, parallel operation, traffic shifting, rollback planning, and AWS decommissioning. Major migration costs include AWS data egress, duplicated infrastructure during parallel runs, cross-cloud latency, and forgotten AWS resources after cutover.
Oct 22, 2026
4,456 words in the original blog post.
Recent disclosures involving Railway, CircleCI, and Heroku illustrate that deployment platforms can expose customer secrets when decryption systems, sessions, tokens, logs, or unmanaged copies are compromised. The article recommends maintaining a single secret-management source of truth, providing platforms with secret references rather than values, using short-lived and tightly scoped credentials, monitoring access and usage, avoiding secret logging, and regularly rehearsing rotation procedures. It explains that Qovery separates its SaaS control plane, which stores configuration and optionally customer-provided secrets, from customer-owned Kubernetes data planes that retain workloads, data, logs, metrics, TLS keys, and runtime secrets. Qovery-stored secrets are encrypted per organization, masked after entry, scoped through RBAC, audited, and delivered only to required deployments, though the company acknowledges that any platform capable of decrypting secrets retains some exposure. For more sensitive credentials, Qovery supports references to AWS Secrets Manager, Parameter Store, and GCP Secret Manager through an in-cluster External Secrets Operator, or direct runtime retrieval by applications using IAM roles, keeping secret values outside Qovery’s control plane. Doppler and Infisical can also synchronize secrets into Qovery, while the company cites SOC 2 Type II certification, compliance with HIPAA, DORA, and GDPR, annual penetration testing, and open-source cluster components as elements of its security approach.
Oct 22, 2026
4,073 words in the original blog post.
SecNumCloud is an ANSSI qualification that applies to specific cloud services and datacentres rather than entire providers, making it a central requirement for French public-sector workloads involving sensitive data under the “Cloud au centre” doctrine and related law. The comparison, dated 22 October 2026, finds that 3DS Outscale, OVHcloud, Cloud Temple, Worldline and Cegedim hold listed qualifications for defined IaaS offerings, while Numspot and S3NS have announced newer qualifications, and Scaleway and Bleu remain in qualification processes. It emphasizes that managed Kubernetes and database services frequently fall outside qualified scopes, with Cloud Temple’s OpenShift PaaS presented as a notable qualified container option, while many users may need to operate Kubernetes themselves on qualified IaaS. OVHcloud is positioned as a broad and comparatively inexpensive option but with qualified private infrastructure distinct from its public-cloud managed services; Outscale and Cloud Temple suit strict ministry, defence, and critical-infrastructure needs; Numspot targets public-sector buyers through its ownership and SecNumCloud-first approach; and Scaleway is highlighted for developer experience despite lacking a qualification at the stated date. The piece also discusses Clever Cloud, Bleu, and S3NS as alternatives for simpler PaaS use cases or Microsoft- and Google-oriented environments, and recommends portable containers, Helm, Terraform, open observability standards, and minimal reliance on proprietary services to avoid cloud-specific rewrites. It presents Qovery, the author’s company, as an optional deployment layer across customer-owned cloud accounts and Kubernetes clusters, while clarifying that Qovery itself is not SecNumCloud-qualified and does not replace the underlying provider’s compliance scope.
Oct 22, 2026
4,043 words in the original blog post.
A six-engineer team can operate staging and separate EU and US production Kubernetes clusters without a dedicated platform team if all environments are generated from a common template and differ only in controlled variables such as region, cloud account, scale, and endpoints. The central operational requirements are standardized cluster lifecycle management, one application definition, consistent environment configuration, centrally managed secrets, scoped deployment permissions, and regional data boundaries, particularly for EU residency requirements. The article argues that Kubernetes upgrade cycles, add-on compatibility, identity, networking, and configuration drift create the main maintenance burden, estimating that four clusters can require roughly a part-time engineer when managed through a DIY GitOps stack using tools such as Argo CD or Flux, Terraform or OpenTofu, Helm or Kustomize, and External Secrets Operator. It presents managed internal developer platforms, including Qovery, as an alternative that handles cluster creation, upgrades, configuration, secrets injection, and RBAC within customer-controlled cloud accounts, while suggesting a hybrid model in which GitOps manages infrastructure add-ons and a platform manages application delivery. A sustainable four-cluster setup includes separate production accounts and databases for EU and US workloads, a defined promotion path from staging through regional production, centralized observability and secrets, restricted production access, and automated shutdown of non-production environments to reduce costs.
Oct 22, 2026
3,985 words in the original blog post.
Sandbox environments are ephemeral, isolated copies of an entire application stack—including services, databases, queues, configuration, and networking—created for individual AI-agent tasks, whereas code sandboxes such as microVMs or containers primarily support fast, isolated code execution and unit-test loops. The approach is intended to let agents validate multi-service changes, database migrations, infrastructure updates, integration tests, and production-bug reproductions against realistic dependencies without affecting shared staging or production systems. On Qovery, these environments are cloned from a blueprint into separate Kubernetes namespaces within a customer’s cloud account, seeded through deployment-pipeline lifecycle jobs, and accessed through APIs, CLI tools, or an MCP server with scoped permissions, RBAC controls, sandbox-specific secrets, and audit logs. Environments can be removed by agents, pull-request events, automated cleanup tasks, or time-to-live policies, though they require minutes to deploy and incur real infrastructure costs. The proposed workflow combines code sandboxes for rapid editing and debugging with full sandbox environments for end-to-end verification, using measures such as containerized databases, spot instances, autoscaling, skipped services, and automatic deletion to manage cost.
Oct 22, 2026
4,243 words in the original blog post.
Running Falcon, Mixtral, Llama, Qwen, and embedding models on a shared Kubernetes GPU pool can reduce idle capacity and operational overhead, provided workloads are sized and scheduled according to VRAM, isolation, and performance requirements rather than assigned one node pool per model. The recommended approach uses the NVIDIA GPU Operator with Kubernetes labels, resource limits, namespaces, priorities, taints, and autoscaling, while GPU sharing can be implemented through time-slicing, MIG hardware partitions, experimental MPS, or application-layer multi-model servers such as vLLM, Triton, Ray Serve, and KServe ModelMesh. Memory planning must account for model weights, KV cache, and runtime overhead, with quantization and context-length caps serving as major controls; for example, several 7–8B models can share an 80 GB GPU, whereas Mixtral 8x7B requires substantial memory despite activating fewer parameters per token. Separate GPU pools are generally warranted only for different GPU types, pricing models such as spot versus reserved capacity, strict physical isolation, or incompatible CUDA driver needs. The discussion also contrasts self-hosted inference, which may become economical at sustained utilization and supports data residency and reserved-capacity discounts, with hosted endpoints that can be simpler and less costly for low or highly variable demand.
Oct 22, 2026
4,048 words in the original blog post.
AWS Trainium2 can support self-hosted AI inference through the Neuron SDK’s native PyTorch backend, TorchNeuron, which uses the torch-neuronx package, a neuron device, and torch.compile rather than CUDA or PyTorch/XLA. The guide describes deploying supported models such as Llama 3, Qwen3, and GPT-OSS on trn2.48xlarge instances through the beta vLLM Neuron plugin, which provides an OpenAI-compatible API, while noting that NxD Inference remains an option for some models but is in maintenance status. Deployment requires Neuron drivers, runtime libraries, the neuronx-cc compiler, and persistent compile caches, since compiled artifacts depend on model shapes, tensor-parallel settings, and SDK versions. On Amazon EKS, workloads use aws.amazon.com/neuron or aws.amazon.com/neuroncore resources, with the Neuron device plugin and scheduler supporting allocation of contiguous cores. Trainium2 Capacity Blocks are presented as potentially less expensive per hour than H100-based P5 instances, but the guide emphasizes that actual cost effectiveness depends on utilization, latency targets, model compatibility, recompilation overhead, and engineering effort. It recommends Trainium2 chiefly for stable, high-volume traffic involving a small number of supported models, while NVIDIA GPUs remain preferable for broad framework compatibility, custom CUDA kernels, newly released architectures, and bursty demand.
Oct 22, 2026
3,772 words in the original blog post.
Google Cloud capacity shortages in 2026 can affect CPU-based virtual machine families as well as GPUs, causing GKE autoscaling failures and pending pods through errors such as ZONE_RESOURCE_POOL_EXHAUSTED, which indicate unavailable physical hardware rather than insufficient quota. Google’s recommended responses include retrying, changing zones or machine types, using Spot or Flex-start capacity, reserving resources, and planning fallback options, while company statements and customer reports suggest supply constraints are broad and persistent. The piece presents AWS as a complementary capacity pool rather than an unlimited alternative, citing its larger infrastructure footprint, wide EC2 instance selection, and independent regions and supply chains, while noting that AWS also faces AI-related constraints. It recommends offloading portable workloads such as CI runners, batch processing, queue workers, preview environments, stateless APIs, and some GPU inference workloads, while retaining databases, data-intensive applications, and latency-sensitive services near their GCP data. It argues that the main obstacle to multicloud bursting is operational complexity across Kubernetes, identity, networking, deployment pipelines, and monitoring, and promotes Qovery as a platform for managing GKE and EKS environments through a shared control plane, allowing teams to clone and deploy environments across clouds while considering cross-cloud latency, data-egress costs, connectivity, and workload suitability.
Oct 21, 2026
3,261 words in the original blog post.
Data-center GPU capacity remains constrained in 2026 as major providers including Google Cloud, CoreWeave, and Nebius report limited or sold-out capacity, particularly for on-demand access to recent GPU generations. The available market spans hyperscalers, GPU-focused neoclouds, regional and developer clouds, and capacity marketplaces, with smaller providers sometimes offering faster access for inference or short-term workloads despite having smaller overall fleets. Managed Kubernetes is presented as a key portability layer because it allows containerized GPU applications, Helm charts, and deployment workflows to run across multiple providers, although GPU types, regions, private connectivity, and compliance capabilities vary. Organizations adding a second cloud must plan how application requests, model weights, and private data will move between environments, using public APIs, VPNs, overlay networks, private interconnects, or cross-cloud networking tools depending on traffic and security requirements. The main challenge is often operational rather than technical, since each provider introduces separate clusters, identity systems, pipelines, monitoring, and on-call responsibilities; the post promotes centralized platforms such as Qovery as a way to manage deployments across clusters while leaving cloud accounts, networking, and commercial agreements under the customer’s control.
Oct 21, 2026
2,447 words in the original blog post.
The guide presents Qovery as a platform for deploying repositories, monorepos, Docker Compose stacks, and multi-service applications to Kubernetes infrastructure running in a customer-owned AWS, GCP, Azure, Scaleway, or existing cluster account, with managed databases, HTTPS, autoscaling, preview environments, deploy-on-merge, and rollback support. It explains how Compose services can be translated into production applications, ingress, persistent storage, internal service discovery, health probes, and managed Postgres or Redis, while recommending provider-managed data services for backups, recovery, and credential handling. The author argues that Qovery reduces the operational work of maintaining Kubernetes, secrets synchronization, CI/CD glue, and multi-environment workflows, and describes agent-oriented installation and deployment tools for Claude Code, Cursor, Codex, and similar assistants. It also acknowledges that native products such as Google Cloud Run, AWS App Runner, and Azure App Service are usually simpler and faster for a single stateless service, while positioning a platform layer as more useful for multiple services, environments, pull-request previews, compliance controls, or multi-cloud deployments.
Oct 21, 2026
5,672 words in the original blog post.
SOC 2 Type II and ISO 27001 Stage 2 auditors assess an organization’s controls over cloud workloads rather than the underlying cloud provider, focusing on system-generated evidence for access management, approved production changes, encryption and key custody, centralized retained logging, and risk and supplier reviews. Under the shared-responsibility model, providers such as AWS, GCP, Azure, and Scaleway handle physical infrastructure and certain managed services, while customers remain responsible for IAM, network settings, patching, encryption configuration, backups, and application security; provider reports and complementary user entity controls help define these obligations. Auditors test complete populations of events over an observation period, so dated records tying access grants, removals, deployments, approvals, logs, and restore tests to named individuals are more important than policies or isolated screenshots. The article recommends migrating before a SOC 2 observation window where possible, using one enforced deployment pipeline, infrastructure as code, environment-specific RBAC, immutable deployment histories, and nonproduction environments without live sensitive data to reduce evidence gaps. It also notes that compliance automation tools can collect evidence but cannot replace operating controls, and presents Qovery as a platform intended to provide governed deployments within an organization’s own cloud account while retaining control of data, keys, regions, and audit logs.
Oct 21, 2026
4,949 words in the original blog post.
A phased hybrid cloud migration keeps on-premises and cloud environments serving production simultaneously while workloads move through six stages: discovery and dependency mapping, landing-zone and network setup, a low-risk pilot, production waves, data-heavy and legacy migrations, and decommissioning. It recommends sequencing work by dependency clusters rather than application lists, retiring unused systems early, moving stateless services first, and migrating tightly coupled applications and databases together unless WAN latency and query volume are low enough to safely separate them. Dedicated private connectivity, non-overlapping IP ranges, shared DNS, identity, observability, and standardized deployment processes should be established before production waves, with interconnect lead times often determining the schedule. Database migrations should use bulk copying, change-data capture, short write freezes, validation checks, reverse replication, and staged traffic shifts, while numeric rollback triggers and rehearsed recovery procedures limit disruption. The plan emphasizes that hybrid operation should have a written end date, typically 12–24 months, because duplicated infrastructure, tooling, operations, and licensing can be more costly than cutovers; it also presents Kubernetes-based platform standardization and Qovery’s deployment tools as ways to reduce operational differences between on-premises and cloud environments.
Oct 21, 2026
5,603 words in the original blog post.
Mistral’s Timestral family of open-weight time-series forecasting models, released in four sizes from 5M to 209M parameters, is presented as suitable for self-hosting because the models can run efficiently on CPUs, with the largest model reportedly producing forecasts in about 140 milliseconds on an eight-thread server. The weights use a Modified MIT license that permits commercial use only for organizations below a $20 million monthly consolidated-revenue threshold, while the inference code is Apache 2.0. Self-hosting is positioned as most useful for sensitive or regulated data, high-volume scheduled forecasting, locality requirements, and stable pinned model versions, whereas hosted APIs are described as more practical for lower-volume workloads due to reduced operational effort. Recommended deployment practices include packaging the model behind an HTTP service, preloading weights, using readiness and startup probes, batching series, pinning software and model revisions, and scaling Kubernetes workloads by concurrency or queue depth rather than CPU utilization. The article argues that CPU instances are generally sufficient, GPUs are needed mainly for high-throughput endpoints, and cost control depends on using scale-to-zero, spot capacity for batch jobs, and avoiding idle GPU infrastructure. It also stresses operational safeguards such as backtesting, input-drift monitoring, authentication, rollback plans, and license review, while promoting Qovery as a platform for deploying such services in a customer’s own cloud account or Kubernetes cluster.
Oct 21, 2026
3,958 words in the original blog post.
EU public-sector cloud procurement in 2026 increasingly focuses not only on EU data residency but also on operational control and insulation from foreign legal demands, requiring software vendors to match infrastructure choices precisely to tender language. The comparison distinguishes AWS European Sovereign Cloud as a separate EU-operated hyperscaler partition, Google Cloud Sovereign Controls as configurable safeguards on standard regions with partner and air-gapped options, and Microsoft Cloud for Sovereignty as a governance layer over commercial Azure regions, while noting that French and German providers such as OVHcloud, Scaleway, STACKIT, IONOS, Open Telekom Cloud, 3DS Outscale, and Cloud Temple may better meet no-US-parent, SecNumCloud, or BSI C5 requirements. AWS is positioned for EU-only operations on a hyperscaler stack, Google for flexible controls and mature Kubernetes, Microsoft for customers deeply integrated with Microsoft 365 and Entra ID, and European providers for stricter national sovereignty clauses despite generally smaller service catalogues. The discussion emphasizes that certifications apply to specific services rather than entire providers, EUCS remains unadopted, and legal protection from US extraterritorial requests remains debated even for EU-based subsidiaries of US companies. It also argues that the largest practical challenge is supporting the same software across several required jurisdictions, recommending portable components such as Kubernetes, PostgreSQL, S3-compatible storage, OIDC, and OpenTelemetry, supported by a common deployment workflow rather than separate cloud-specific platforms.
Oct 21, 2026
5,280 words in the original blog post.
For five-person engineering teams using Kubernetes without a dedicated platform engineer, the guide recommends GitHub Actions or GitLab CI for building and testing alongside a managed bring-your-own-cloud Kubernetes platform for deployments, upgrades, ingress, TLS, autoscaling, and pull-request preview environments. It argues that free GitOps tools Argo CD and Flux CD are mature, portable, CNCF-graduated options for reconciling Git-defined infrastructure, but do not manage cluster lifecycle or the supporting components needed for reliable previews, leaving substantial operational work to the team. Argo CD is presented as better for teams wanting a dashboard and built-in pull-request application generation, while Flux suits terminal-oriented teams seeking a smaller controller footprint and native image automation; both are recommended only when an engineer is willing to own platform operations or compliance requires Git-based declarative state. The guide estimates self-managed GitOps can consume 20–30 percent of an engineer’s time, compared with managed platforms such as Qovery, Northflank, and Porter, which charge platform fees but retain workloads, cloud billing, data control, and cloud discounts in the customer’s own account. It also notes that a simple CI pipeline using Helm can be sufficient for a single trusted team and cluster, while GitOps becomes more valuable with multiple clusters, configuration-drift concerns, or audit requirements.
Oct 21, 2026
5,611 words in the original blog post.
AKS on bare metal Ubuntu, introduced as a public preview in late 2026, enables Kubernetes to run directly on customer-owned Ubuntu servers through Azure Arc, but it has no SLA, is limited to single-node clusters in East US, and is not intended for production use. While direct hardware access can benefit GPU, storage, latency-sensitive, data-residency, and consistently high-utilization workloads, it shifts responsibility for physical infrastructure, firmware, provisioning, storage, networking, spare capacity, and hardware incidents from Azure to the customer. The comparison with managed AKS on Azure VMs emphasizes that virtualization overhead is generally modest for typical applications, whereas managed cloud services provide elastic capacity, automated node replacement, built-in load balancing, and SLA-backed control planes. The discussion concludes that teams without dedicated platform or hardware operations expertise should generally use managed AKS and first pursue right-sizing, cloud commitments, autoscaling, and non-production shutdowns before buying hardware. It also distinguishes a Kubernetes cluster from a developer platform, presenting tools such as Qovery as a way to add deployment self-service, preview environments, RBAC, and cost controls to existing cloud or on-premises clusters without taking over responsibility for the underlying hardware.
Oct 20, 2026
3,246 words in the original blog post.
Anyscale’s general availability on Azure AKS expands its managed Ray service across AWS, GCP, and Azure, but the discussion argues that Ray itself remains portable because it is open source and can run on Kubernetes through KubeRay. It distinguishes Ray as the distributed computing framework, KubeRay as its Kubernetes operator, and Anyscale as a commercial managed platform, noting that lock-in is more likely to arise from proprietary job APIs, workspaces, identity integrations, storage conventions, and data location than from Ray or a particular cloud. A portable approach uses shared KubeRay RayCluster, RayJob, and RayService manifests, common container images, cloud-specific configuration overlays for GPU infrastructure and identity, and storage URIs supplied through configuration rather than application code. GPU availability, drivers, quotas, interruption behavior, and accelerator types remain cloud-specific concerns, while moving large datasets and checkpoints across clouds can create substantial egress costs, making data locality important. The piece contrasts managed Anyscale, self-managed KubeRay, cloud-native ML platforms such as SageMaker and Vertex AI, and an internal developer platform, concluding that the appropriate choice depends on Kubernetes expertise, operational needs, cloud commitments, compliance requirements, and the real need for multi-cloud deployment.
Oct 20, 2026
3,778 words in the original blog post.
OpenTofu and Terraform remain broadly compatible infrastructure-as-code tools in 2026, sharing HCL configurations and state files, with migrations between them generally requiring only a binary and CI change. The comparison argues that licensing is rarely the decisive issue for internal teams: Terraform’s BUSL-1.1 mainly restricts competing commercial services, while OpenTofu’s MPL-2.0 license and Linux Foundation governance appeal to organizations seeking vendor independence. Instead, the central challenge for small teams without dedicated infrastructure engineers is maintaining modules, provider upgrades, Kubernetes version lifecycles, state, drift controls, and policy processes. OpenTofu is presented as the default for teams willing to manage their own infrastructure code, while Terraform with HCP Terraform suits organizations requiring vendor support, Sentinel policies, or existing HashiCorp integrations. Pulumi is positioned for developers preferring general-purpose languages, Crossplane for Kubernetes-native operations, and Spacelift or env0 for managed execution, approvals, and policy controls around existing IaC. For teams lacking anyone to own cluster and module maintenance, the article recommends a BYOC internal developer platform such as Qovery, while retaining Terraform or OpenTofu for broader shared infrastructure such as landing zones, DNS, and data services.
Oct 20, 2026
4,098 words in the original blog post.
Migrating from Netlify to Kubernetes involves replacing several integrated services, including static hosting and CDN delivery, serverless functions, databases, TLS and DNS, deployment pipelines, and pull-request preview environments. The recommended approach is to migrate only when control, compliance, private networking, persistent databases, or long-running background workloads require infrastructure in an organization’s own cloud account, rather than for cost savings, since even a minimal managed Kubernetes setup can exceed Netlify pricing. The proposed sequence is to inventory all Netlify-specific configuration and add-ons, select a managed Kubernetes service and managed database, containerize backend functions into API services, workers, and scheduled jobs, migrate databases through replication with connection pooling, and move the frontend and DNS only after validation with low TTLs and rollback options. Static frontends can remain on Netlify while APIs and databases move to Kubernetes, which the article presents as a practical hybrid option for many smaller teams. It also emphasizes rebuilding developer-experience capabilities such as Git-based deployments, TLS, and preview environments through a DIY Kubernetes toolchain or managed platforms such as Qovery, while noting that such platforms reduce but do not eliminate operational responsibility.
Oct 20, 2026
4,159 words in the original blog post.
Qovery has launched a publicly accessible Console demo at console-demo.qovery.com that lets prospective users explore its full infrastructure-management interface without creating an account, providing payment details, or connecting cloud credentials. The demo uses fictional sample organizations and production-like data, including clusters, services, deployments, logs, metrics, audit records, and pull-request preview environments, while explaining how actions requiring a real cloud account would work. Visitors can explore independently or follow seven one-to-two-minute guided tours covering infrastructure provisioning, application deployment, preview environments, observability, Copilot deployment troubleshooting, incident-response automations, and security controls. The offering is aimed at product evaluation, whereas existing customers can use Qovery Training for structured onboarding, and users ready to deploy their own applications can begin a 14-day credit-card-free trial.
Oct 19, 2026
700 words in the original blog post.
EKS to AKS: How to Migrate Kubernetes Workloads When You Move to Azure for Better Enterprise Pricing
Migrating Kubernetes workloads from Amazon EKS to Azure AKS is presented as primarily a cloud-platform reconfiguration rather than a simple Kubernetes move: standard manifests and Helm charts are largely portable, but AWS-specific identity, networking, storage, registry, secrets, monitoring, and application integrations must be rebuilt for Azure. AKS is not inherently much cheaper than EKS beyond its potentially free control plane, with meaningful savings instead depending on negotiated Azure enterprise agreements, reserved capacity, savings plans, and licensing benefits, while stranded AWS commitments and the expense of operating both clouds during migration can offset early savings. The recommended approach is an eight-stage process of inventorying dependencies, establishing an Azure landing zone, adapting manifests, moving images and data, running EKS and AKS in parallel, shifting traffic through weighted DNS, and retaining EKS for two to four weeks as a rollback option before decommissioning it. Stateful applications, especially databases and single-writer volumes, determine the migration timeline because they require replication, dual-write strategies, validation, or planned downtime, whereas stateless workloads can move rapidly. A mid-sized migration is estimated at six to fourteen weeks, and the article argues that multi-cloud platform tools such as Qovery can simplify parallel deployment but do not eliminate data migration, managed-service replacement, or application code changes.
Oct 19, 2026
4,343 words in the original blog post.
Qovery Training is now generally available to all Qovery customers as a browser-based sandbox that replicates the Qovery Console using fake data, allowing users to deploy, roll back, delete environments, and troubleshoot issues without affecting cloud accounts or production systems. It offers 18 role- and task-focused tracks comprising 68 lessons for software engineers, DevOps, platform, and security teams, as well as migration paths from Heroku, Vercel, and Cloudflare to AWS. Users can configure the sandbox to resemble their own infrastructure, including cloud providers, frameworks, databases, queues, workers, and services, while imported configurations replace sensitive values with placeholders. Lessons combine guided tasks, automated completion checks, contextual help, a simulated terminal with common infrastructure tools, and an AI agent for exercises involving deployment analysis and service dependencies. The platform also includes deliberately broken troubleshooting scenarios and training spaces where administrators can select or create lessons, invite team members, and monitor progress, aiming to reduce onboarding time and dependence on internal Qovery experts before engineers receive production access.
Oct 19, 2026
1,627 words in the original blog post.
EU teams evaluating sovereign GPU capacity should distinguish between training and inference: both constitute personal-data processing under GDPR, but training uses a finite, auditable corpus while inference continuously generates prompts, outputs, embeddings, logs, feedback, and other potentially personal data. The article argues that paying more for EU-based or sovereign training is most justified when the corpus contains regulated, personal, or contractually restricted information, whereas public, licensed, synthetic, or genuinely anonymized data can often be trained on less expensive capacity elsewhere. It notes that EU-headquartered providers such as Scaleway and OVHcloud may offer H100 rates below US hyperscaler list prices, while the larger sovereignty premium often appears in specialized sovereign-control arrangements, limited availability, and delays accessing newest GPUs. “Sovereign” can range from an EU region run by a US hyperscaler to EU-controlled cloud offerings and EU-headquartered operators, with each providing different protections but none removing GDPR, security, contractual, or governance obligations. Because model weights, checkpoints, vector stores, and observability systems can also retain personal data, the article recommends assessing their residency and transfer risks alongside the GPU location. It presents a common split architecture in which data is sanitized in the EU, training occurs on the most suitable available capacity when permissible, and model artifacts, production inference, logs, backups, and retrieval systems remain in an EU region under the organization’s own cloud account.
Oct 19, 2026
4,621 words in the original blog post.
Aleph Alpha’s Kolibri 1 is presented as an Apache 2.0-licensed, open-weight 78.1-billion-parameter mixture-of-experts model that can be commercially self-hosted, with 3.46 billion parameters active per token and fp8 weights requiring roughly 78 GB of GPU memory before KV-cache overhead. The guide argues that organizations seeking European data sovereignty should distinguish EU-owned or EU-operated cloud providers and dedicated sovereign-cloud offerings from ordinary European regions of US hyperscalers, which it says may remain exposed to the US CLOUD Act. For production deployment, it recommends Kubernetes with GPU node pools, the NVIDIA GPU Operator, vLLM with tensor parallelism, persistent Hugging Face weight caching, shared-memory configuration, TLS, and API authentication, typically using two H100 GPUs or one H200 for the published fp8 model. It emphasizes that GPU inference economics depend chiefly on utilization and idle capacity, advising queue-based autoscaling, cached weights to reduce cold starts, scale-to-zero for nonproduction environments, and benchmarking of latency and throughput before capacity commitments. The article also positions Qovery as a bring-your-own-cloud platform that can manage Kubernetes operations in an organization’s own account while leaving model serving to vLLM, and concludes that self-hosting is most compelling for high-volume workloads or strict control and jurisdiction requirements rather than lightly used deployments.
Oct 19, 2026
4,803 words in the original blog post.
NVIDIA Nemotron 3.5 Super VL 120B is an open-weights, commercially usable multimodal mixture-of-experts model that processes text, images, and video, with 120 billion total parameters but 12 billion active per token and a context window of up to one million tokens. Running its BF16 version requires roughly 240 GB of VRAM for weights alone, plus memory for KV cache, vision encoding, and operational headroom, leading NVIDIA to recommend a single node with four H100 GPUs or two GB200 GPUs; the NVFP4 version reduces weight memory to about 60 GB but should be independently evaluated for visual-task accuracy. The article advises using a Kubernetes stack consisting of compatible NVIDIA GPU drivers and device plugins, a serving engine such as vLLM or NVIDIA NIM, a GPU-enabled Deployment configured for intra-node tensor parallelism, and an authenticated OpenAI-compatible endpoint. It emphasizes keeping all GPUs for a replica within one NVLink-connected node, staging large model files to avoid slow restarts, providing sufficient shared memory and startup time, monitoring queue depth and GPU cache metrics, and avoiding driver, CUDA, or inference-engine version mismatches. Self-hosting can support data residency, customization, predictable latency, and high sustained throughput, but its economics depend heavily on utilization because an eight-H100 cloud node can cost roughly $40,000 per month when continuously running; hosted APIs are generally more economical for low or bursty demand, while scale-to-zero pools, committed-use discounts, spot capacity, and quantization can reduce self-hosting costs.
Oct 19, 2026
4,526 words in the original blog post.
PSD2, its strong customer authentication rules, the EBA outsourcing guidelines, and DORA do not prohibit fintechs from hosting payment systems on public clouds such as AWS, Google Cloud, Azure, or Scaleway; instead, they require firms to retain accountability and demonstrate control over security, outsourcing, resilience, and exit arrangements. Strong customer authentication depends on two independent factors, dynamic linking of payments to a specific amount and payee, protected authentication data, and sound key-management and audit-log practices rather than on-premise infrastructure. DORA, effective from January 2025, adds requirements for ICT risk management, third-party provider registers, incident reporting, resilience testing, concentration-risk management, and documented, tested exit plans for critical services. The article argues that cloud portability is most difficult for proprietary managed services such as event streams, serverless runtimes, and identity systems, while containers, Kubernetes, and infrastructure-as-code can support more credible migration and recovery exercises. It presents bring-your-own-cloud approaches, including Qovery’s platform, as a way to keep contracts, data, encryption keys, logs, and operational control in the fintech’s own account, while emphasizing that such arrangements do not themselves deliver compliance or remove the firm’s responsibility for governance, reporting, testing, and oversight.
Oct 19, 2026
3,825 words in the original blog post.
Moving an application from a PaaS such as Heroku, Render, Railway, Fly.io, or App Engine to Google Kubernetes Engine primarily involves replacing the managed services the PaaS provided, including routing, TLS certificates, secrets, databases, migrations, logging, autoscaling, scheduled jobs, static egress, and rollback mechanisms, rather than changing the application container itself. The recommended approach for small teams without Kubernetes specialists is GKE Autopilot, which shifts node provisioning, patching, upgrades, and capacity management to Google, while Cloud SQL, Artifact Registry, Secret Manager, Cloud NAT, Gateway API, and managed observability fill other platform functions. A proposed migration sequence covers configuration inventory, image building, infrastructure provisioning, database migration, workload deployment, traffic setup, parallel operation, and delayed decommissioning, with database cutover identified as the principal irreversible step. A simple web application with PostgreSQL may take one to three weeks to migrate, while more complex systems can require one to three months, and the article argues that GKE generally becomes financially attractive only when PaaS costs exceed roughly $500 to $1,000 per month after engineering time is considered. It also notes that Cloud Run may be a simpler option for stateless HTTP services within its limits, while teams choosing GKE still need clear ownership for recurring work such as upgrades, certificate renewal, scaling tuning, secret rotation, backup testing, and incident response. Internal developer platforms, including the author’s company Qovery alongside alternatives such as Northflank and Porter, are presented as ways to retain self-service deployment workflows and reduce operational overhead while running workloads in a company’s own cloud account.
Oct 18, 2026
4,011 words in the original blog post.
Governed AI coding-agent access to Mistral Large 4 requires separate controls for model inference and cloud infrastructure: short-lived, scoped credentials rather than exposed API keys, audit trails that link model calls and deployments to a human owner, and hard spending limits that deny requests rather than merely issue alerts. The proposed four-layer approach combines an LLM gateway or secret broker for credential management and token budgets, cloud-provider identity and logging tools for inference governance, correlated session IDs across agent, model, and deployment logs, and infrastructure controls such as role-based access, approval gates, environment auto-stop, preview-environment expiration, and instance guardrails. The piece argues that hosting Mistral through Azure AI Foundry, Google Vertex AI, or AWS Bedrock can integrate inference with existing IAM, billing, and audit systems, although critical logs such as Bedrock invocation logs and Google Cloud Data Access logs must be enabled manually. It positions Qovery as an infrastructure-action layer that lets agents deploy into customer-owned AWS, GCP, Azure, Scaleway, or Kubernetes environments under per-environment permissions, with its MCP server read-only by default and write access requiring explicit configuration. The recommended implementation begins with read-only permissions, restricts writes to non-production environments, assigns one expiring credential per agent and project, tests revocation and spending controls, and combines credential-management products and model gateways with deployment governance to address both security and cloud-cost risks.
Oct 18, 2026
4,211 words in the original blog post.
Migrating a containerized application from Azure App Service to Azure Kubernetes Service (AKS) generally does not require changing the application image, but it transfers responsibility for operational functions such as upgrades, node management, ingress, TLS, secrets, autoscaling, monitoring, and rollback planning to the team or a managed platform layer. For small teams, the recommended approach is to first assess whether Azure Container Apps can meet requirements, since it avoids direct Kubernetes management and may be more suitable for fewer than roughly ten services without needs for Kubernetes APIs, Helm charts, advanced networking, or node-level control. A proposed AKS migration spans inventory and ownership assignment, cluster provisioning with security and monitoring, manifest creation, edge and DNS configuration, secrets migration to Key Vault, gradual weighted traffic shifting while retaining App Service as a rollback target, and establishing ongoing operational ownership. The discussion distinguishes tools such as kompose, which converts Docker Compose files into initial manifests, and Argo CD or Flux, which automate GitOps deployments but do not operate clusters, from internal developer platforms including Qovery, Porter, and Humanitec, which provide varying levels of self-service and cluster management. It argues that AKS’s principal cost is often engineering time rather than control-plane pricing, and advises avoiding big-bang cutovers, moving managed databases into clusters, relying on a single Kubernetes expert, and postponing observability or upgrade planning until after production traffic is moved.
Oct 18, 2026
3,754 words in the original blog post.
European fintech and healthtech teams seeking to run GPU inference without exporting regulated data must treat EU residency as an end-to-end data-path requirement, covering GPU compute, prompts, vector stores, object storage, backups, logs, traces, and telemetry rather than merely selecting an EU cloud region. The guide distinguishes data residency from legal sovereignty, noting that US-headquartered cloud providers may remain subject to the US CLOUD Act despite EU hosting, while GDPR transfers can still be lawful through DPAs, SCCs, or the EU-US Data Privacy Framework. It compares managed EU inference APIs, self-hosted models in a customer’s hyperscaler account, EU-headquartered GPU clouds, and on-premises deployments, balancing sovereignty, operational complexity, capacity, cost, and DORA exit requirements. It advises enforcing region restrictions and default-deny egress through cloud policies, maintaining in-region encryption keys and audit logs, documenting subprocessors and processing activities, and conducting DPIAs and reversibility tests where required. The guide also highlights scarce EU H100 and A100 capacity, the importance of early quota requests, GPU utilization and operational expenses in cost planning, and the potential role of bring-your-own-cloud deployment platforms such as Qovery in operating workloads while leaving infrastructure and billing under the customer’s control.
Oct 18, 2026
4,354 words in the original blog post.
Hetzner’s GEX63 dedicated server combines an NVIDIA RTX PRO 6000 Blackwell Max-Q GPU with 96 GB of VRAM, 128 GB of system memory, NVMe storage, and a flat €999 monthly price plus setup fee, positioning it as a potentially economical option for steady self-hosted AI inference. The guide recommends Ubuntu 24.04, a single-node K3s cluster, and NVIDIA’s GPU Operator to expose the GPU to Kubernetes, followed by deployment of vLLM or Ollama with persistent model caching, appropriate health probes, TLS, and authentication. Its VRAM can support approximately 32B-parameter models in BF16 or 70B-class models in 4-bit quantization, although KV-cache demands from long contexts and concurrent requests remain a primary capacity constraint. Because Blackwell uses CUDA compute capability sm_120, the server requires recent drivers, CUDA 12.8 or later, and PyTorch or vLLM builds containing sm_120 kernels; otherwise, containers may fail with a “no kernel image is available” error. The comparison argues that the server can cost less than comparable always-on cloud GPU rentals at high utilization, while cloud providers remain preferable for bursty workloads, interruptible batch processing, managed services, or redundancy needs. It also emphasizes operational responsibilities such as pinning drivers and kernels, monitoring GPU health and ECC or Xid errors, handling out-of-memory failures, securing public endpoints, and planning for the single node as a point of failure, while presenting Qovery as an optional deployment platform for adding self-service workflows to a self-managed cluster.
Oct 17, 2026
5,276 words in the original blog post.
A unified Kubernetes policy set and CI/CD workflow across AWS, GCP, Azure, and private clusters is presented as achievable when teams set the portability boundary at the Kubernetes API: manifests, RBAC, admission policies, rollout strategies, and observability can be shared, while cloud-specific concerns such as IAM, load balancers, storage classes, node pools, networking, and managed databases should be isolated in per-cluster configuration overlays. The proposed architecture uses a common policy engine such as Kyverno or OPA Gatekeeper, Pod Security Admission as a baseline, pull-based GitOps through Argo CD or Flux, and a cloud-neutral CI process that builds, scans, signs, and promotes immutable image digests without retaining cluster credentials. Argo CD ApplicationSets or equivalent mechanisms can distribute one application definition to labeled clusters, while External Secrets Operator and per-cluster SecretStores provide a common secrets interface across cloud-native secret services. To limit operational drift, the approach recommends fleet-wide Kubernetes version targets, centrally managed add-on bundles, consistent cluster labels, staged policy enforcement, and shared infrastructure abstractions through Terraform wrapper modules, Cluster API, or Crossplane. It argues that the ongoing burden of upgrades, compatibility testing, policy maintenance, and platform operations can outweigh initial implementation work, making managed internal developer platforms such as Qovery a potential alternative for organizations without sufficient dedicated platform engineering capacity.
Oct 17, 2026
4,822 words in the original blog post.
For teams running Kubernetes without a dedicated platform function, GKE Autopilot is presented as the simplest managed-cluster option because Google handles nodes, scaling, patching, and automatic upgrades while billing primarily by pod resource requests, whereas EKS Auto Mode offers comparable AWS automation but still requires more work around VPC design, IAM, and version planning. Standard GKE and classic EKS managed node groups provide greater control but demand significantly more Kubernetes and cloud-operational expertise, including node management, add-on maintenance, upgrades, and incident response. The comparison argues that managed control planes address only part of the challenge, with developer self-service, CI/CD, preview environments, secrets, RBAC, networking, and cost visibility representing the larger operational burden. It recommends choosing the cloud where a team already has expertise, credentials, and spending commitments, then adding a platform layer such as Qovery, Porter, or Northflank for self-service deployments in a company-owned cloud account. For simpler stateless HTTP services, Cloud Run or AWS App Runner may avoid Kubernetes entirely, while Kubernetes becomes more appropriate for needs such as background jobs, stateful workloads, custom networking, or compliance controls. Costs depend on cluster fees, pod-versus-node billing, cloud commitments, utilization, and network egress, rather than on a universally cheaper choice between AWS and GCP.
Oct 17, 2026
4,789 words in the original blog post.
PCI DSS v4.0.1 requires fintechs moving cardholder-data processing to AWS, Google Cloud, Azure, Scaleway, or Kubernetes to treat their deployed workloads as a separately assessed cardholder data environment rather than inheriting cloud-provider compliance. Providers’ attestations cover only specified infrastructure services and regions, while the fintech must document shared responsibilities, inventory and monitor third parties, maintain current data-flow and network diagrams, and prove scope and segmentation, with service providers facing several six-month review and testing cycles. Core obligations include making stored PAN unreadable, encrypting data in transit, prohibiting post-authorization storage of sensitive authentication data, enforcing MFA and least-privilege access, retaining protected audit logs for 12 months with recent logs readily accessible, and conducting quarterly vulnerability scans alongside regular penetration and segmentation tests. The article emphasizes reducing scope by using hosted payment pages, tokenization, and strong isolation rather than attempting to harden an entire shared environment, especially because Kubernetes namespaces alone are not considered sufficient segmentation. It also explains that platforms such as Qovery may operate Kubernetes in a customer’s own cloud account, preserving customer control over logs, keys, and network boundaries, but do not confer PCI compliance, which remains subject to assessment by a QSA, acquirer, or card brand.
Oct 17, 2026
4,510 words in the original blog post.
Migrating from DigitalOcean App Platform to Kubernetes can preserve existing Git-based development workflows because source triggers, CI tests, container builds, image registries, and rollback practices generally remain unchanged, while only the deployment step must be replaced. The recommended approach is to export the App Platform specification and translate its services, workers, jobs, environment variables, routing, health checks, and scaling settings into Kubernetes resources such as Deployments, Services, Ingresses, Jobs, Secrets, and ConfigMaps, while leaving managed databases outside the cluster. Teams are advised to migrate incrementally by establishing cluster infrastructure, ingress, TLS, secret management, and observability, moving low-risk stateless services first, operating both platforms in parallel, gradually shifting traffic, and migrating workers and scheduled jobs last. Kubernetes introduces operational responsibilities previously managed by App Platform, including certificate renewal, autoscaling configuration, resource sizing, logs and metrics, cluster upgrades, security patching, and preview-environment management. DOKS offers the least disruptive destination for organizations remaining on DigitalOcean, while EKS, GKE, AKS, Scaleway, and self-managed clusters may suit broader cloud or infrastructure requirements. The comparison of tools presents Kustomize and Argo CD as high-control GitOps options, Backstage as a portal requiring complementary deployment tooling, Humanitec and Port as platform-engineering products, and Qovery as a managed internal developer platform intended to retain git-push deployments and pull-request previews on Kubernetes.
Oct 17, 2026
4,037 words in the original blog post.
LightOnOCR-3 is an Apache 2.0-licensed family of open-weight document OCR vision-language models from LightOn that can be self-hosted on Kubernetes rather than accessed through per-page managed APIs such as AWS Textract, Google Document AI, Azure AI Document Intelligence, or Mistral OCR. The guide explains that the 0.8B and 1B variants can run on a single 24 GB NVIDIA L4 or A10-class GPU and recommends serving the model through vLLM’s OpenAI-compatible API, with persistent Hugging Face cache storage, GPU resource limits, node-taint tolerations, and startup and readiness probes to support reliable Kubernetes deployment. It notes that page resolution, batching, concurrency, CPU capacity for PDF rasterization, and GPU VRAM management materially affect throughput and cost. Self-hosting may be economically favorable at predictable, high volumes, with an estimated break-even near 390,000 pages monthly compared with low-cost plain OCR APIs but closer to 12,000 pages for expensive form and layout extraction services, although infrastructure and engineering costs must also be considered. Managed APIs remain more suitable for low or uneven demand, vendor SLAs, and specialized prebuilt parsers, while self-hosting offers greater control over prompts, capacity, data residency, and compliance controls by keeping documents within an organization’s own cloud environment or on-premises cluster.
Oct 16, 2026
4,398 words in the original blog post.
Vercel Ship 2026 highlighted the company’s expansion from frontend hosting toward broader application infrastructure, with general availability for Python, container support, Vercel Services, and Daemons for long-running workers, alongside new AI-agent tools such as Connect, Agent, Drives, and AI Gateway controls. The post argues that governance has become central as Vercel reports that agent-triggered deployments now account for more than half of deployments, prompting greater emphasis on scoped credentials, approvals, budgets, identity controls, and auditability. It focuses particularly on Vercel Bring Your Own Cloud, now generally available for Enterprise customers on AWS, which runs Node.js and Python Vercel Functions as AWS Lambda workloads in a customer account but retains routing, TLS termination, static-asset delivery, and request transit on Vercel infrastructure; it does not support containers, Services, Edge Runtime, other cloud providers, or FedRAMP requirements. The author contrasts this limited BYOC model with Qovery’s platform, which it says deploys full stacks—including applications, containers, workers, databases, and AI agents—within customer-controlled cloud accounts or Kubernetes clusters across AWS, GCP, Azure, and Scaleway, while acknowledging that Vercel remains well suited to CDN-focused Next.js frontends and that organizations may combine Vercel for frontend delivery with another platform for backend workloads.
Oct 16, 2026
3,172 words in the original blog post.
Deploying the same containerized application stack to AWS and GCP is best achieved through layered architecture rather than trying to make the clouds identical: Kubernetes and Helm provide a portable application layer, while OpenTofu, Terraform, Pulumi, or Crossplane manage cloud-specific infrastructure through a shared codebase and per-cloud providers. The main differences can usually be limited to parameters for identity, ingress, storage, databases, secrets, registries, DNS, and node types, although proprietary managed data services such as AWS Aurora and Google AlloyDB remain difficult to standardize. Kubernetes and Helm are suited to application runtime and packaging, IaC tools address resource provisioning, Crossplane offers continuous Kubernetes-native reconciliation for mature platform teams, and Ansible is better for configuration and sequencing than modeling cloud infrastructure. Teams without dedicated platform engineering resources may use an internal developer platform such as Qovery to deploy a single environment definition into their own AWS or GCP accounts, while retaining Terraform or OpenTofu for organizational networking and other non-container infrastructure. The recommended approach is to maintain one source of truth, test both cloud targets continuously, avoid cross-cloud traffic, and account for separate operational responsibilities such as Kubernetes upgrade schedules and cloud-specific data-service capabilities.
Oct 16, 2026
3,943 words in the original blog post.
Cloud IAM and tagging governance is presented as a combination of preventive controls that block noncompliant resources, corrective tools that remediate drift after creation, and generative platforms that produce compliant infrastructure by default, with the central challenge being the ongoing effort required to author and maintain policies. Native AWS, Azure, and GCP organizational controls provide a free baseline but have coverage and workflow limitations, while Turbot Guardrails is positioned for large, audited multi-cloud estates, Cloud Custodian for engineer-managed corrective automation, and OPA/Gatekeeper or Kyverno for Kubernetes-focused preventive controls. StackGen and ControlMonkey embed governance in Terraform workflows, whereas CloudBolt, Flexera One, and CloudCheckr emphasize FinOps, tagging, inventory, and chargeback rather than IAM prevention. The author, Qovery’s co-founder, describes Qovery as a low-operations internal developer platform that applies IAM scoping, RBAC, and tags to resources it creates but does not audit an entire existing estate, and recommends combining native controls, a compliant developer deployment path or IaC generator, and drift-detection or remediation tools such as Cloud Custodian or Steampipe.
Oct 16, 2026
3,924 words in the original blog post.
AWS Transform is presented as an agentic AI service for modernizing .NET, mainframe, VMware, and selected custom workloads, with transformation paths limited to AWS and coverage focused primarily on code transformation and workload migration rather than estate assessment or post-migration operations. The comparison frames cloud modernization as four distinct stages—assessment, transformation, migration, and ongoing operation—and argues that alternatives should be selected according to the organization’s main constraint. CAST is positioned for large-scale application portfolio analysis, Flexera One for license management and multi-cloud cost governance, Azure Migrate and GitHub Copilot for Azure-focused Microsoft estates, Google Cloud Migration Center and Migrate to Containers for GKE targets, IBM watsonx Code Assistant for Z and OpenText for COBOL modernization or re-platforming, and Modelcode for large-scale AI-assisted code translation. Qovery is described as a complementary internal developer platform for operating migrated applications in customer-owned AWS, Azure, GCP, Scaleway, or existing Kubernetes environments, rather than assessing portfolios or transforming legacy code. The article emphasizes that tool licensing is only part of migration costs, with cloud spending, licensing obligations, and the staffing required to operate Kubernetes often becoming more consequential after cutover.
Oct 16, 2026
3,679 words in the original blog post.
NIS2 does not prohibit regulated workloads from running on hyperscalers such as AWS, Azure, or Google Cloud; instead, it requires covered entities to manage and demonstrate cybersecurity controls for the systems they configure and operate above the provider’s infrastructure. Articles 20 and 21 emphasize management accountability and risk-management measures including access controls, MFA, encryption, logging, incident response, backups, recovery testing, secure development, vulnerability management, environment segmentation, and supply-chain oversight, while Article 23 requires early warning of significant incidents within 24 hours, notification within 72 hours, and a final report within one month. The article argues that a cloud provider’s shared-responsibility model leaves customers responsible for identities, network configuration, keys, deployment processes, audit evidence, and much of Kubernetes security. It distinguishes NIS2 from GDPR, DORA, and national sovereignty schemes, noting that data-localization or sovereign-cloud obligations generally arise from other rules rather than NIS2 itself. It also presents bring-your-own-cloud platforms as a model that can retain data, keys, logs, and cloud-account control within the regulated entity while requiring assessment of the platform vendor as part of the supply chain, and promotes Qovery’s BYOC deployment approach as supporting these objectives.
Oct 16, 2026
4,117 words in the original blog post.
Kubernetes self-service deployments can be implemented without giving developers kubeconfigs by using Git as the only developer write interface, a pull-based GitOps controller such as Argo CD or Flux as the cluster’s sole write identity, and admission controls such as Kyverno, OPA Gatekeeper, Pod Security Admission, ResourceQuota, and LimitRange to enforce security and resource rules. The proposed model adds developer abstractions like constrained Helm charts, Crossplane XRDs, or Score specifications, self-service portals such as Backstage, Port, Humanitec, or Qovery, and tenant isolation through namespaces, RBAC, quotas, vCluster, or separate clusters. It argues that Git review alone cannot prevent unsafe manifests, so policies and schema-constrained templates are needed to block privileged workloads, excessive resource requests, untrusted images, and cross-namespace deployments while still providing developers scoped visibility into logs, events, rollout status, restarts, and rollbacks. Building this stack internally can require sustained platform-team ownership for tooling, policies, secrets, previews, and Kubernetes upgrades, whereas commercial platforms may reduce operational work. The article presents Qovery as a managed, bring-your-own-cloud option that combines deployment workflows, per-environment permissions, preview environments, and cluster operations, while acknowledging that custom platform APIs, bespoke operators, and complex tenancy needs may favor tools such as Crossplane, Backstage, Argo CD, Kyverno, and vCluster.
Oct 15, 2026
4,459 words in the original blog post.
For software vendors selling to French public-sector buyers in 2026, SecNumCloud qualification is decisive when a tender explicitly cites ANSSI qualification or the cloud de confiance doctrine, because it applies to specific offerings and scopes rather than entire providers and includes security, EU operational control, and protection from extraterritorial laws such as the US CLOUD Act. The shortlisted options are OVHcloud, 3DS Outscale, Cloud Temple, NumSpot, S3NS, Orange Business, and AWS European Sovereign Cloud, although only the first six have cited SecNumCloud-qualified offerings; AWS’s separate EU-operated cloud in Germany may meet requirements for EU residency and operational control but is not accepted for tenders requiring SecNumCloud. OVHcloud, Outscale, and Cloud Temple are positioned for established qualified infrastructure, while NumSpot, S3NS, and Orange Business provide newer trusted-cloud offerings with varying degrees of managed-service, Kubernetes, GPU, and PaaS support. The practical deployment cost depends chiefly on application architecture: Kubernetes-based applications using portable components can generally be retargeted relatively quickly, whereas applications reliant on proprietary hyperscaler services may require significant redesign. Qovery is presented as a non-qualified deployment and portability layer that can deploy workloads into customer-owned accounts or Kubernetes clusters, including those hosted by qualified providers, but compliance remains dependent on the underlying cloud offering.
Oct 15, 2026
4,239 words in the original blog post.
AWS Proton reached end of support on October 7, 2026, ending its control plane, console, API, template registry, and self-service abstractions while leaving already provisioned CloudFormation stacks, Terraform state, workloads, IAM roles, and other resources in customers’ AWS accounts operational. The proposed migration approach begins by exporting template bundles and inventorying environments, service templates, instances, components, and provisioning modes, then moving underlying infrastructure definitions into Git and importing existing resources rather than recreating them. It identifies five main destinations: direct infrastructure-as-code and CI workflows, AWS-native combinations such as CodePipeline and Service Catalog, self-built developer portals using Backstage or Port with Crossplane or Terraform, managed internal developer platforms including Qovery, Humanitec, and Northflank, and Kubernetes GitOps tools such as Argo CD or Flux. The recommended choice depends primarily on available platform-engineering capacity, with managed platforms presented as more suitable for small teams and custom platforms as more viable for larger teams. A three-stage, roughly 10-week plan for a 30-service estate recommends freezing new Proton template versions, re-owning and validating infrastructure code, and replacing the developer self-service layer last, while running old and new paths in parallel to reduce production risk.
Oct 15, 2026
3,850 words in the original blog post.
For teams self-hosting open-weight LLMs in 2026, the comparison argues that vLLM is generally the best default because its open licensing, broad accelerator support, continuous batching, PagedAttention, quantization, and prefix caching can reduce cost per token, while NVIDIA NIM is better suited to organizations already paying for NVIDIA AI Enterprise and seeking supported, pre-optimized deployments rather than the lowest GPU cost. Hugging Face TGI remains a practical production option for teams embedded in the Hugging Face ecosystem, SGLang is positioned as especially competitive for agent and RAG workloads with shared prompt prefixes, and TensorRT-LLM may deliver strong NVIDIA-specific performance but requires substantial engineering effort for compilation and tuning. The central cost metric is sustained tokens per hour at a defined latency target, since GPU cost per million tokens depends on hourly instance price divided by usable throughput, but the article stresses that utilization, right-sizing, autoscaling, quantization, and shutting down idle non-production GPUs often affect total spending more than the serving engine itself. It also notes that self-hosting becomes more economical primarily at predictable, high utilization, whereas hosted APIs are often cheaper and simpler for low-volume or bursty demand, and it presents orchestration tools such as Ray Serve and KServe, as well as Qovery’s managed deployment layer, as operational complements rather than inference engines.
Oct 15, 2026
4,089 words in the original blog post.
Consistent compliance across AWS, GCP, and sovereign EU clouds can be achieved by defining controls at the shared Kubernetes and delivery-pipeline layer rather than separately in each provider’s console, using a single framework such as ISO/IEC 27001:2022 or SOC 2 as a control spine. The approach centers on a version-controlled control-to-cloud matrix, standardized RBAC and admission policies, and one production pipeline that records approvals, deployments, identities, environments, and target clouds in a unified audit trail. Auditors typically require evidence of access controls, approved changes, data residency, and continuous operation throughout the audit period, while cloud-provider certifications cover only the provider’s infrastructure responsibilities rather than customers’ configurations, identities, data, and applications. Data residency should be demonstrated through region-pinned configuration, preventive policies, and live inventories that also account for backups, logs, CI systems, and third-party services, while residency, operational sovereignty, and jurisdictional exposure should be treated as distinct concepts. The text recommends combining provider-native guardrails, Kubernetes-native preventive controls, centralized immutable logging, CSPM tools for detection, and compliance-automation platforms for framework mapping; it presents Qovery as a platform intended to standardize deployment, RBAC, and Git-based change records across supported cloud accounts and Kubernetes clusters.
Oct 15, 2026
4,110 words in the original blog post.
Migrating a Spring Boot application from Cloud Foundry to Kubernetes is presented as primarily a platform-contract change rather than a rewrite of Java business logic, replacing CF buildpacks, VCAP_SERVICES, routes, scaling, logging, and health checks with container images, Kubernetes Secrets and ConfigMaps, Ingress or Gateway API routing, deployments and autoscaling, and stdout plus Actuator probes. The recommended approach is to remove Cloud Foundry-specific integrations such as spring-cloud-connectors, VCAP_SERVICES parsing, cloud profiles, and reliance on local storage or sticky sessions, then use standard Spring Boot environment-based properties and externalize state through managed databases, Redis, object storage, or persistent volumes. Paketo Buildpacks can preserve a Dockerfile-free build workflow, while careful JVM memory tuning, startup probes, graceful shutdown configuration, resource quotas, secrets management, observability, and traffic cutover planning are emphasized as major operational responsibilities previously handled by Cloud Foundry. The guide advises phased migrations that keep both platforms active until validation is complete, noting that individual services may move quickly but larger estates require substantial work on shared infrastructure, data migration, routing, security, and deployment practices. It outlines several target models, including raw Kubernetes, Korifi for CF-style workflows, commercial platforms such as Tanzu or OpenShift, and internal developer platforms such as Qovery, which is promoted as a way to retain push-to-deploy workflows, preview environments, and self-service services on Kubernetes.
Oct 14, 2026
4,244 words in the original blog post.
Reflection AI’s announced Beam is a 501-billion-parameter sparse Mixture-of-Experts model with 23 billion active parameters per token, planned for release under the permissive Apache 2.0 license, enabling commercial self-hosting, modification, and redistribution once weights become available. Although sparsity reduces inference computation, all experts must remain in memory, requiring roughly 1 TB of VRAM at BF16, 500 GB at FP8, or 250 GB at 4-bit quantization before cache and runtime overhead, making an 8xH200- or MI300X-class node a practical FP8 baseline and multi-node deployments likely at higher precision or long contexts. The discussion compares cloud hardware options across AWS, GCP, Azure, Scaleway, and neocloud providers, emphasizing GPU availability, high-speed interconnects, storage for large checkpoints, and the performance advantages of keeping parallelism within a single NVLink-connected node. It recommends production engines such as vLLM, SGLang, or TensorRT-LLM, with Kubernetes-based deployments requiring careful management of GPU drivers, networking, autoscaling, weight caching, observability, and upgrades. Self-hosting can lower per-token costs only under sustained high utilization, while hosted APIs are generally more economical for variable or low-volume demand; data residency, regulatory constraints, proprietary fine-tuning, and version control may nevertheless justify operating models in a company’s own cloud account.
Oct 14, 2026
5,030 words in the original blog post.
Maintaining cloud-pricing leverage does not necessarily require duplicated production environments; instead, the proposed “credible leverage” model uses a primary cloud alongside a small, always-on secondary footprint for staging, preview environments, and a reference workload, supported by documented and regularly tested exit procedures. Portability depends chiefly on standardized container images, Kubernetes, infrastructure-as-code tools such as OpenTofu or Terraform, portable data choices such as Postgres or MySQL, and quarterly deployment-and-restore drills that provide measurable evidence for procurement negotiations. The main operational challenges are identity, networking, egress, managed data dependencies, observability, compliance, and staffing, while proprietary databases, serverless services, and cross-cloud data transfer create the strongest forms of lock-in. The piece distinguishes infrastructure provisioning, Kubernetes fleet management, developer deployment platforms, and GitOps tools as complementary layers, comparing products including Qovery, Rafay, SUSE Rancher, Red Hat ACM, GKE Enterprise, Northflank, and CloudBolt according to their roles. It argues that a single deployment control plane can reduce duplicated workflows by sharing pipelines, access controls, and environment definitions across clouds, while noting that multi-cloud is financially unsuitable when its added cost exceeds likely discounts, particularly for smaller organizations or teams reliant on proprietary cloud services.
Oct 14, 2026
3,728 words in the original blog post.
SOC 2, HIPAA, and internal audit requirements largely overlap around access controls, encryption, logging, change management, incident response, and vendor oversight, allowing organizations migrating legacy insurance or healthcare-adjacent applications to build one control matrix and evidence pipeline rather than separate compliance programs. The frameworks differ primarily in their assessment models and specific obligations: SOC 2 Type II evaluates controls over time, HIPAA has no formal certification but requires safeguards and Business Associate Agreements for PHI-related services, and internal audits emphasize approvals, segregation of duties, and insurer-specific obligations. The discussion identifies common migration risks including manual approvals, shared privileged credentials, inadequate audit trails, and production PHI appearing in non-production environments, recommending synthetic data, automated governance, immutable deployment histories, and identity-based access controls. It argues that a compliant cloud stack typically combines a cloud provider, infrastructure-as-code tools such as Terraform or OpenTofu, a deployment governance platform, compliance-evidence automation, and, where expertise is limited, a managed cloud partner. The proposed migration sequence begins with controls and a cloud landing zone, validates evidence generation through a low-risk non-PHI service, then moves regulated workloads while allowing sufficient time for the SOC 2 observation period. Qovery is presented as one possible deployment and environment-governance layer that operates within a customer’s own cloud account, complementing rather than replacing auditors, compliance tools, infrastructure-as-code, or managed security providers.
Oct 14, 2026
4,240 words in the original blog post.
For teams without dedicated infrastructure engineers, the recommended starting point is usually a managed PaaS or a bring-your-own-cloud (BYOC) internal developer platform, adding infrastructure-as-code (IaC) only for resources those platforms cannot manage. Terraform, OpenTofu, and Pulumi offer broad control but require ongoing ownership of state, modules, provider upgrades, secrets, drift detection, and CI/CD processes, creating a maintenance burden that can distract small teams from product work. Terraform and its MPL-2.0 fork OpenTofu are positioned as strong choices for teams seeking the largest provider ecosystem, hiring pool, and library of examples, while Pulumi may suit teams already using languages such as TypeScript, Python, Go, or C# because it supports familiar development and testing practices and can use Terraform providers. Managed PaaS products such as Vercel, Render, Railway, and Heroku can enable fast deployment for conventional web applications but place workloads in vendor accounts, limiting access to cloud credits, committed-use discounts, direct network integration, and some compliance options. BYOC platforms such as Qovery, Porter, and Northflank aim to provide PaaS-style deployment workflows while retaining workloads, billing, cloud discounts, and compliance boundaries within the customer’s own account, though they still require IaC for bespoke networking, identity, DNS, and shared data infrastructure. The article argues that mature organizations often use a hybrid model in which platforms manage repeatable application deployments and IaC manages custom foundational infrastructure, with the best choice depending on infrastructure ownership, workload complexity, compliance needs, portability requirements, and existing engineering skills.
Oct 14, 2026
3,742 words in the original blog post.
European sovereign cloud options differ substantially in ownership, certifications, service compatibility, and pricing transparency: AWS European Sovereign Cloud is operated by an EU-controlled AWS entity and publishes calculator pricing, STACKIT publishes list prices for its OpenStack-based platform, while Microsoft-based Bleu and Google Cloud-based S3NS generally require custom quotes. S3NS had obtained SecNumCloud 3.2 qualification by December 2025, whereas Bleu was still pursuing qualification in late 2026; AWS offers a separate EU-controlled control plane, and STACKIT operates German and Austrian regions but lacks hyperscaler API compatibility. Available comparisons indicate AWS’s sovereign region costs roughly 15% more than AWS Frankfurt for several services, while STACKIT’s comparable compute pricing can meet or undercut AWS list prices, although smaller service catalogs, limited availability-zone depth, onboarding requirements, support fees, egress, disaster recovery design, and added platform engineering may have a greater effect on total cost than compute rates. The account argues that standard EU cloud regions with residency controls are usually sufficient for GDPR-focused workloads, while SecNumCloud or similar sovereign offerings are most justified by explicit public-sector, critical-infrastructure, regulatory, or contractual requirements. It recommends verifying certification status, discount eligibility, service roadmaps, data-export terms, and support costs directly with vendors, while using portable architectures, account ownership, right-sizing, and selective hybrid deployment to manage cost and reduce lock-in.
Oct 13, 2026
2,963 words in the original blog post.
EU AI Act rules for general-purpose AI models became enforceable by the European Commission on 2 August 2026, exposing GPAI providers to fines of up to 3% of worldwide turnover or €15 million, while most organizations that internally self-host unmodified Qwen, DeepSeek, or Mixtral models remain deployers with narrower duties involving staff AI literacy, user transparency, and retention of model licences and cards. Provider obligations arise when organizations market models under their own name or substantially modify them, requiring technical documentation, downstream information, copyright and text-and-data-mining policies, training-data summaries, and, for systemic-risk models, evaluations, incident reporting, and cybersecurity controls. The central operational recommendation is to build compliance evidence into infrastructure from the outset by pinning model versions and hashes, archiving provenance, logging prompts and outputs under a defined retention policy, attributing usage by team, and proving the physical inference region. It describes a self-hosted production stack built around runtimes such as vLLM or SGLang, Kubernetes-based orchestration and autoscaling, and gateways such as LiteLLM, arguing that queue depth, concurrency, KV-cache pressure, and time-to-first-token are better scaling signals than GPU utilization. GPU capacity planning must account for total model parameters, KV-cache demands, context length, and concurrency, making large mixture-of-experts and DeepSeek models multi-GPU workloads despite lower active-parameter counts. Self-hosting can improve data residency, auditability, version control, and access to cloud-account discounts, but becomes cost-effective relative to hosted APIs mainly at sustained high utilization; the text presents Qovery as a BYOC platform layer for deployment, environments, and access control rather than an inference runtime or compliance solution.
Oct 13, 2026
8,304 words in the original blog post.
Mistral AI’s reported €3 billion Series D funding round and release of Mistral Large 4 are presented as strengthening Europe’s ability to deploy frontier-class AI under EU jurisdiction, particularly through Mistral’s EU-hosted La Plateforme or self-hosted open weights rather than US hyperscaler marketplaces. The piece argues that using Mistral through Amazon Bedrock, Azure AI Foundry, or Google Vertex AI may provide model choice but does not eliminate potential US CLOUD Act exposure, whereas EU-owned providers or customer-controlled hardware offer higher levels of legal or technical sovereignty. It describes Mistral Large 4 as a multimodal mixture-of-experts model with an OpenAI-compatible API, positioning it as suitable for many business uses such as retrieval, extraction, summarization, classification, coding assistance, and tool-calling agents, while advising organizations to evaluate it against their own accuracy, latency, and cost requirements. The article identifies OVHcloud, Scaleway, Open Telekom Cloud, IONOS, and Exoscale as viable destinations for containerized applications built with Kubernetes, Postgres, and S3-compatible storage, estimating that such migrations can take 6 to 16 weeks, but stresses that AWS-specific services such as Lambda, DynamoDB, Step Functions, Cognito, CloudFormation, and IAM policies require substantial rebuilding. It recommends adopting portable infrastructure components, maintaining tested exit plans for regulatory resilience, accounting for operational and migration costs, and using platform tooling such as Qovery or alternatives to preserve developer workflows across cloud environments.
Oct 13, 2026
7,235 words in the original blog post.
Five-person engineering teams can operate production infrastructure without a dedicated DevOps hire by adopting platforms that automate provisioning, deployments, TLS, secrets, logging, rollbacks, scaling, and upgrades. The comparison divides options into managed PaaS platforms such as Vercel, Render, Railway, and Fly.io, which offer the fastest setup but run workloads in vendor accounts; self-hosted tools such as Coolify and Dokploy, which are inexpensive but require teams to maintain servers and respond to incidents; and BYOC platforms including Qovery, Convox, and Kuberns, which provide self-service deployment workflows while keeping workloads in the customer’s cloud account. Vercel is presented as especially suitable for Next.js and frontend applications, Render and Railway for small full-stack products, Fly.io for globally distributed low-latency workloads, and Coolify or Dokploy for low-cost side projects and internal tools. The article argues that BYOC becomes more attractive when teams need compliance controls, VPC isolation, cloud credits, committed-spend discounts, or greater control over data and infrastructure, while emphasizing that total cost includes cloud spending and residual engineering effort rather than subscription fees alone. It recommends evaluating candidates through real deployments, pull-request preview environments, rollback and backup tests, visibility into logs and metrics, portability, and security and access-control capabilities.
Oct 12, 2026
3,104 words in the original blog post.
What Does Kubernetes On-Premise Really Cost in 2026? A Line-by-Line Comparison With EKS, GKE and AKS
Running Kubernetes on-premise involves far more than server purchases, including hardware amortization, colocation, power, redundant networking and storage, software subscriptions, disaster recovery, refresh cycles, and 2–5 platform or SRE engineers for continuous operations. Managed Kubernetes services such as Amazon EKS, Google GKE, and Azure AKS charge roughly $0.10 per cluster-hour for their control planes, but their primary costs come from compute, storage, load balancers, NAT gateways, data transfer, and operational staff. The comparison depends chiefly on utilization and workload characteristics: on-premise infrastructure may become less expensive for stable workloads operating above roughly 60–70% utilization for several years, particularly for egress-heavy, GPU-intensive, or data-residency-constrained workloads, while cloud services generally favor bursty, short-lived, geographically distributed, and development workloads because capacity can be scaled down. The discussion cautions against comparing cloud list prices only with server costs, noting that both approaches can incur substantial waste through idle infrastructure, overprovisioning, or unmanaged cloud resources. Hybrid strategies can place predictable workloads on owned or bare-metal capacity while retaining cloud elasticity, although maintaining separate tools and operating models can add complexity. It recommends reducing costs in either environment through standardized resource sizing, automated shutdown of nonproduction systems, ephemeral preview environments, autoscaling, cost visibility tools, and unified deployment and access-control workflows; Qovery is presented as a platform intended to provide such workflows across cloud and on-premise Kubernetes environments.
Oct 12, 2026
3,537 words in the original blog post.
Migrating applications from managed platforms such as Heroku, Render, Vercel, Fly.io, or App Engine to Kubernetes requires coordinating four distinct layers rather than adopting a single migration tool: infrastructure provisioning, data and state transfer, GitOps-based deployment, and developer self-service. Terraform or OpenTofu are presented as the default choices for provisioning clusters, while Crossplane suits teams that want infrastructure managed through Kubernetes APIs; databases require separate migration methods such as pg_dump, logical replication, or AWS DMS. Velero is useful only for moving workloads between existing Kubernetes clusters, not for PaaS-to-Kubernetes migrations because PaaS sources have no Kubernetes state to back up. Argo CD and Flux provide continuous GitOps reconciliation but do not replace features such as push-to-deploy workflows, preview environments, rollback controls, secrets management, and managed add-ons, which often require an internal developer platform, a custom Backstage-based portal, or commercial self-service tools. Moving Next.js applications from Vercel adds further work around standalone builds, distributed ISR caching, image optimization, edge functionality, CDN configuration, and build-time environment variables. A typical migration proceeds through application-parity discovery, cluster and networking setup, parallel deployment of stateless services, rehearsed data cutover, and eventual decommissioning, taking roughly four to eight weeks for small estates and longer for complex or regulated systems. Although Kubernetes can reduce infrastructure costs at scale, savings depend on controlling idle capacity, networking charges, operational complexity, and the potentially substantial cost of platform engineering staff.
Oct 12, 2026
4,297 words in the original blog post.
Spot instances can reduce compute costs substantially but require workloads and infrastructure designed to withstand abrupt node reclamation, with warning periods varying from about two minutes on AWS to roughly 30 seconds on GCP and Azure. Suitable candidates include replicated stateless services, asynchronous workers, CI runners, retryable batch jobs, and non-production environments, while primary databases, stateful singletons, uncheckpointed long-running jobs, and license-bound software should remain on on-demand or reserved capacity. The recommended production model combines a diversified spot fleet across instance types and availability zones with an on-demand baseline for critical system and stateful components, plus automatic fallback when spot capacity is unavailable. Kubernetes safeguards such as multiple replicas, PodDisruptionBudgets, topology spreading, termination handling, readiness probes, graceful SIGTERM shutdowns, workload scheduling controls, and interruption-aware node draining are presented as essential. The piece compares Cluster Autoscaler, Karpenter, cloud-native offerings, and commercial tools such as Spot Ocean, Cast AI, and Qovery, noting trade-offs between operational effort, automation, cloud support, and pricing. It advises a staged rollout beginning with non-production workloads and progressing through async jobs to carefully monitored stateless production services, while emphasizing that spot savings depend on compute’s share of total spending and should complement rightsizing, auto-stop policies, and reserved-capacity discounts.
Oct 11, 2026
3,257 words in the original blog post.
Organizations leaving VMware due to Broadcom’s subscription and licensing changes should treat the transition primarily as a deadline-driven migration problem, choosing rehosting to cloud virtual machines for stateful, vendor-packaged, Windows, or poorly documented workloads and reserving container replatforming for stateless, frequently deployed services that can benefit from autoscaling and self-service delivery. Rehosting through tools such as AWS MGN, Azure Migrate, and Google Migrate to Virtual Machines can preserve existing operating systems and configurations with lower near-term risk, but may retain inefficiencies and increase costs without later right-sizing. Containerization can lower long-term infrastructure costs and improve deployment velocity, but requires investment in Kubernetes operations, CI/CD, security, observability, and ongoing platform maintenance. The recommended approach is a hybrid strategy: rapidly rehost the majority of workloads to meet renewal deadlines, pilot containers with a small high-traffic segment, retire unused systems, and fund further modernization from resulting savings. Migration plans should account for licensing constraints, loss of VMware overcommit capacity, committed-use cloud discounts, and a three-to-six-month period of paying both VMware and cloud providers during cutover and rollback preparation.
Oct 11, 2026
2,747 words in the original blog post.
Progressive delivery on Kubernetes is presented as a three-layer practice that combines a delivery or GitOps engine, a rollout controller, and a traffic provider to release changes gradually, measure real production signals, and automatically promote or roll back deployments. The recommended open-source default is Argo CD for Git reconciliation paired with Argo Rollouts for canary or blue/green orchestration and metric-gated analysis, while Flux plus Flagger is positioned as a natural alternative for organizations already using Flux or service meshes. Argo CD and Flux do not perform canary analysis themselves, and Istio, Linkerd, NGINX Ingress, Gateway API, Traefik, and AWS ALB are traffic-routing layers rather than CI/CD tools; a service mesh is optional unless teams need features such as mTLS, request-level routing, or built-in telemetry. Argo Rollouts offers broad provider support, experiments, a dashboard, and a Rollout resource that replaces Kubernetes Deployments, whereas Flagger wraps existing Deployments and suits mesh- or Flux-oriented environments. Commercial platforms such as Harness, Spinnaker, Codefresh, and Octopus Deploy add managed analysis, governance, support, and approval workflows at the cost of pricing and potential lock-in, while Qovery is described as an internal developer platform that operates clusters, environments, pipelines, and access controls but still relies on Argo Rollouts or Flagger for canary analysis. Successful implementations also require a metrics backend, backward-compatible database migrations, consideration of low traffic, session affinity, parallel-version costs, and ongoing ownership of Kubernetes upgrades, making the central choice between building an open-source stack and purchasing operational support dependent on platform-team capacity, compliance needs, scale, and existing infrastructure.
Oct 11, 2026
4,890 words in the original blog post.
Azure Kubernetes Service and Amazon Elastic Kubernetes Service offer broadly equivalent, CNCF-conformant Kubernetes capabilities, making an organization’s existing cloud footprint, identity system, skills, commitments, and networking needs more important than nominal service differences. Both charge about $0.10 per cluster hour for production control planes, although AKS provides a no-SLA free tier for non-production; in practice, compute, NAT gateways, load balancers, cross-availability-zone traffic, egress, and logging typically dominate costs, with AWS cross-AZ transfer and Azure log ingestion highlighted as potential surprises. AKS may be simpler for teams already using Microsoft Entra ID and Azure tooling, while EKS better fits AWS-centric teams despite a steeper IAM, IRSA, and workload-identity learning curve; EKS also benefits from Karpenter, whereas AKS’s Azure CNI Overlay can reduce pod IP exhaustion concerns. Neither service eliminates ongoing operational work such as Kubernetes upgrades, node patching, add-on compatibility, RBAC administration, observability, backups, and cost optimization, so small teams are advised to enable automated modes, use infrastructure-as-code and GitOps, and consider developer self-service platforms or internal developer platforms. The recommended practical approach is to select the cloud already supported by organizational skills and commitments, automate operations aggressively, plan networking and identity choices carefully, and focus effort on reducing developer-facing and day-two management burdens.
Oct 11, 2026
5,208 words in the original blog post.
NIST Cybersecurity Framework 2.0 is a voluntary, outcome-based framework for organizations rather than a certification program for software vendors, so deployment platforms cannot legitimately be “NIST certified.” Instead, organizations should evaluate vendors through evidence such as current SOC 2 Type II reports, ISO 27001 or FedRAMP status where applicable, mappings to NIST SP 800-53 controls, identity and access architecture, data-flow diagrams, exportable audit logs, vulnerability-management commitments, and contractual incident-response and subprocessor terms. Deployment platforms primarily affect CSF areas involving governance, asset visibility, access control, data and platform security, monitoring, and recovery, while customers remain responsible for application security, endpoint protection, policies, personnel, and physical security. The article argues that deployment architecture is especially significant: vendor-hosted PaaS products place workloads and related controls within the vendor’s boundary, whereas bring-your-own-cloud platforms such as Qovery, Humanitec, and OpenShift can preserve an organization’s existing cloud account, keys, network boundaries, and native logging, though they may retain more operational responsibility. It recommends treating migration as a continuing risk-management exercise by assessing vendors before contracting, testing evidence and restore processes in nonproduction, documenting residual risk, and regularly reviewing controls as platforms and requirements change.
Oct 11, 2026
3,922 words in the original blog post.
For mid-sized engineering teams, actively running the same production workloads across multiple cloud providers is generally presented as more costly and operationally complex than beneficial, while maintaining the ability to migrate between providers is often a more practical goal. The decision should be driven by concrete external requirements such as customer contracts, regulatory data-residency obligations, acquisitions, temporary GPU shortages, uniquely differentiated services, or major pricing commitments rather than broad concerns about lock-in or resilience. Key costs extend beyond compute to duplicated platform engineering, IAM, networking, CI/CD, observability, compliance, on-call expertise, cross-cloud data egress, and weaker volume discounts. The proposed approach is to assess readiness through infrastructure-as-code, containerization, centralized telemetry, cost allocation, and dedicated platform capacity, then use a six-factor score to decide whether to remain single-cloud, conduct a pilot, or proceed. For most teams, a portable single-cloud model using tools such as Kubernetes, Terraform or OpenTofu, OpenTelemetry, FinOps and FOCUS is recommended, with any second-cloud strategy tested through a 90-day pilot measuring normalized costs, engineering effort, delivery speed, and incident impact.
Oct 10, 2026
5,235 words in the original blog post.
European public-sector cloud procurement in 2026 should prioritize verifiable legal and compliance requirements, required service capabilities, end-to-end workload costs, and tested exit plans rather than provider branding. European providers such as OVHcloud, Scaleway, Open Telekom Cloud, IONOS, and StackIT may offer advantages in EU jurisdiction, national certification alignment, and lower or unmetered data-egress costs, while AWS, Azure, and Google Cloud provide broader managed-service catalogues, stronger multi-availability-zone footprints, committed-use discounts, and more mature data, AI, and GPU platforms. The comparison distinguishes data residency, legal sovereignty, operational control, and technical portability, noting that EU data location alone does not resolve exposure to foreign laws or provider administrative access. It recommends validating the precise scope of certifications such as France’s SecNumCloud, Germany’s BSI C5, ISO standards, and HDS health-data hosting requirements, rather than relying on general sovereignty claims. To reduce lock-in, organizations are encouraged to use portable technologies including containers, Kubernetes, PostgreSQL, S3-compatible storage, Terraform, and documented migration processes, while the EU Data Act is expected to further reduce switching barriers from January 2027. The article presents Qovery as an optional bring-your-own-cloud developer platform intended to standardize deployments across sovereign providers, hyperscalers, and on-premises Kubernetes environments without altering the customer’s cloud contracts or compliance responsibilities.
Oct 10, 2026
4,401 words in the original blog post.
Migrating from a managed PaaS such as Heroku, Render, Fly.io, or Azure App Service to Azure Kubernetes Service requires coordinating four distinct layers: infrastructure provisioning, continuous delivery, state and data migration, and developer experience. The discussion emphasizes that containers are generally straightforward to move, while PaaS-provided capabilities such as ingress, TLS, autoscaling, secrets, managed databases, logging, preview environments, and rollbacks must be redesigned and operated on AKS. Terraform or Bicep, Azure Migrate, and Azure Container Registry support provisioning; Argo CD and Flux handle GitOps delivery but require existing manifests; Azure Database Migration Service supports database moves; and Velero is useful only when the source environment is already Kubernetes. For developer self-service, the options presented include Azure Container Apps for teams without deep Kubernetes requirements, Azure Red Hat OpenShift, a self-built Backstage and Crossplane platform, or commercial offerings such as Qovery. A recommended migration sequence includes inventorying dependencies, containerizing services, provisioning reproducible infrastructure, replacing PaaS add-ons with Azure managed services, rehearsing data migrations, operating both platforms in parallel, and shifting traffic gradually. Although AKS control-plane costs can be low, the analysis stresses that compute, observability, upgrades, security patching, on-call responsibilities, and platform-engineering staffing determine total cost, while Azure discounts and reservations may make AKS more economical at larger scale.
Oct 10, 2026
4,392 words in the original blog post.
Qovery’s migration from ingress-nginx to Envoy Gateway across hundreds of managed Kubernetes clusters reached 94% completion by October 2026, seven months later than its original March deadline, largely because the company chose not to force unvalidated production changes on customers. Its dual-stack rollout model, which ran NGINX and Envoy behind separate load balancers with test hostnames, DNS-based cutovers, and reversible settings, enabled rapid rollback and successfully limited the impact of most issues. Challenges included an internal default-setting change that disrupted deployments, complex certificate and ACME validation behavior, Envoy’s stricter defaults for timeouts, headers, encoded slashes, retries, compression, and HTTP/2, as well as Gateway API policies that replace rather than merge with broader policies. Operational complications involved client IP forwarding, insufficient Envoy replicas, excessive logging, DNS failures, GKE API limitations, and behavior changes across Envoy Gateway upgrades, including an outage caused by Lua extensions no longer being enabled by default. The experience led Qovery to emphasize explicit configuration, documentation testing, planned policy precedence, staging and changelog reviews for every minor upgrade, shared ownership, and quieter release windows, while the remaining clusters transition before NGINX and its load balancers are removed.
Oct 10, 2026
2,788 words in the original blog post.
GPU pricing comparisons across AWS, GCP, Azure, and Scaleway should rely on first-party calculators or pricing APIs, supplemented by multi-cloud comparison tools and hardware specifications that account for differences such as VRAM and NVLink versus PCIe interconnects. The cited October 2026 prices show substantial variation, with AWS H100 capacity estimated at about $6.88 per GPU-hour, Azure at about $12.29, and Scaleway’s PCIe H100 at €2.73, while comparable inference GPUs such as L4 and L40S are cheaper but vary by provider availability and VM design. The article argues that hourly GPU cost alone is insufficient for inference decisions because utilization, model fit, batching, quantization, serving software, and throughput determine the more relevant metrics of cost per request or per token. It also emphasizes that regional capacity, quota approvals, commitments, spot-instance interruption risk, data transfer, storage, Kubernetes control-plane fees, NAT charges, image pulls, and idle non-production environments can materially alter total costs. A recommended process is to define workload requirements, request quotas early, compare equivalent hardware in eligible regions, gather current provider rates, and benchmark an identical real workload on multiple clouds before making a commitment. Qovery is presented as a deployment platform rather than a GPU seller or pricing tool, enabling teams to run the same inference service in their own cloud accounts and measure actual cost performance across providers.
Oct 10, 2026
4,259 words in the original blog post.
DORA, the EU Digital Operational Resilience Act applicable since January 2025, requires financial entities to retain responsibility for ICT third-party risk even when using cloud providers, with core obligations including a complete register of ICT contractual arrangements, Article 30 contract provisions, timely major-incident reporting, due diligence and concentration-risk assessments, and documented, tested exit strategies for critical or important functions. No single product or provider delivers compliance automatically: GRC platforms such as OneTrust, ServiceNow IRM and IBM OpenPages manage registers and contract workflows; AWS, Azure and Google Cloud provide contractual addenda, resilience services and audit evidence; observability and incident tools support detection, impact assessment and reporting deadlines; and portable infrastructure approaches support demonstrable exit testing. The article emphasizes that provider addenda and cloud migration do not replace the entity’s own criticality assessments, regulatory reporting, audit evidence, or recovery and exit tests, while noting that only about 6.5% of nearly 1,000 registers in a 2024 voluntary ESA dry run passed all data-quality checks. It argues that Kubernetes-based deployments, open-source-compatible databases, infrastructure as code, and BYOC models can make portability more practical, while clarifying that DORA does not mandate multi-cloud arrangements but does require a credible, assessed and tested alternative. Qovery is presented as an engineering-layer tool for Kubernetes-native portability, deployment audit trails, RBAC and reproducible testing environments, but not as a system for registers, contract management, regulatory filings, or DORA certification.
Oct 10, 2026
4,127 words in the original blog post.
Qovery’s DevOps Team Extension is a recurring add-on service for Growth, Business, and Enterprise customers that provides a named engineer, continuous read-only infrastructure assessments, monthly reviews, biweekly working sessions, and business-hours escalation for growing teams without an in-house DevOps or SRE function. The service evaluates roughly 150 controls across reliability, security, delivery, performance, and cost, identifying issues such as publicly accessible databases, shared credentials, insufficient replicas, weak alerting, and inefficient resource use, while relating findings to claimed compliance frameworks including SOC 2, GDPR, ISO 27001, HIPAA, and DORA. Engagements begin with a free assessment and collaborative review, followed by a 90-day remediation plan and reassessment, with all changes requiring customer approval. Qovery positions the offering as a lower-cost way to access experienced infrastructure guidance before hiring a full-time specialist, combining AI-assisted analysis with engineers who can apply system-specific context and practical production experience.
Oct 10, 2026
2,037 words in the original blog post.
Qovery’s DevOps Team Extension engagement with an unnamed growing B2B SaaS company began with a read-only review of roughly 150 infrastructure controls across three AWS EKS clusters, identifying sound foundational practices but gaps in Kubernetes workload alerting, redundancy, production isolation, database exposure, shared credentials, and burstable database capacity. A joint review with the customer corrected several automated findings, uncovered configuration and Terraform isolation issues that API checks could not detect, and highlighted the absence of service criticality classifications as the most valuable strategic finding. The team agreed on a 90-day plan covering alerting, database improvements, safer preview-environment cleanup, internal-tool access controls, and a review of production service resilience. A proposed three-node high-availability Typesense cluster was successfully tested but ultimately reconsidered after discussion revealed that search was noncritical because the application could fall back to PostgreSQL, while memory capacity—not availability—was the main constraint. The recommendation therefore shifted to application-level logical sharding across smaller independent Typesense instances, limiting outages to subsets of customers, reducing memory costs compared with full dataset replication, and allowing selected customers to receive dedicated high-availability clusters if needed.
Oct 10, 2026
3,463 words in the original blog post.
Migrating from Amazon ECS to EKS generally preserves container images and CI build processes, including testing, ECR pushes, and caching, but requires replacing the final deployment step and manually recreating Kubernetes equivalents for ECS-managed capabilities such as IAM roles, ingress, service discovery, logging, secrets, autoscaling, and networking. The comparison argues that no reliable tool converts ECS task definitions directly into production-ready Kubernetes manifests: kompose only converts Docker Compose files and can provide an initial draft, while Argo CD and Flux manage GitOps reconciliation after manifests have been created and stabilized. AWS-native EKS options, including Auto Mode, Blueprints, Karpenter, and related controllers, provide flexibility and control but require significant platform engineering ownership for cluster add-ons, upgrades, and deployment standards. Qovery is presented as an alternative for teams without dedicated Kubernetes platform expertise, operating EKS in customers’ own cloud accounts while retaining existing CI build stages and handling application deployment, infrastructure resources, upgrades, and preview environments. The recommended migration approach is to inventory dependencies, establish cluster add-ons first, migrate a low-risk stateless service, use weighted traffic for parallel ECS and EKS operation, and defer stateful or complex workloads until later, while recognizing that EKS’s main additional cost is often operational expertise rather than its control-plane fee.
Oct 09, 2026
3,779 words in the original blog post.
Multi-cloud GPU orchestration is presented as three distinct layers rather than a single product category: capacity brokering to locate GPUs across providers, a Kubernetes-based runtime to execute workloads, and a developer platform for self-service environments, access control, and deployment workflows. SkyPilot is positioned as an open-source option for cross-cloud capacity acquisition and failover, while Ray handles distributed computation within existing clusters, Kueue and Volcano provide Kubernetes queueing, Run:ai and Nvidia tooling manage GPU sharing and fleet configuration, and platforms such as OpenShift AI and Qovery address higher-level operational consistency. The discussion emphasizes that portable workloads depend less on scheduler choice than on standardized containers, CUDA and driver versions, Kubernetes configuration, checkpointing, storage, IAM, and infrastructure automation. GPU scarcity, large price differences between hyperscalers and specialized GPU providers, and spot-market availability can support a multi-cloud strategy for substantial batch training workloads, although egress charges and operational complexity may offset savings. It argues that commercial commitments such as reservations and long-term discounts create stronger lock-in than technical APIs, recommending that organizations retain ownership of cloud accounts, commit only baseline demand, use portable runtime components, and treat one primary cloud with a secondary burst provider as a more practical approach than equal use of three clouds.
Oct 09, 2026
3,933 words in the original blog post.
Deployment policy can be enforced at several points in the Kubernetes delivery lifecycle, with managed platform layers generating compliant configurations before they are written and OPA Gatekeeper or Kyverno rejecting non-compliant resources at Kubernetes admission time. Gatekeeper uses Rego and can extend OPA-based policy evaluation to Terraform, CI, and application authorization, while Kyverno uses Kubernetes-native YAML policies and adds features such as resource generation and image-signature verification. Both serve as cluster-wide backstops but cannot directly govern cloud IAM, account placement, database settings, or other controls outside the Kubernetes API server. Tools such as Crossplane, Backstage, Port, and Argo CD address infrastructure provisioning, developer self-service, or GitOps delivery rather than replacing per-resource policy enforcement. The recommended approach is a layered model combining a platform golden path, admission control, GitOps delivery, and CI checks, with policies initially run in audit mode, supported by compliant templates and measured for developer friction before enforcement is enabled.
Oct 09, 2026
3,938 words in the original blog post.
How to Let Claude Code and Codex Deploy Self-Hosted Models Without Losing Control of Cost and Access
Platform teams can let AI coding agents such as Claude Code and Codex deploy self-hosted open-weight models without granting them cloud credentials by combining four distinct layers: a model-serving system such as vLLM or Ray Serve, LiteLLM as an OpenAI-compatible gateway for virtual keys, budgets, rate limits, routing, and logging, infrastructure provisioning through Terraform, OpenTofu, or an internal platform, and OpenCost or Kubecost for GPU-cost attribution. The recommended security model limits agents to opening pull requests or calling narrowly scoped platform APIs, while CI or the platform retains credentials, applies environment-specific RBAC, records audit trails, and enforces review gates for production. Cost control requires both gateway-level token and request caps and infrastructure-level safeguards including GPU quotas, node-pool size limits, auto-stop or scale-to-zero for non-production services, and daily anomaly alerts. Self-hosting can be economical under sustained, predictable utilization but becomes costly when GPUs are idle, making hosted APIs preferable for bursty workloads or tasks requiring frontier-model quality. A hybrid architecture can route routine, high-volume work to local models while sending complex reasoning to hosted providers through the same gateway. Kubernetes-native containers, manifests, and OpenAI-compatible endpoints support portability across cloud and on-premises GPU environments, though this flexibility requires more operational ownership than managed services.
Oct 08, 2026
3,622 words in the original blog post.
Moving from a managed PaaS to AWS, Azure, Google Cloud, Scaleway, or a self-managed Kubernetes environment does not invalidate an ISO/IEC 27001 certification or restart its three-year cycle, because certification applies to an organization’s information security management system rather than its hosting provider. The migration does, however, require updates to the Statement of Applicability, risk assessment, supplier records, asset inventory, architecture documentation, and planned-change records, along with notification to the certification body. Cloud providers’ own ISO certifications can support supplier due diligence and cover underlying physical infrastructure, but customers remain responsible for workload-level controls such as IAM, patching, configuration management, logging, environment separation, and change approvals. The discussion identifies ten Annex A controls particularly affected by deployment practices, emphasizing that audit difficulties commonly arise when evidence is fragmented across cloud consoles, CI systems, Terraform, and informal approval channels. It distinguishes cloud infrastructure providers, compliance automation tools, cloud-security posture management products, and deployment control planes, arguing that the latter can centralize access controls, environment separation, and deployment audit trails but cannot replace broader ISMS governance or certification. It also stresses that auditors assess evidence over time, so organizations should establish and operate new controls well before an audit window, ideally for a full quarter.
Oct 08, 2026
6,549 words in the original blog post.
Argo CD is presented as a capable open-source GitOps reconciler whose license-free model can still create substantial operational work through controller scaling, high availability, security patching, RBAC, credential management, and upgrades. The comparison groups alternatives into hosted Argo CD services such as Akuity and Codefresh GitOps, vendor-supported or cloud-integrated offerings including Red Hat OpenShift GitOps, Azure Arc/AKS Flux extensions, and EKS add-ons, CI/CD-integrated choices such as Harness and GitLab, and internal developer platforms such as Qovery and Plural. Hosted Argo options reduce control-plane operations while retaining existing Argo manifests and workflows, whereas Flux offers a smaller controller-oriented architecture but still requires operational ownership unless managed through a cloud wrapper. The text distinguishes GitOps reconciliation of arbitrary Kubernetes resources from broader delivery outcomes such as git-based deployments, reproducible environments, rollbacks, developer self-service, and pull-request preview environments. It argues that platforms like Qovery can provide the latter without requiring teams to author Argo Application resources, while noting that Argo CD or Flux remain better suited for large-scale reconciliation of arbitrary CRDs and platform-level Kubernetes resources.
Oct 08, 2026
3,936 words in the original blog post.
OVHcloud’s 2026 pricing changes reportedly raise costs for some recent dedicated-server configurations by up to 87% on new orders, while existing customers may face smaller renewal increases tied mainly to memory and storage, alongside Public Cloud instance increases. The article argues that organizations should first verify renewal quotes against current list prices, negotiate consolidated and capped multi-year terms, and reduce waste from idle non-production environments, overprovisioned infrastructure, unused resources, and unclaimed discounts before considering migration. It emphasizes comparing providers using total workload cost rather than compute prices alone, since egress, storage, managed services, support, and commitment discounts can substantially alter costs across OVHcloud, AWS, Google Cloud, Azure, Scaleway, and Hetzner. It presents containers, Kubernetes, infrastructure-as-code, and customer-owned cloud accounts as ways to improve portability and negotiating leverage, while acknowledging that managed databases, networking, IAM, and proprietary services can still make moves difficult. The article also notes that migrations require inventorying systems, building parallel environments, transferring data, and managing cutover risk, and promotes Qovery as a platform for deploying workloads across customer-owned cloud accounts or existing Kubernetes clusters.
Oct 08, 2026
3,719 words in the original blog post.
Qovery built an AI agent to perform initial investigations of Sentry alerts, helping engineers distinguish actionable bugs from harmless noise and issues requiring further review. Triggered by alerts routed through a Slack relay, separate frontend and backend agents examine error details and relevant code, then respond in the original Slack thread and Sentry with either a draft pull request, an explanation and archived alert, or a set of hypotheses and recommended human checks. Examples include identifying and proposing a fix for a table-sorting crash caused by a missing value, dismissing a browser-extension-blocked tracking request as noise, and flagging a GitHub authentication failure for credential review. Engineers remain responsible for validating diagnoses, testing proposed changes, and deciding whether to merge pull requests, while the agent does not make production decisions independently. Hosted as Qovery Agent Tasks in the company’s cloud account, the system can use controlled infrastructure access and service context such as deployment history, logs, and metrics, though Qovery is still evaluating its effectiveness and has not yet measured time savings.
Oct 08, 2026
956 words in the original blog post.
Kubernetes 1.34 is approaching end of upstream support, after which security patches cease, while managed-service deadlines and costs differ across Amazon EKS, Google GKE, and Azure AKS: EKS offers paid extended support before automatic upgrades, GKE support depends on release channels, and AKS provides LTS through its Premium tier. The recommended upgrade process is to first scan live resources and Helm releases for deprecated APIs using tools such as Pluto or kubent, verify add-on compatibility and PodDisruptionBudget settings, rehearse in production-like staging, and then upgrade the control plane one minor version at a time, followed by node pools, add-ons, and smoke testing before retiring old nodes. Most upgrade failures stem not from the control-plane change but from removed APIs, stalled node drains, incompatible controllers, node-image changes, and infrastructure-as-code version drift. Because managed platforms generally do not allow skipping minor versions or reliable downgrades, resilient rollback plans rely on retaining old node pools or using blue-green clusters. The piece estimates that maintaining supported versions can require one to three engineering days per cluster upgrade and argues that teams operating several clusters should consider provider automation, internally built CI/CD and GitOps workflows, or managed platform services to reduce recurring operational work.
Oct 08, 2026
4,239 words in the original blog post.
FISA Section 702 can apply to European data hosted by US-headquartered cloud providers because its reach depends on a provider’s US jurisdiction and ability to access plaintext, rather than server location, making standard EU regions insufficient protection on their own. Unlike the CLOUD Act, which governs law-enforcement demands and allows limited legal challenges, Section 702 supports foreign-intelligence collection targeting non-US persons abroad, generally without customer notice or customer-side remedies. The discussion argues that meaningful risk reduction depends on eliminating US-controlled access through customer-held external encryption keys, confidential computing, EU-only operations, careful review of support, telemetry, backup, and control-plane paths, and selecting providers without a US nexus where necessary. Sovereign offerings from AWS, Microsoft, and Google may reduce practical exposure through European entities, staff, or partner-operated key management, but their legal effectiveness depends on whether a US-controlled entity can still access plaintext. For GDPR compliance, organizations should document these risks in Transfer Impact Assessments, account for the still-contested EU-US Data Privacy Framework, and maintain portability, while the text promotes self-hosted deployment control planes such as Qovery as a way to avoid placing sensitive deployment data in third-party SaaS tenancy.
Oct 07, 2026
5,118 words in the original blog post.
Mistral Large 4 is presented as a roughly 1.05-trillion-parameter mixture-of-experts model with 49 billion active parameters per token, designed for frontier coding and multimodal use but, as of October 2026, not yet downloadable and without published final license terms for commercial self-hosting. Its full weights require substantial VRAM despite sparse activation, estimated at about 2.1 TB in bf16, 1.05 TB in FP8, or 525 GB at 4-bit quantization, making practical deployments dependent on multi-GPU H100 or H200 nodes and careful allowance for long-context KV cache memory. The recommended architecture uses vLLM or, for comparison, SGLang as the model-serving engine, with a LiteLLM or Portkey gateway providing authentication, budgets, logging, fallback routing, and protocol translation for Claude Code, while Codex can use an OpenAI-compatible gateway directly. The analysis argues that self-hosting is primarily justified by data residency, version control, rate-limit independence, and sufficiently steady demand rather than by cost alone, since an eight-H100 node can cost about $40,000 monthly when run continuously and hosted APIs generally remain cheaper at low utilization. It emphasizes scale-to-zero, discounted or spot capacity, quantization, cached weights, private networking, monitoring, and hybrid routing that sends routine coding tasks to self-hosted models while reserving difficult tasks for hosted frontier services. Different infrastructure tools are assigned distinct roles, including SkyPilot for capacity sourcing, KubeAI or KEDA for autoscaling, vLLM for inference, LiteLLM or Portkey for gateway governance, and Qovery or similar platforms for Kubernetes operations, access controls, and deployment workflows.
Oct 07, 2026
4,349 words in the original blog post.
Qovery describes a pilot AI agent built with its Agent Tasks platform to streamline vulnerability triage for SOC 2 compliance, including a commitment to remediate critical issues within 10 days. Scheduled to read open findings and due dates from Vanta, the agent prioritizes vulnerabilities by urgency, maps affected packages to code across 16 GitHub and GitLab repositories, identifies duplicate root causes or fixes awaiting deployment, and prepares tested changes when possible. It opens only verified draft pull or merge requests and posts concise Slack reports identifying urgent items, review-ready fixes, and cases requiring human judgment. The system is constrained from modifying Vanta records, default branches, releases, or merges, while engineers retain responsibility for reviewing, approving, and deploying every change. Qovery says it is still evaluating the appropriate schedule, measurable time savings, and the categories of fixes the agent can handle reliably before adopting the approach more broadly.
Oct 07, 2026
1,801 words in the original blog post.
Cloud portability is presented as a measurable switching cost, defined by the engineering time and operating expense required to run a real workload on a second cloud or Kubernetes target rather than as a binary architectural claim. The proposed approach scores individual services from 0 to 3 across compute, data, networking, identity and secrets, and CI/CD plus infrastructure as code, with the weakest layer determining practical migration difficulty; proprietary data services, data-transfer volumes, undocumented IAM policies, and networking configurations are identified as more significant obstacles than containerized compute. A two-week portability drill is recommended in which teams select a representative production service, codify its environment with Terraform or OpenTofu and Helm, deploy it elsewhere, replay realistic traffic, measure costs over a billing cycle, and publish deployment time, a cost report, and an owned list of breakages. Cost assessment should combine provider list prices, infrastructure-plan estimates, measured post-deployment allocation, and engineering costs such as dual-running systems, migration work, and delayed roadmap activity, while accounting for commonly overlooked items including egress, NAT gateways, cross-zone transfer, load balancing, observability, backups, and idle non-production resources. The discussion distinguishes portability from continuously operating multi-cloud infrastructure, argues for periodic drills as both risk preparation and contract-negotiation leverage, and recommends portable containers, infrastructure as code, open data interfaces where appropriate, and provider-neutral identity patterns; it also positions Qovery as a deployment layer that can preserve a consistent developer workflow across cloud accounts and Kubernetes clusters while requiring separate tools for cost analysis and migration assessment.
Oct 07, 2026
4,920 words in the original blog post.
Migrating from Red Hat OpenShift to Amazon EKS can eliminate OpenShift’s per-core subscription costs, though organizations must still pay for EKS control planes, underlying AWS compute, observability, registries, CI/CD, and any replacement platform engineering work. The migration primarily involves converting OpenShift-specific resources such as DeploymentConfigs, Routes, ImageStreams, BuildConfigs, SecurityContextConstraints, projects, and oc-based workflows to Kubernetes Deployments, Ingress or Gateway APIs, Amazon ECR, external CI pipelines, Pod Security Admission, namespaces with RBAC, and kubectl. A key technical risk is container images that rely on OpenShift’s arbitrary UID security model, which may fail on EKS unless they are made explicitly non-root and compatible with appropriate filesystem permissions. The recommended approach is a parallel, phased migration that inventories workloads and operators, establishes an EKS landing zone, pilots a stateless application, validates measurable performance and reliability criteria, migrates build systems, moves stateful services last, and aligns decommissioning with Red Hat renewal dates. Because EKS provides Kubernetes rather than OpenShift’s integrated developer console, builds, GitOps, monitoring, registry, and tenancy features, teams must either assemble and operate alternatives such as Argo CD, Backstage, Helm, and Prometheus tooling, adopt an internal developer platform, or consider ROSA if retaining the managed OpenShift experience is more important than removing its service fees.
Oct 07, 2026
5,469 words in the original blog post.
Deploying open-weight models such as Mistral, Llama, Qwen, DeepSeek, or gpt-oss internally is presented as a four-layer architecture: an inference server such as vLLM, SGLang, or TGI; a Kubernetes-based orchestrator such as KubeAI or Ray Serve; an OpenAI-compatible gateway such as LiteLLM for routing, keys, budgets, and usage tracking; and a deployment governance layer such as Qovery or internally maintained Terraform and CI workflows. Claude Code and Codex can connect through custom endpoints, although Claude Code has limitations for non-Claude backends and Codex may require protocol translation, making the gateway a central integration and control point. The discussion emphasizes that agent-facing keys should be revocable, budget-limited, and separated from cloud credentials or Kubernetes access, while deployment agents should use tightly scoped templates, RBAC, quotas, ephemeral environments, audit logging, and human production approval to reduce risks from prompt injection and excessive permissions. GPU idle time is identified as a larger cost problem than hourly pricing, with scale-to-zero, nonproduction auto-stop, environment expiration, batching, quantization, and selective spot use suggested as major cost controls. Self-hosting is generally portrayed as economically viable only at sustained high token volumes and for requirements such as data residency, model pinning, air-gapped systems, or lower latency; for most teams, a hybrid approach is recommended, using self-hosted models for routine high-volume tasks and hosted frontier models for more demanding agentic coding work.
Oct 07, 2026
4,572 words in the original blog post.
BSI C5 is Germany’s government-issued cloud security criteria catalogue, but compliance is demonstrated through a C5-Testat, an independent auditor’s attestation under ISAE 3000 rather than a public certificate. It is legally required mainly for German federal cloud procurement, while private SaaS providers encounter it as a commercial requirement from enterprises, regulated financial institutions, healthcare organizations, and public-sector tenders. Type 1 reports assess whether controls are suitably designed at a point in time, whereas Type 2 reports assess their effective operation over roughly six to twelve months and are increasingly expected by buyers. C5 overlaps with ISO 27001 and SOC 2 but adds mandatory disclosures on jurisdiction, data-processing locations, subprocessors, certifications, and government-access requests, with reports typically shared under NDA. Although cloud providers such as AWS, Azure, and Google Cloud can cover underlying infrastructure controls within their attested scope, SaaS companies remain responsible for complementary controls including access management, reviewed deployment changes, logging and retention, encryption, patching, portability, incident response, and accurate data-location documentation. The central preparation challenge is producing continuous, auditable operational evidence through infrastructure as code, controlled production pipelines, centralized logging, and well-defined cloud and subprocessor boundaries rather than merely maintaining policies.
Oct 06, 2026
4,460 words in the original blog post.
Hetzner does not offer a native managed Kubernetes service, but its cloud infrastructure, Kubernetes cloud-controller-manager, and CSI driver can support production clusters operated either by third parties such as Syself or Cloudfleet or by teams using tools such as Kube-Hetzner and Talos. It can be a cost-effective option for small teams because compute and egress costs are generally much lower than on AWS, Google Cloud, or Azure, and there is no separate managed-control-plane charge, but these savings shift responsibility for control-plane availability, etcd backups, upgrades, monitoring, disaster recovery, and stateful services onto the customer or provider. Hetzner is presented as suitable for cost-sensitive, primarily stateless workloads in its limited European, US, and Singapore regions when a team has Kubernetes expertise and accepts self-managed data services, while hyperscalers remain better suited to organizations needing managed databases, workload identity, broader compliance certifications, extensive global coverage, or stringent recovery objectives. The article recommends avoiding manually maintained kubeadm clusters, validating storage, network, regional, and compliance constraints, and considering hybrid setups that combine Hetzner compute with managed services elsewhere; it also distinguishes cluster operation from application delivery, noting that platforms such as Qovery can deploy to existing Hetzner clusters but do not manage their Kubernetes control planes.
Oct 06, 2026
3,367 words in the original blog post.
STACKIT Kubernetes Engine, IONOS Managed Kubernetes, and Open Telekom Cloud CCE provide conformant managed Kubernetes control planes in EU data centers, making them attractive to organizations prioritizing EU jurisdiction, GDPR-related transfer concerns, and certifications such as BSI C5 and ISO 27001. Kubernetes workloads and vendor-neutral tools including Helm, Argo CD, Flux, Kyverno, OPA Gatekeeper, Prometheus, Grafana, KServe, and vLLM generally operate consistently across all three, while differences arise in GPU availability, managed-service breadth, Terraform and Crossplane coverage, storage and load-balancer configuration, and control-plane pricing. STACKIT is positioned as strongest for sovereign GPU options and a broad managed-database catalog, IONOS emphasizes simpler pricing and a free control plane, and Open Telekom Cloud offers a comparatively deep service stack and Telekom-backed footprint. The discussion argues that organizations must still assemble or adopt a developer platform for CI/CD, secrets, RBAC, preview environments, upgrades, cost controls, and policy enforcement, with options ranging from DIY CNCF tooling to Rancher, OpenShift, and bring-your-own-cluster platforms such as Qovery. EU-hosted infrastructure can reduce third-country transfer exposure, but complete data sovereignty depends on auditing the full stack, including CI runners, observability services, SaaS integrations, and AI APIs; migrations should therefore focus on replacing cloud-specific dependencies, validating storage and data recovery, and establishing quotas and non-production shutdown policies.
Oct 06, 2026
3,201 words in the original blog post.
Bleu, S3NS, and Scaleway serve distinct roles for French SaaS scaleups navigating cloud sovereignty and compliance requirements: Bleu is an Orange–Capgemini joint venture offering a Microsoft Azure-compatible environment, S3NS is a Thales-majority venture offering Google Cloud-compatible services, and Scaleway is an independent Iliad-owned European public cloud with its own infrastructure, self-service access, public pricing, and managed Kubernetes. As of 6 October 2026, the text states that S3NS’s PREMI3NS offer is SecNumCloud 3.2 qualified for IaaS, CaaS, and PaaS, while Bleu and Scaleway remain in qualification progress; it stresses that qualification applies to specific offers and scopes rather than entire companies. The recommended choice depends chiefly on an explicit contractual or regulatory need, such as a SecNumCloud clause, public-sector, OIV/OSE, or health-data requirements, rather than GDPR alone, which can generally be met through EU hosting regions and suitable transfer and security safeguards. Scaleway is presented as best suited to teams seeking transparent public pricing and European independence, whereas Bleu and S3NS may reduce migration friction for organizations already using Azure or Google Cloud but require enterprise contracts and negotiated pricing. The text also argues that portability through containers, Kubernetes, Terraform, and compatible storage should be maintained until compliance needs require migration, and that preserving developer features such as automated deployments and preview environments often requires adding or building a platform layer rather than relying on raw Kubernetes alone.
Oct 06, 2026
4,515 words in the original blog post.
AWS App Runner is presented as a simple deployment option for single-container, request-driven HTTP services, but with fixed limits including a maximum of 4 vCPUs and 12 GB of memory per service, no sidecars, GPUs, batch jobs, StatefulSets, or advanced traffic-management features. The piece argues that existing users should plan alternatives because it states AWS has closed App Runner to new customers and does not intend to add features, while noting that low-traffic and early-stage workloads may still fit the service well. It identifies larger compute requirements, multi-container needs, GPU or queue-based workloads, advanced policies such as mTLS and canary releases, and sustained utilization costs as primary reasons to consider Amazon EKS. EKS offers Kubernetes flexibility, shared node capacity, GPU support, custom autoscaling, and more deployment controls, but shifts responsibility for networking, observability, upgrades, security, and cluster operations to the customer. The proposed migration follows five phases: inventorying current services, adapting containers for Kubernetes probes and shutdown behavior, establishing a production EKS environment with IAM and ingress components, gradually shifting traffic through weighted DNS routing, and retiring App Runner resources after validation. It also emphasizes that migration risks commonly arise from health checks, secrets, IAM permissions, networking, scaling behavior, and operational maintenance rather than Kubernetes manifests, and it suggests internal developer platforms such as Qovery as one way to preserve a simplified developer workflow on EKS.
Oct 05, 2026
4,251 words in the original blog post.
Migrating a small application from Heroku need not require a full rewrite because its Procfile, buildpacks, and configuration variables are generally portable, while add-ons, release-phase behavior, scheduled jobs, logging, and database migration require more deliberate work. Dokku, Railway, and Fly.io offer the closest path for retaining Procfile and buildpack workflows, whereas Render, DigitalOcean App Platform, Northflank, and Qovery typically require a Dockerfile or prebuilt container image, with each Procfile process becoming an independently configured service. The article recommends exporting configuration variables early, explicitly recreating release migrations and operational integrations, testing for memory leaks that Heroku’s regular dyno restarts may have concealed, and keeping Heroku available during a staged DNS cutover. Costs for a web process, worker, and roughly 10 GB PostgreSQL database vary substantially: self-hosted Dokku is cheapest in direct spending but shifts server maintenance to the team, managed PaaS options reduce operations work, and Qovery becomes more economical chiefly at larger scale or when organizations benefit from using their own cloud accounts and committed-use discounts. A short Dockerfile or an OCI image produced with Cloud Native Buildpacks can improve portability across future platforms, while the preferred destination ultimately depends on whether the team prioritizes minimal migration changes, managed convenience, low infrastructure costs, multi-region deployment, or control of its own cloud environment.
Oct 05, 2026
3,494 words in the original blog post.
Migrating a complex microservices application from a managed platform such as Heroku, Render, or Fly.io to AWS, GCP, Azure, Scaleway, or a self-managed Kubernetes environment requires separate planning for infrastructure provisioning, application delivery, data migration, and replacement of platform features such as add-ons, schedulers, log drains, configuration management, and preview environments. Infrastructure tools including OpenTofu, Terraform, Pulumi, and Crossplane create cloud resources but do not replace developer workflows, while Argo CD and Flux support Kubernetes GitOps for teams with existing platform expertise; internal developer platforms such as Qovery, DuploCloud, and LocalOps aim to combine provisioning and deployment workflows in a customer-owned cloud account. The central technical risk is database cutover rather than application deployment, with streaming PostgreSQL logical replication or cloud migration services enabling a short write freeze and typically less than 10 minutes of downtime, whereas dump-and-restore approaches can create outages lasting hours. Successful cutovers also require advancing database sequences, validating data and extensions, lowering DNS TTLs well in advance, shifting traffic gradually using weighted routing, testing graceful shutdown behavior, and retaining the original environment for rollback. The cited migration timelines range from roughly 6 to 12 weeks when using a platform that manages provisioning and delivery to 4 to 9 months for teams building infrastructure and GitOps capabilities themselves, with developer-experience preservation and early inventory of hidden dependencies identified as major determinants of success.
Oct 05, 2026
4,915 words in the original blog post.
Heroku alternatives should be assessed by how their total costs scale across production, databases, bandwidth, builds, and especially staging and preview environments, rather than by entry-level compute prices alone. The comparison groups Render, DigitalOcean App Platform, and Upsun as managed platforms with relatively simple fixed or structured pricing; Railway and Fly.io as usage-based options suited to spiky or edge-oriented workloads; and Northflank or Qovery as platforms that can support bring-your-own-cloud deployments. It argues that Heroku remains useful for small teams seeking minimal operational work but can become costly as applications require higher dyno tiers, add-ons, and multiple non-production environments. Managed PaaS services offer convenience but may limit access to cloud-provider discounts, while BYOC models let organizations pay cloud providers directly and use commitments or reservations, though they introduce account, Kubernetes, and control-plane costs that may not suit small deployments. Migration planning should address runtime compatibility, database cutovers, add-on replacements, and preservation of developer workflows such as automated deployments and pull-request previews. The recommended approach is to trial two or three candidates with a representative non-critical service and compare actual invoices, deployment experience, and operational requirements before committing.
Oct 05, 2026
2,967 words in the original blog post.
AWS App Runner, Amazon ECS, and Amazon EKS address different container needs: App Runner is positioned for small numbers of stateless HTTP services requiring rapid deployment, ECS on Fargate as the default for most AWS production workloads without Kubernetes-specific needs, and EKS for teams that need Kubernetes capabilities such as Helm charts, operators, service meshes, event-driven scaling, or multi-cloud portability. Their costs differ in billing structure, with App Runner charging for provisioned memory and active CPU, Fargate charging per vCPU and memory usage, and EKS adding a cluster control-plane fee plus node costs; networking, load balancing, logging, idle environments, and engineering time may be larger expenses than compute. The central distinction is operational ownership: App Runner abstracts infrastructure heavily, ECS requires management of tasks and supporting AWS services, while EKS requires ongoing responsibility for Kubernetes upgrades, add-ons, networking, identity, autoscaling, ingress, RBAC, and developer workflows. The discussion argues that teams adopting EKS should explicitly choose a platform-operating model, whether through internal platform staff, managed Kubernetes providers, or developer platforms such as Qovery, while AWS Copilot is highlighted as a simpler ECS-focused deployment option. For ECS-to-EKS migrations, it recommends inventorying services and dependencies, translating workloads gradually, running both environments in parallel with weighted traffic routing, and avoiding simultaneous migration of stateful data services.
Oct 05, 2026
4,331 words in the original blog post.
One Control Plane for Kubernetes on AWS, GCP, Azure, Scaleway and On-Prem: 8 Options Compared (2026)
Managing Kubernetes across AWS, GCP, Azure, Scaleway, and on-premises environments generally requires combining tools rather than relying on a single control plane, because cluster lifecycle, security governance, and application delivery have distinct requirements. The comparison groups eight options into fleet managers such as Spectro Cloud Palette, Rancher, and Platform9; infrastructure-as-code platforms including Crossplane and Spacelift; and application delivery platforms such as Qovery, Devtron, and Northflank. Fleet tools are positioned for bare metal, edge, and cluster operations, while delivery platforms focus on developer self-service, CI/CD, and preview environments; Crossplane and Spacelift address declarative infrastructure but do not provide complete application delivery. Consistent cross-cloud governance is best achieved through Kubernetes-native admission controls such as Kyverno or OPA/Gatekeeper, deployed through GitOps tools like Argo CD or Flux, while portable CI/CD separates standardized container builds from cloud-specific deployment details. The text argues that consolidation savings primarily come from reducing duplicate pipelines, platform-engineering effort, and idle non-production resources rather than avoiding relatively modest managed-cluster control-plane fees, and recommends choosing platforms based on fleet complexity, on-premises needs, provider coverage, existing infrastructure practices, and whether cloud billing and data residency must remain in the organization’s own accounts.
Oct 04, 2026
3,739 words in the original blog post.
AI agents can assist cloud migrations most reliably with discovery and dependency mapping, code and infrastructure-as-code translation, characterization test generation, documentation, and post-migration cost analysis, where they produce reviewable reports, diffs, or plans. They remain poorly suited to high-context decisions such as target architecture, data cutover, network and identity design, compliance accountability, and defining the operating model, which require human ownership. The recommended control model is “propose-approve-apply,” using read-only agent credentials, pull requests, policy-as-code checks, staged validation in ephemeral environments, approval requirements scaled to production risk, immutable audit trails, and deterministic rollback rather than direct production access. The article argues that no single migration tool covers every task, so teams commonly combine assessment services, code assistants, Terraform and policy tools, and a developer platform such as Qovery or a self-managed Backstage, Argo CD, and Terraform stack. Although vendors report substantial productivity improvements, independent research suggests outcomes vary and can be offset by review, remediation, and incident costs, making net benefit dependent on governance and validation. Lasting modernization also requires a post-migration self-service developer path, standardized deployment practices, cost controls, lifecycle ownership, and RBAC, while regulated or air-gapped deployments require particular attention to model hosting, prompt data residency, and comprehensive auditing.
Oct 04, 2026
3,619 words in the original blog post.
ISO/IEC 27001 certification applies to an organization’s information security management system rather than transferring from any cloud or deployment vendor, so platform selection primarily determines which Annex A controls a SaaS company must operate and evidence itself. Certified hyperscalers such as AWS, Google Cloud, Azure, and Scaleway provide infrastructure-level assurances and powerful security primitives but leave customers responsible for IAM design, logging retention, configuration, environment separation, and change records, while managed PaaS providers such as Heroku, Render, and Vercel reduce operational burden at the cost of control granularity. Compliance automation products including Drata, Sprinto, and Vanta collect evidence, monitor controls, and support access reviews but do not enforce deployment-pipeline controls. The comparison identifies 11 deployment-related Annex A controls, particularly access management, privileged access, logging, secure development, environment separation, and change management, and argues that auditors commonly seek proof of production access, approvals, deployment history, log retention, secrets management, and isolation between environments. For many mid-sized SaaS companies, the proposed approach combines a certified cloud provider, a compliance automation tool, and a standardized deployment layer such as Qovery or a self-built Backstage and Argo CD stack, with responsibilities, retention settings, and control owners established before migration.
Oct 04, 2026
4,389 words in the original blog post.
Platform teams seeking to let coding agents such as Claude Code and Codex deploy self-hosted open-source models are advised to use a layered architecture rather than a single product: vLLM for high-throughput production inference, Kubernetes or KServe for deployment and scaling, LiteLLM for model routing, virtual keys, budgets, and rate limits, and an internal developer platform or policy-controlled infrastructure tooling for audited deployment automation. The central recommendation is to give agents a narrowly scoped deployment interface instead of broad cloud credentials, using environment-specific RBAC, pull-request workflows, OIDC-based ephemeral credentials, and deployment logs to limit risk. GPU costs should be controlled through both infrastructure measures, including scale-to-zero node pools, right-sizing, spot capacity, and automatic shutdown of nonproduction environments, and gateway-level budget caps and rate limits. Self-hosting can be economical for predictable, sustained, high-volume workloads or where data residency and private-network requirements matter, but hosted APIs are generally cheaper for intermittent or low-volume use, making a hybrid approach practical for many organizations.
Oct 04, 2026
2,631 words in the original blog post.
French public-sector SaaS procurement in 2026 is shaped by separate requirements for sovereignty, health-data hosting, privacy, accessibility, and tender documentation, with eligibility often determined before price or product quality are assessed. The “cloud au centre” doctrine directly requires French State bodies to use approved cloud paths and, for sensitive data, SecNumCloud-qualified services protected from non-EU extraterritorial laws; local authorities and hospitals are not legally bound in the same way but commonly include similar conditions in tenders. SecNumCloud is an ANSSI qualification for a specific cloud offering rather than for applications deployed on it, so vendors remain responsible for their software, administrative access, support staff, subprocessors, data location, and reversibility arrangements. HDS certification independently applies to personal health data, while RGPD obligations cover lawful processing and international-transfer safeguards; neither HDS nor EU data residency alone establishes sovereignty. French-qualified providers such as OVHcloud, 3DS Outscale, Cloud Temple, Numspot, and S3NS can meet relevant sovereignty requirements, whereas the main AWS, Azure, and Google Cloud platforms generally offer broader managed-service ecosystems but lack SecNumCloud qualification and retain exposure to US jurisdiction. Although some French providers can be less expensive for comparable workloads, their narrower service catalogues may create operational and engineering costs. The recommended practical approach for many SaaS companies is to maintain one portable codebase and deploy it into a buyer’s or qualified provider’s cloud account, supported by clear evidence of certification scope, security controls, EU-based operations, accessibility, subcontractors, and data-exit plans before a tender closes.
Oct 03, 2026
5,359 words in the original blog post.
For a 10-person engineering team, container deployment choices generally fall into five categories: serverless containers such as Google Cloud Run and AWS ECS Fargate, vendor-hosted PaaS offerings including Upsun, Render, and Railway, lightweight orchestration with HashiCorp Nomad, self-operated managed Kubernetes such as EKS, GKE, and AKS, and BYOC platforms such as Qovery, Northflank, and Porter that operate Kubernetes within a customer’s cloud account. The comparison argues that managed Kubernetes still leaves teams responsible for substantial cluster work, including node upgrades, networking, ingress, autoscaling, observability, and security configuration, making it the most operationally demanding option despite provider-managed control planes. Serverless platforms are presented as the lowest-operations and often lowest-cost choice for bursty, stateless HTTP workloads, while vendor PaaS products prioritize fast deployments, managed databases, and preview environments but may limit cloud-account control, data residency options, and access to committed-spend discounts. BYOC Kubernetes is positioned as a middle path for teams expecting GPU workloads, Helm charts, operators, compliance requirements, or larger multi-environment deployments, although it retains cloud infrastructure costs alongside platform fees. Nomad remains a simpler but narrower alternative suited mainly to organizations already invested in HashiCorp tooling. The recommended decision process is to choose based on the primary constraint—such as speed, workload type, compliance, cloud billing ownership, or portability—and validate shortlisted services through a practical one-week deployment and incident-response test rather than selecting Kubernetes by default.
Oct 03, 2026
5,268 words in the original blog post.
For large microservices migrations to AWS, the central choice is often less about Fargate versus EKS than about who will operate the resulting platform over time. EKS on EC2 generally offers greater control, support for DaemonSets, GPUs, custom networking, bin-packing, Spot capacity, and Savings Plans, making it more economical for dense, steady workloads, while Fargate reduces node-management work and can suit bursty, isolated, or small-team workloads but has higher per-pod pricing and feature limitations. Many organizations use a mixed EKS model, placing baseline services on EC2 node groups and bursty or tenant-isolated workloads on Fargate. Migration requires separate tools for discovery, packaging, data transfer, cluster-state restoration, and incremental traffic cutovers, including AWS Migration Hub, App2Container, Velero, AWS DMS, DataSync, and routing options such as Istio, App Mesh, load balancers, or Route 53. The discussion distinguishes supporting tools such as policy engines and API-deprecation scanners from actual migration tools, and recommends phased, rollback-capable migrations that move low-risk stateless services before stateful systems. It also compares post-migration platform options, including OpenShift, Cloud Foundry/Korifi, internally built Kubernetes platforms, and Qovery, emphasizing that upgrades, IAM, networking, observability, CI/CD, cost controls, and non-production environment sprawl can require a dedicated platform team or managed-service subscription long after the initial migration ends.
Oct 03, 2026
3,316 words in the original blog post.
For teams of 3–10 engineers moving from a managed PaaS to Kubernetes, the central recommendation is to choose portable tools that communicate through the Kubernetes API and install via Helm or Operators, rather than cloud-specific runtimes such as AWS App Runner, ECS/Fargate, Azure Container Apps, or Google Cloud Run. It organizes a practical stack into CI with GitHub Actions or GitLab CI, infrastructure provisioning with Terraform or OpenTofu, delivery through Argo CD, Flux, or a platform layer, a developer-facing interface such as Qovery, Backstage, or Humanitec, and observability through tools including Prometheus, Grafana, Datadog, or Honeycomb. The discussion stresses that leaving a PaaS transfers responsibility for pipelines, networking, TLS, secrets, autoscaling, monitoring, database recovery, rollbacks, upgrades, security patching, and on-call support to the team, making recurring operational effort more consequential than basic cloud infrastructure costs. It advises migrating incrementally by proving the full pipeline on a low-risk service, retaining the former platform during cutover, moving databases last, and measuring outcomes using DORA metrics. Safe recovery requires both rapid application rollback through immutable image versions and Kubernetes deployment revisions, and careful database practices such as expand-contract migrations and tested point-in-time restores. While it presents Kubernetes portability and bring-your-own-cloud platforms such as Qovery as useful for teams seeking cost control, private networking, compliance, or reduced vendor lock-in, it also concludes that teams without such concrete requirements may be better served by remaining on a managed PaaS.
Oct 02, 2026
5,648 words in the original blog post.
Small engineering teams can reduce time spent on cloud operations by using either a managed platform-as-a-service, which runs applications in a vendor-owned account, or a bring-your-own-cloud internal developer platform that automates infrastructure within the team’s own AWS, GCP, Azure, Scaleway, or Kubernetes environment. The discussion argues that building VPCs, IAM roles, load balancers, Kubernetes clusters, and database configurations manually with Terraform or similar tools is generally better suited to teams with dedicated platform expertise, because ongoing maintenance includes upgrades, security patches, costs, backups, and incident response. For production PostgreSQL, it recommends managed offerings such as Amazon RDS or Aurora, Google Cloud SQL, Azure Database for PostgreSQL, Neon, or Crunchy Bridge, emphasizing automated backups, point-in-time recovery, high availability, monitoring, and tested restoration procedures over self-hosted databases. It distinguishes deployment platforms, infrastructure-as-code frameworks, Kubernetes control planes, and IaC governance tools, noting that each solves different operational problems. The article presents Qovery as a BYOC option that combines git-based deployments, managed cloud databases, preview environments, and automated networking while retaining customer ownership of cloud billing, data residency, discounts, and access to standard Kubernetes and cloud APIs. It also advises teams to manage costs through disposable preview environments, auto-stopping nonproduction workloads, right-sizing resources, and maintaining an escape path through inspectable infrastructure and portable underlying services.
Oct 02, 2026
3,958 words in the original blog post.
Managing GitOps across multiple Kubernetes clusters is less constrained by reconciliation than by governance consistency, including uniform sync settings, RBAC, secret handling, approvals, and auditable deployment records. Argo CD and Flux are both CNCF-graduated, production-oriented tools, with Argo CD favoring a UI-driven, hub-capable model using ApplicationSets, SSO, and AppProjects, while Flux emphasizes lightweight, composable controllers managed through Git, Kustomize, and Terraform. Neither provides built-in policy enforcement or a complete approval audit trail, so the recommended approach combines generated configuration from a central control-plane repository, admission controls such as Kyverno or OPA Gatekeeper, and separate approval mechanisms such as protected pull requests, GitHub Environments, or commercial pipeline tools. A defensible audit record joins Git and pull-request history, Kubernetes and cloud audit logs showing actual changes, and approval data tied to individual identities, ideally correlated through commit identifiers. Commercial platforms such as Harness offer integrated policy, approval, and audit capabilities on top of Argo CD, while internal developer platforms such as Qovery are presented as complementary control layers for application self-service, environment-level permissions, and centralized deployment history, leaving Argo CD or Flux to manage infrastructure components and cluster reconciliation.
Oct 02, 2026
4,092 words in the original blog post.
How to Migrate GPU Inference to a European Sovereign Cloud: Providers, Costs, and the Platform Layer
Migrating GPU inference to a European sovereign cloud requires separating GPU capacity selection from the deployment platform layer, with providers such as Scaleway, OVHcloud, Nebius, IONOS, StackIT, and T-Systems offering varying H100, H200, or newer GPU options. The piece argues that hosting workloads in a European region operated by a US hyperscaler provides data residency but not complete jurisdictional sovereignty because US parent companies may remain subject to the CLOUD Act, while European providers can better align legal control, operations, and data processing with EU requirements. It states that inference performance should generally remain comparable because the underlying NVIDIA hardware is identical, and EU-based serving can reduce latency for European users, although interconnect quality, storage speed, GPU availability, and autoscaling capabilities require validation. Recommended migrations use portable containers, Kubernetes, NVIDIA GPU Operator, open serving runtimes such as vLLM, TGI, or Triton, EU-based supporting services, infrastructure as code, and dual-running production traffic before cutover. Published European GPU prices are described as competitive with AWS, but utilization, batching, quantization, scale-to-zero policies, and reservation planning have greater effects on total cost than hourly rates alone. The article presents cloud-agnostic platforms such as Qovery, alongside tools including KServe, SkyPilot, and NVIDIA Run:ai, as ways to preserve deployment workflows across providers while maintaining a direct cloud contract, and emphasizes documenting every component, subprocessor, access path, log, backup, and control-plane interaction to demonstrate sovereignty to auditors.
Oct 01, 2026
5,399 words in the original blog post.
European organizations are selectively relocating certain workloads from AWS, Azure, and Google Cloud in response to concerns about U.S. extraterritorial access laws such as the CLOUD Act, EU requirements under the Data Act, DORA, and NIS2, provider-concentration risk, and less predictable egress, licensing, and committed-spend costs. Rather than pursuing full cloud exits, they commonly move regulated personal, health, citizen, public-sector, identity, collaboration, backup, CI, and non-production workloads while retaining deeply integrated analytics, machine-learning, and global delivery services on hyperscalers. Destinations include European providers such as OVHcloud, Scaleway, IONOS, Hetzner, StackIT, and Aruba; sovereign-cloud offerings from U.S. hyperscalers; self-managed Kubernetes on European infrastructure; and hybrid architectures that place regulated data in EU-controlled environments while retaining other compute where practical. The text argues that waived exit-egress fees and forthcoming Data Act restrictions on switching charges have reduced transfer-cost barriers, leaving migration engineering, replacement of managed services, and operational staffing as the main expenses. It recommends classifying workloads by risk, reducing proprietary dependencies, using Kubernetes and common deployment workflows for portability, beginning with non-production systems, and regularly testing documented provider-exit plans to support regulatory and operational resilience.
Oct 01, 2026
4,742 words in the original blog post.
France’s HDS certification is required for organizations hosting personal health data on behalf of third parties and is issued by COFRAC-accredited certification bodies under the ANS framework, while the ANS register is presented as the authoritative source for verifying certified legal entities, service scopes, regions, sites, and expiry dates. As of October 2026, commonly used certified platforms include OVHcloud, Scaleway, 3DS Outscale, Clever Cloud, Scalingo, Exoscale, NumSpot, Cloud Temple, AWS, Microsoft Azure, and Google Cloud, although certification applies only to specified activities and services rather than an entire provider catalogue. The framework distinguishes six hosting activities, covering physical infrastructure, virtual infrastructure, platform hosting, system administration, and backups, meaning teams using certified IaaS or managed Kubernetes may still retain responsibility for operating platforms, administering systems, and managing backups. HDS is separate from GDPR, ISO 27001, SecNumCloud, and HIPAA; SecNumCloud additionally addresses sovereignty and protection from extraterritorial law. The discussion advises health-tech teams to examine certificates and annexes, verify exact regions and services, establish contractual and audit evidence, avoid real patient data in non-production environments, and select managed PaaS, self-managed Kubernetes, or managed cloud services based on their ability to assume operational and compliance responsibilities. Qovery is described as a non-certified deployment control layer that operates within a customer’s existing cloud account or Kubernetes cluster, rather than replacing an HDS-certified host or the customer’s own obligations.
Oct 01, 2026
4,243 words in the original blog post.