October 2026 Summaries

13 posts from Aiven

Filter
Month: Year:
Post Summaries Back to Blog
Aiven Managed Agents enable users without coding experience to build AI agents by describing tasks in everyday language, connecting required integrations, testing the result, and deploying agents on demand or on schedules. Agents combine an AI model, system instructions, a recurring task prompt, and tools provided through Model Context Protocol integrations such as Notion, Slack, GitHub, Jira, web search, and Aiven services. The example creates a Halloween-themed Slack bot that reads weekly Notion tasks and posts a brief ghost-themed update every Monday, illustrating how Aiven generates editable instructions and tool selections from a written description. The process requires requesting limited-availability access, configuring integrations and least-privilege permissions, running real tests in a safe destination, refining prompts or instructions as needed, and selecting a deployment schedule. Aiven manages the underlying infrastructure while allowing users to choose models, communicate with agents through chat, schedule recurring runs, and scope access through role-based MCP tokens.
Oct 31, 2026 1,398 words in the original blog post.
Aiven Runtime offers an Apache Airflow template for scheduling and orchestrating workflows beyond basic cron-style jobs, using Python-defined directed acyclic graphs (DAGs) to manage task order, dependencies, conditions, dynamically generated tasks, and reactions to external data. The template deploys Airflow in Standalone mode through a Compose file and Dockerfile, combining the web server, scheduler, triggerer, and related components in one container while using PostgreSQL to store workflow state and the LocalExecutor to avoid requiring Redis, Valkey, or Celery. DAG files remain in a version-controlled repository, and users can deploy examples such as an Astronaut ETL workflow through Aiven Runtime, then monitor them in Airflow’s web interface. The setup is intended for cloud-based experimentation, development, and moderate workloads, including scheduled data pipelines and automation of Aiven Runtime applications through the Aiven API. Its main limitations are single-machine execution without distributed workers, a development-oriented Simple Auth Manager that stores credentials in the container, and ephemeral logs unless external logging is configured; larger production deployments are instead recommended to use Kubernetes and the Airflow Helm chart.
Oct 28, 2026 1,188 words in the original blog post.
Josep Prat, a Senior Engineering Director at Aiven and maintainer of Apache Kafka, describes how open-source maintainership is earned by individuals through community contributions rather than owned by their employers. He says this independence makes an employer’s treatment of open source essential to his decision to remain there, contrasting Aiven’s approach with a previous workplace where permission to contribute to open source was difficult to obtain. Although Aiven dissolved its dedicated Open Source Program Office in October 2024 and moved its members into product teams, Prat views the change as distributing open-source work more broadly rather than ending it. He notes that Aiven, despite being smaller than major technology companies, employs substantial Kafka maintenance contributors, and emphasizes that maintainers must prioritize the project’s interests over those of any employer to retain community trust.
Oct 27, 2026 798 words in the original blog post.
At All Things Open 2026, conversations among open-source contributors, technology companies, and Aiven representatives suggested that community-driven software remains important despite concerns that AI could encourage contributors to work in isolated forks. Panelists and project maintainers emphasized that contributors, upstream collaboration, and transparent security processes remain central to projects such as Valkey, including efforts to manage CVE reports and coordinate with vendors and AI providers. Attendees were more focused on dependable data storage, transfer, governance, and infrastructure than on AI coding trends, reflecting the growing data demands created by AI workloads. Aiven presented its open-source-based services, tools, and deployment options as ways for organizations to build durable data systems while retaining flexibility across clouds, hardware, models, and agent environments. The discussion also highlighted the risks of depending on a single AI model or provider, advocating interoperable approaches such as a multi-model gateway and architectures that use LLMs for development rather than as indispensable production dependencies.
Oct 27, 2026 963 words in the original blog post.
Logical replication can enable PostgreSQL major-version upgrades or migrations from Aiven, RDS, Cloud SQL, Azure, or self-managed servers to Aiven with only seconds of application downtime, while preserving the original server as a rollback option. Unlike physical replication or in-place pg_upgrade, it transfers decoded table-row changes across versions and providers, allowing a new target database to be built, synchronized, verified, and activated after a brief cutover. The process requires copying the schema in advance, identifying tables that need replica identities, monitoring lag and retained WAL, verifying data with counts and hashes, freezing or carefully coordinating DDL changes, copying large objects separately, and synchronizing sequences before PostgreSQL 19. Aiven provides logical WAL settings, helper functions, and a one-command migration path through aiven-db-migrate when prerequisites are met, while accompanying scripts support more customized migrations. PostgreSQL 18 improves conflict reporting and replication operations, and PostgreSQL 19 is expected to add native sequence replication, though DDL and large-object migration will still require separate handling; in all cases, removing replication slots after migration is essential to prevent unbounded WAL retention on the source.
Oct 27, 2026 5,747 words in the original blog post.
Aiven Managed Agents is a limited-availability managed platform for deploying production AI agents near an organization’s data infrastructure, addressing challenges such as unmanaged local agents, unclear costs, broad credentials, missing operational ownership, and vendor lock-in. Built on an open-source core and supporting any model, it provides durable execution, persistent memory stored in the customer’s PostgreSQL environment, MCP-based connections to databases, APIs, and services, and triggers including manual requests, schedules, webhooks, and events. The service includes access controls, secrets management, audit logs, tracing, and per-agent cost tracking, while Aiven operates scaling, upgrades, security, and monitoring. Users can create agents from service-specific templates, natural-language descriptions, or manual configurations, connecting integrations such as Slack, GitHub, Jira, Sentry, and custom MCP servers. Typical applications include SRE agents that monitor, investigate, and remediate approved infrastructure issues, internal operations automation, and customer-facing embedded product agents; an example slow-query agent analyzes PostgreSQL query statistics and sends a Slack report.
Oct 20, 2026 1,621 words in the original blog post.
Production AI agents require more than core reasoning logic, needing adaptable infrastructure for model access, orchestration, observability, retrieval, user interfaces, and automation. The proposed open, swappable stack uses LiteLLM to abstract model providers, LangGraph to define agent workflows, Temporal to resume failed tasks, Langfuse to trace decisions and resource use, Streamlit and Hasura for user and API access, n8n and Airflow for scheduled or event-driven workflows, and pgvector or OpenSearch for data retrieval. Aiven Runtime is presented as a platform offering deployable templates for these components alongside organizational data, aiming to reduce integration work and avoid dependence on proprietary tools. For a support-ticket agent, the suggested setup retrieves ticket context, drafts responses, routes cases requiring human review, and provides resilience, monitoring, and an interface for support staff.
Oct 20, 2026 873 words in the original blog post.
Inference, rather than model training, accounts for most engineering work in deployed LLM systems and resembles familiar data-engineering tasks such as moving data, managing latency, caching, monitoring, and evaluating results. A documentation assistant typically combines a lightweight encoder that converts document fragments and questions into embeddings for semantic search with a larger decoder that generates answers from retrieved text, often using hybrid vector and keyword search to handle both meaning and exact technical identifiers. This retrieval-augmented generation approach keeps answers current, supports citations, and is generally more suitable for changing documentation than fine-tuning a decoder, which tends to improve style more reliably than factual recall and cannot directly cite sources. The discussion explains tokens, context limits, cosine similarity, embedding spaces, multilingual retrieval trade-offs, and the need to use the same encoder for documents and queries, while emphasizing that retrieval quality, source conflicts, reindexing, caching, and observability determine system reliability. It also covers low-temperature sampling for technical support, quantisation as a way to reduce a 7B model’s memory footprint from roughly 14 GB to about 4 GB while improving generation speed at some cost to complex reasoning, and performance measures such as time to first token, tokens per second, throughput, streaming, and p95 or p99 latency. Because retrieval does not eliminate hallucinations, the proposed system treats groundedness as a measurable property in which claims trace to retrieved sources and uses refusals when documentation lacks a relevant answer.
Oct 19, 2026 5,488 words in the original blog post.
Aiven presents a configurable log-management architecture centered on OpenSearch for searching and visualizing logs, with Apache Kafka added as an optional buffer-and-routing layer when organizations need multiple destinations, resilience to indexing bursts, or the ability to replay logs after mapping and parsing changes. In this model, applications and log forwarders write once to Kafka, which retains events while consumers send copies to OpenSearch, object storage, ClickHouse, or other destinations without requiring application redeployments. Aiven provides managed Kafka and OpenSearch services, built-in integrations for forwarding Aiven service logs to either platform, an OpenSearch sink connector, OpenSearch Dashboards, and infrastructure-management options including Terraform, APIs, CLI tools, and Kubernetes operators. OpenSearch alone may suit stable, single-destination log workloads where occasional loss and replay are acceptable, while Kafka is recommended for bursty volumes, multi-team ownership, durable retention, multi-destination routing, and uninterrupted ingestion during OpenSearch maintenance.
Oct 16, 2026 1,587 words in the original blog post.
Aiven has launched its open-source AI Gateway to help organizations run AI coding agents and other high-volume workloads across multiple model providers, with an emphasis on lower-cost open models. The announcement argues that always-on, increasingly autonomous engineering agents could consume hundreds of millions of tokens daily, making proprietary models prohibitively expensive at scale compared with alternatives such as DeepSeek V4.1 Flash. It contends that open models are no longer merely budget options, citing their growing share of token usage and ecosystem-driven improvements similar to those behind technologies such as Kafka and PostgreSQL. During early access, Aiven is offering free usage and $300 credits to approved organizations, while planned additions include model routing and failovers, usage-based billing, configurable budgets and alerts, and bring-your-own-key support.
Oct 15, 2026 527 words in the original blog post.
Observability is presented as a data-management challenge requiring different backends for metrics, logs, and traces rather than a single all-purpose product. Aiven for Metrics, based on Thanos and compatible with Prometheus, supports alerting and long-term time-series analysis but is less appropriate for high-cardinality labels; Aiven for OpenSearch enables full-text log, audit, and security-event searches, though indexing can become costly at large volumes; and Aiven for ClickHouse supports rapid SQL analysis of high-volume, high-cardinality structured logs and traces, but does not replace search or native alerting tools. A common architecture collects telemetry through OpenTelemetry, buffers it in Kafka, routes it to the appropriate backend, and uses Grafana for unified visualization. The approach emphasizes Aiven’s use of upstream open-source Thanos, OpenSearch, and ClickHouse to support portability, local testing, and reduced vendor lock-in, while allowing organizations to adopt services incrementally as telemetry needs expand.
Oct 10, 2026 1,453 words in the original blog post.
Aiven’s open-source Model Context Protocol server gives AI assistants such as Claude, Cursor, and VS Code structured, permission-controlled access to live PostgreSQL services, allowing them to inspect schemas, query data, analyze logs and WAL status, recommend configuration changes, and perform approved operations. Using an Apache Airflow metadata database as an example, the account shows how the assistant identified a standard Airflow schema, reviewed successful DAG runs, distinguished harmless monitoring-log noise from a collation warning, and assessed an inactive replication slot as a future WAL-retention risk rather than an immediate issue. It also recommended connection pooling through PgBouncer after observing potentially constrained connection limits, while clearly separating documentation-backed guidance from its own inference. The demonstration then used the MCP to provision Kafka resources and configure a Kafka Connect JDBC source connector that transforms Airflow’s JSON data into individual astronaut-observation messages streamed from PostgreSQL to Kafka as new DAG runs occur. Safety controls include read-only mode, service-scoped tools, approval requirements for changes, and inheritance of the user’s existing Aiven permissions, while planned support for additional services such as OpenSearch and ClickHouse is intended to expand the platform’s agent-assisted workflows.
Oct 06, 2026 3,066 words in the original blog post.
A Kafka defect affecting versions 3.9 through 4.3 could leave removed controllers recorded in cluster metadata as “ghost” controllers, preventing clusters from enabling newer features such as Share Groups or upgrading their metadata version. Although Kafka 4.4 introduced a native controller-unregistration command, affected clusters faced a circular dependency because outdated ghost-controller registrations blocked the metadata upgrade required to use that command. Aiven developed patches for Kafka 4.0–4.3 that reuse existing controller-registration metadata to mark permanently offline, non-voter controllers as decommissioned, allowing patched clusters to ignore them during feature and metadata-version validation without introducing incompatible metadata records. The patch was also carried forward into version 4.4 so that clusters could transition safely to Kafka’s official unregistration mechanism and eventually remove the obsolete registrations. Aiven plans to deploy the fix automatically through maintenance updates for users of its managed platform, while patches are available for organizations maintaining their own Kafka forks.
Oct 01, 2026 1,228 words in the original blog post.