August 2026 Summaries

ClickHouse has introduced an in-product onboarding flow for its Managed Postgres offering, which combines NVMe-backed PostgreSQL for transactional workloads with ClickHouse for real-time analytics and managed replication between the two. The new experience replaces an initially empty operational dashboard with a stateful, outcome-focused path designed to help users work with data quickly: provision a Postgres service, migrate or ingest data through ClickPipes, run an initial query in the SQL Console, and configure change data capture replication into ClickHouse for analytics. Users can migrate an existing Postgres database, connect through standard language clients, load sample data, or create tables manually, while an AI Assistant can help generate SQL. The final analytics step supports both direct ClickHouse queries and the pg_clickhouse extension, which lets users access ClickHouse tables through PostgreSQL syntax; the company cites benchmark improvements of roughly three to seven times on a one-million-row dataset. The four-step design intentionally emphasizes early hands-on value and progress rather than attempting to teach every platform feature.

QuintoAndar, Latin America’s largest housing platform, is rebuilding its customer data platform on ClickHouse Cloud to process nearly 900 million raw monthly events from 14 million users and support roughly 300 million API requests. The change replaces a complex Lambda architecture that relied on separate Postgres and Databricks pipelines, reducing duplicate-event processing, costly Delta Lake merges, and maintenance of multiple codebases while preserving low-latency access for internal services and analytics. Although Hightouch Customer Studio did not natively support ClickHouse, the team used ClickHouse Managed Postgres and its pg_clickhouse extension to expose ClickHouse tables through a Postgres-compatible interface, allowing Hightouch to run analytical queries in ClickHouse while retaining its operational sync-state tables in Postgres. This approach enabled QuintoAndar to keep its existing marketing activation tooling, maintain access controls, integrate with Trino, Spark, and Superset, and make ClickHouse available to other third-party tools that support PostgreSQL.

ClickHouse has become a widely adopted storage and query engine for observability systems, particularly for high-volume logs, traces, and event-style metrics, due to its columnar architecture, compression, scalable distributed execution, SQL interface, and ability to handle high-cardinality data efficiently. The discussion argues that its strengths enable lower-cost retention of full-fidelity telemetry, flexible investigation workflows, and consolidation of multiple data types, while recent full-text indexing improves log search capabilities and emerging PromQL support seeks to address conventional metrics use cases. However, it emphasizes that a database alone cannot deliver effective observability, which also depends on collection, schema design, correlation, interfaces, alerting, and operational workflows; ClickStack is presented as an integrated OpenTelemetry-based platform intended to provide these layers. ClickHouse may be less suitable for small teams seeking a fully managed, low-maintenance experience or for observability vendors with specialized needs that justify owning a proprietary datastore, and its Prometheus compatibility remains experimental and incomplete. As AI agents increasingly query telemetry autonomously, the need for low latency, concurrent queries, expressive SQL, and long-term unsampled data may reinforce ClickHouse’s advantages, although the authors frame its current position as a strong but unfinished role in an evolving observability ecosystem.

ClickHouse’s .NET driver has advanced significantly since its 1.0 release, adding type-safe POCO-based inserts and queries, JSON-column serialization, configurable parameter type resolution and formatting, and read-value conversion to give developers more control over data handling. Version 1.3 also expands support for multidimensional arrays, ValueTuples, server-side Identifier parameters, per-query compression settings, application tagging, and improved schema-probe avoidance and buffer management to reduce round trips and garbage-collection pressure. Correctness fixes include improved DateTime time-zone inference, fixed-offset zones, composite serialization, Variant null handling, and enum parsing. The project now includes an official Entity Framework Core provider with LINQ translation, migrations, ClickHouse-specific DDL configuration, and bulk inserts, alongside integrations for Serilog, .NET Aspire, and Semantic Kernel. Future work includes performance-focused version 1.4 features such as reduced allocations and pluggable compression, followed by a native TCP client intended to use ClickHouse’s columnar protocol for higher throughput, progress reporting, and persistent connections.

PostgreSQL 19, currently in beta, adds the `WAIT FOR LSN` command to let sessions wait until write-ahead log progress reaches a specified location, enabling read-your-writes consistency on asynchronous replicas without requiring every primary commit to wait for replica replay. After a write, an application or intermediary can obtain the primary’s current WAL insert LSN and issue `WAIT FOR` on a standby before reading, ensuring the relevant changes have been replayed while avoiding application-side polling. The command supports replay, flush, and write wait modes, optional timeouts, and a `NO_THROW` option that returns statuses such as success, timeout, or not in recovery. It must be executed as a top-level, snapshot-free utility command because holding a snapshot while waiting could block WAL replay and create a self-deadlock. The feature preserves asynchronous write performance by imposing delay only on reads that require freshness, and it may be especially useful for connection poolers or protocol-aware proxies that can automatically track LSNs and route consistent reads to replicas.

ClickGap is an autonomous QA agent developed by ClickHouse to review merged pull requests in its performance-critical C++ database, where high merge volume and extensive existing CI still leave gaps involving untested behavior, regressions, and performance losses. In its first five months, the bot filed roughly 500 issues and opened about 200 test-coverage pull requests, with many findings resolved through linked fixes, after reproducing defects such as unintended semantic changes, memory-spill failures, ineffective skip indexes, and query slowdowns. The system validates findings through real builds, executed reproducers, coverage mutation tests, adversarial model review, deterministic checks, and ten evidence gates before it can create a public issue or pull request. It also bisects regressions, tests affected supported releases, and avoids attributing defects to a pull request unless the evidence supports a causal connection. ClickGap’s design emphasizes maintainer trust, incorporating safeguards against spam, false positives, flawed CI interpretation, excessive filing, and repeated mistakes, while using human feedback to update a curated memory system. The project argues that specialized knowledge, measurable validation, deterministic automation, concise reports, and continuous outcome-based evaluation are more important than simply making an AI agent appear capable.

Suprema Gaming, a Brazilian gaming group serving more than 500,000 players across 70 countries, migrated most of its analytics workloads from Snowflake to ClickHouse Cloud to support real-time, AI-driven operations. The company replaced batch-oriented pipelines that left data up to four hours old with change-data-capture ingestion and dbt-based transformations, achieving roughly one-minute p95 data lag, millisecond query responses, 27-times-faster mart reads, and a reported 62% reduction in warehouse spending. Before each production cutover, Suprema validated metrics against source systems to the cent, using AI-assisted discrepancy detection and version-controlled dbt tests to identify issues such as ingestion gaps and distorted test-account data. Its platform incorporates role-based access, query-time PII masking, auditing, and least-privilege controls so AI agents can be treated as governed data consumers. The resulting agentic system, led by the conversational assistant Senna and supported by specialized agents for modeling, validation, development, and freshness monitoring, provides business users with live, cross-product data while maintaining traceability through monitored conversations, SQL queries, costs, and latency.

ClickHouse Cloud’s LogHouse observability platform now ingests 50 million OpenTelemetry events per second and stores 177 PiB of uncompressed data, requiring an ingestion architecture that can withstand unpredictable bursts and database backpressure without data loss or excessive permanent capacity. Initial agent-to-gateway collectors relied on in-memory queues and failed during outages, while a local disk-based write-ahead log improved durability but created slow, FIFO backlog recovery, operational complexity from persistent volumes, and delays to fresh telemetry. Rather than operate Kafka or another streaming platform solely for this use case, the team built a stateless OpenTelemetry collector design that routes data directly to ClickHouse under normal conditions and automatically fails over to S3-compatible blob storage when ClickHouse is unhealthy. Blob-storage event notifications are delivered through lightweight queues such as SQS, which separate catch-up collectors use to replay overflow data into ClickHouse after recovery, ensuring live data is not blocked behind historical backlog. A staging outage test showed that traffic switched automatically to S3 and later resumed direct ClickHouse ingestion while catch-up collectors drained stored data independently. Although the approach requires region-specific buckets, queues, notifications, and more complex collector configuration, it has scaled from 10 million to 50 million events per second, reduced operational burden and costs, and is being extended to Clickstack Cloud’s managed OTLP ingestion service.

ClickHouse’s monthly update highlights the 26.7 release, which introduces more efficient hash joins, automatic multi-table join ordering, positional text indexes for faster phrase searches, and the new EXPLAIN ANALYZE capability, alongside performance contributions from community member Manuel Raimann. The company also announced that database researcher Andy Pavlo has joined to establish ClickHouse Labs, an applied research group focused on advancing ClickHouse and its Managed Postgres service. Other featured developments include an interactive ClickBench playground supporting more than 100 database systems, Managed Postgres improvements in onboarding, notifications, observability, backups, and extensions, and technical guidance on TTL configuration, mixed Postgres and ClickHouse architectures, AI-agent reliability, observability, and vector search. The update also promotes a global Open House Roadshow, virtual training, and numerous regional events across the Americas, EMEA, and APAC.

ClickHouse Managed Postgres treats supporting services such as connection poolers, backup agents, metrics exporters, Prometheus, log collectors, and watchdogs as part of the database failure model because their resource consumption can impair PostgreSQL availability. While PostgreSQL shared buffers are protected through boot-reserved huge pages, other database memory needs share system capacity with auxiliary processes, so supporting Go services are grouped in a cgroup v2 slice with GOMEMLIMIT runtime targets, memory.high pressure controls, and memory.max hard limits to contain leaks and ensure any cgroup OOM event does not select PostgreSQL processes. Resource controls also extend to reduced CPU weights for backups, fixed backup buffer sizes, log collector memory limits, curated metric exports to prevent high cardinality, disk thresholds, and session termination during disk-full emergencies. The disk watchdog preserves replication and monitoring users while ending ordinary application sessions, aiming to maintain recovery and observability functions under pressure. These default controls isolate supporting workloads while keeping backup, replication, monitoring, and PostgreSQL within defined resource budgets.

Shopify replaced separate metrics, logs, and tracing vendors with Observe, an internally built observability platform powered by self-hosted ClickHouse that unifies metrics, logs, traces, profiles, and exceptions. Designed to support a global infrastructure of roughly 500 Kubernetes clusters and 1.5 million pods, the platform ingests about 50 million events per second normally and up to 100 million during Black Friday–Cyber Monday, while making telemetry queryable in under a minute. Shopify reported a 16-fold out-of-the-box query-performance improvement with ClickHouse, exceeding 30-fold at peak, while gaining more predictable costs and control over its technology roadmap. To manage durability, flexible schemas, correlations, and ongoing migrations, the team uses Kafka buffering, materialized views, typed columns, metadata indexes, and an in-house Kubernetes operator, alongside compression and storage optimizations. Shopify is now expanding ClickHouse use into continuous profiling, AI-assisted internal analysis, and merchant-facing analytics, although its engineering director said that, if starting today, the company would likely consider ClickHouse Cloud to reduce the operational burden of self-hosting.

Sai Srirampur’s POSETTE 2026 talk examines how storage constraints can contribute to PostgreSQL scaling issues such as slow ingestion, inconsistent tail latency, autovacuum backlog, checkpoint contention, and logical replication lag once workloads exceed available memory and require frequent physical I/O. In a benchmark using identical PostgreSQL clusters with a 3.3-billion-row table and concurrent random updates, local instance-store NVMe achieved a median 16,030 TPS versus 1,734 TPS on baseline gp3 EBS configured for 3,000 IOPS, while median transaction latency declined from 36.9 ms to 4.0 ms; NVMe also completed vacuuming and replication-backlog recovery faster in this specific test. Profiling attributed most of the difference to lower page-read and WAL fsync wait times rather than CPU work, allowing NVMe-backed systems to keep more CPU capacity productive. Because local NVMe data is lost when a node fails, the proposed production design combines it with quorum-based synchronous replication across availability zones, independently stored base backups and continuous WAL archiving for point-in-time recovery, and capacity planning based on available instance storage.

Sony LIV rebuilt its analytics infrastructure around ClickHouse Cloud to support real-time quality-of-service and quality-of-experience monitoring for its large-scale OTT platform, particularly during unpredictable live sports traffic spikes. Its previous fragmented stack of batch pipelines, Elasticsearch, BigQuery, and separate telemetry stores struggled with latency, cost, scaling, and correlating data across playback, CDN, application, clickstream, and advertising systems. The new centralized architecture streams events from client applications, streaming services, CDN infrastructure, and ad platforms through Amazon Kinesis and ClickPipes into ClickHouse, which supports operational dashboards, product analytics, SRE troubleshooting, ad analytics, and playback-quality monitoring. Sony LIV reports that queries previously requiring tens of seconds or minutes now run in sub-second to low-single-digit-second times, improving incident response and visibility into buffering, playback failures, regional degradation, device issues, and CDN anomalies. The company also cites improved storage economics, reduced operational fragmentation, and better handling of major-event traffic, while planning to expand ClickHouse use for behavioral analytics, predictive autoscaling, anomaly detection, enhanced observability, and real-time advertising intelligence.

PostgreSQL 19, still in beta at the time of writing, introduces a range of monitoring and observability enhancements intended to make lock contention, WAL activity, maintenance operations, replication behavior, and transaction wraparound risks easier to diagnose. Key changes include enabling lock-wait logging by default, allowing log verbosity to be configured by process type, separating autoanalyze logging from autovacuum logging, and adding full-page-image byte metrics to WAL statistics, query plans, and maintenance logs. New wait events expose WAL write stages and file-, pipe-, and program-based COPY I/O, while messages received through logical replication, foreign data wrappers, or dblink now receive standard local log formatting. The release also adds pg_get_multixact_stats() for inspecting multixact usage and disk consumption, raises the advance warning threshold for transaction and multixact wraparound from 40 million to 100 million transactions, and introduces additional system views. Monitoring-tool maintainers are advised to update log parsers, metric collectors, wait-event dictionaries, configuration validation, and alert thresholds to accommodate the new behavior and fields.

ClickHouse Terraform provider version 3.25.0 introduces beta support for managing ClickStack observability resources, including dashboards, alerts, sources, saved searches, connections, and webhooks, across both ClickHouse Cloud Managed ClickStack and self-hosted deployments. By storing these configurations in version control, teams can review changes, reproduce environments, and deploy resources through standard Terraform planning and apply workflows rather than manually maintaining UI-based copies. ClickStack support was incorporated into the existing ClickHouse/clickhouse provider to avoid a separate provider and to preserve consistent releases and authentication, while self-hosted installations use their own endpoint and personal API key. API improvements, including stable identifiers, structured validation errors, and dashboard JSON validation during Terraform planning, support reliable resource management, although validation may be deferred to apply when older ClickStack versions lack the endpoint. Users can create new dashboards using JSON-based Terraform resources, import existing dashboards individually or generate configurations from exports, but should avoid simultaneously managing the same dashboard in Terraform and the UI because UI edits are not reported as drift and may be overwritten.

ClickHouse and Hud have announced an integration intended to help engineering teams safely develop, deploy, and maintain increasing amounts of AI-generated code by combining service-level observability with function-level runtime analysis. ClickStack uses OpenTelemetry traces, logs, and infrastructure data to identify affected services, deployments, and endpoints, while Hud’s Runtime Code Sensor automatically captures detailed behavior within application functions, linking production issues to source code and recent changes without manual instrumentation. Shared trace IDs connect the platforms’ datasets, allowing engineers or coding agents to move from broad operational alerts to code-level causes, or from a problematic function to relevant logs and traces. The combined approach supports pre-deployment risk assessment based on real production behavior, post-release regression monitoring, automated triage and remediation workflows, and informed rollback or fix decisions, with the goal of making AI-assisted software delivery faster while retaining operational confidence.

Recent ClickHouse Grafana plugin updates, developed with Grafana Labs and community contributors, aim to make common observability investigations faster without requiring SQL while retaining one-click access to editable generated queries. Version 4.20 or later introduces single-table data source configuration and a compact query mode for searching logs, applying type-aware filters, reviewing query history, and viewing live SQL, with upcoming support for custom fields. Users can now investigate logs interactively through expandable OpenTelemetry attributes, clickable inclusion or exclusion filters, highlighted-text filtering, and improved dashboard ad hoc filters across Map columns and multiple tables. The plugin also includes ready-to-use OpenTelemetry Logs Explorer, Traces Explorer, and service dashboards, plus guided editors that automatically generate SQL for dashboard variables and deployment or rollback annotations. Additional changes improve schema detection, query efficiency, caching, trace lookup, Grafana 13 compatibility, SQL validation, OpenTelemetry instrumentation, and propagation of Grafana user identity to ClickHouse for policies, quotas, and query attribution.

bitdrift has joined ClickHouse’s House Mates program as a mobile observability partner for ClickStack, integrating mobile-device telemetry with ClickHouse-based backend observability. ClickStack combines ClickHouse storage, the HyperDX interface, and an OpenTelemetry Collector to collect, query, visualize, and alert on logs, metrics, and traces, while bitdrift’s Capture SDK for iOS, Android, and React Native records detailed unsampled telemetry locally in a fixed-size device buffer and retrieves relevant sessions on demand. The integration uses OpenTelemetry trace context and shared trace IDs to connect mobile events such as user actions, crashes, slow screens, and network requests with backend distributed traces, services, databases, and infrastructure activity in ClickStack. This allows engineers to investigate an issue from an affected user’s device session through the corresponding backend request path without manually correlating timestamps across separate tools. The integration works with self-managed ClickStack and ClickHouse Cloud, with Android tracing supported through OkHttp integrations and iOS tracing through URLSession.

iFood, Latin America’s largest food delivery platform, rebuilt its in-house security platform around ClickHouse Cloud after its Databricks-based approach became too costly and slow for high-volume log ingestion, long-term retention, and incident investigations across more than 130 TB of data. While Databricks remains central to the company’s lakehouse for BI and data science, ClickHouse Cloud delivered 9–16 times faster queries at an expected 40–50% lower cost, improved data freshness from hourly batches to two-to-ten-minute ingestion through AWS S3 and ClickPipes, and enabled security alerts to run about every 10 minutes. The security team uses SQL, Querybook notebooks, ClickStack dashboards, and Langfuse to support investigations and monitor AI workflows, with more than 500 employees trained on Langfuse company-wide. ClickHouse has also enabled agentic threat hunting, allowing AI sub-agents to investigate multiple hypotheses across tens of terabytes of historical logs in parallel, reducing work that previously took an analyst a week to about two hours. iFood plans to expand the platform by migrating CDN and API logs from OpenSearch, extending retention from seven days to six months at roughly one-quarter of the cost.

pg_clickhouse v0.10.0 advances ClickHouse pushdown support for PostgreSQL analytic workloads, increasing full TPC-H query pushdown coverage from 12 to 16 of 22 queries and delivering major speedups for queries such as Q17, which fell from 32.7 seconds in an earlier version to 37 milliseconds. The release adds correlated subquery and safe IN/NOT IN pushdown, including safeguards that preserve PostgreSQL’s three-valued NULL semantics despite ClickHouse’s different default behavior, though this correlated-subquery support requires ClickHouse 25.8 or later. It also broadens support for functions, aggregates, partitionwise aggregation, date and time operations, JSON, arrays, window functions, and other SQL features, while making builtin function pushdown explicitly mapped to avoid semantic inconsistencies. Driver changes replace an earlier C++ client with a streaming C-based library, unify binary encoding across HTTP and binary paths, improve memory use, add compression and TLS settings, expand type support, and resolve concurrency and reliability issues. New query and procedure interfaces support arbitrary row-returning queries and effect-only commands against configured ClickHouse servers, while six TPC-H queries remain blocked mainly by join-tree deparsing limitations; future work also includes broader function coverage, UPDATE and DELETE support, and UNION pushdown.

ClickStack’s June–July update focuses on trace usability, broader metrics integration, dashboard enhancements, and automation capabilities shaped largely by user feedback. Trace viewing now includes stable per-service colors, improved handling of short spans, expandable trace controls, a minimap for navigating large traces, and visible OpenTelemetry span links that let users move between related traces. Metrics improvements include quantile and sum support for exponential histograms, a more efficient default metrics-table schema, experimental connections to external Prometheus-compatible endpoints, and an opt-in Datadog receiver that supports gradual migration from existing observability tooling. Dashboards gain cascading filters, categorical bar charts, event-pattern tiles, kiosk mode, external links from tables, improved layout interactions, and service-map metric coloring, while alerts can require conditions to persist across consecutive evaluation windows before firing. Performance work routes autocomplete and filter-value queries through ClickHouse text indexes when available, and API and MCP server additions enable programmatic or agent-driven management of saved searches, webhooks, sources, metrics, dashboards, and alerts.

Musinsa, Korea’s largest fashion platform with 16.4 million members and operations in 13 countries, built a customer data platform in 2025 to create marketing audiences from customer profiles and behavioral data, later adding an AI assistant that answers marketers’ natural-language statistical questions while explaining its methodology. Its Audience Engine manages roughly 1.1 billion user-to-audience mappings, creating scaling, workload contention, and operational challenges in its self-hosted ClickHouse environment, where compute and EBS storage were coupled and multiple jobs competed on a single cluster. After moving to ClickHouse Cloud in AWS’s Seoul region, Musinsa retained much of its existing configuration while gaining managed operations, separated S3-backed storage and compute, workload-specific resources, automatic idling, and vertical autoscaling. The company reports reducing storage costs by 86.5% and total cost of ownership by up to 71.4%, while replacing Databricks Auto Loader and an external Spark cluster with ClickPipes for simpler, more scalable real-time ingestion.

Physical Intelligence develops robotics foundation models intended to generalize across environments, robot embodiments, and tasks, relying on large and diverse datasets including raw training data, operational records, annotations, metadata, and telemetry. As its data grew from tens to hundreds of billions of rows, a single Amazon RDS PostgreSQL instance could no longer efficiently support both strongly consistent transactional workloads and high-cardinality analytical queries, particularly because flexible JSONB-based annotations complicated querying. The company adopted a combined PostgreSQL and ClickHouse Cloud architecture, retaining PostgreSQL for ACID-compliant operational data while replicating data through ClickPipes into ClickHouse for OLAP, search, telemetry, and large-scale annotation analysis. ClickHouse features such as columnar storage, ReplacingMergeTree, materialized views, read-write isolation, and compute-storage separation improved query performance and scalability, enabling researchers to explore data through an internal go/data application that can identify whether specific concepts, tasks, or robot experiences appear across the dataset.

ClickHouse 26.7 introduces 61 features, 112 performance optimizations, and 329 bug fixes, with major work focused on query execution, joins, vector search, observability, federation, and usability. New ordered aggregation optimizations can stream and stop early for suitable GROUP BY, ORDER BY, and LIMIT queries, producing benchmarked reductions of up to 313× in runtime and 592× in peak memory, while join enhancements allow build-side keys to prune unread probe-side granules, compact hash-table row references, and add the DPsub join-ordering algorithm. Vector-search updates expand QBit with Int8 quantization, configurable dimension strides, randomized Hadamard transforms, and quantization codecs for approximate candidate selection followed by exact rescoring. Experimental position-aware text indexes accelerate phrase matching, and EXPLAIN ANALYZE now executes queries while reporting planning and execution times, data flow, memory use, and observed parallelism. Other additions include persistent Remote and RemoteSecure engines, unified URL access across storage backends, DateTime64 support from years 0000 through 9999, the groupFormat aggregate function, experimental SQL-defined executable UDF drivers, an expanded embedded SQL web workspace with version-matched local documentation, JIT-compiled regular expressions, and improved macOS performance and functionality.

ClickHouse Managed Postgres has introduced a range of post-beta improvements aimed at strengthening its unified Postgres and ClickHouse platform for transactional and analytical workloads. A redesigned onboarding flow supports sample data, PostgreSQL migration through ClickPipes, built-in SQL querying, replication to ClickHouse, and cross-system queries through pg_clickhouse. New notifications alert users when disk usage remains above 85%, with future alerts planned for replica lag, uptime, and failovers, while observability additions include searchable server logs, expanded service and query metrics, a Prometheus-compatible endpoint, and an OpenAPI slow-query API. Backup and restore performance has been improved for large deployments, and extension updates include pg_re2 for faster, compatible regular expressions, enhanced query pushdown in pg_clickhouse, and more efficient pg_stat_ch monitoring exports. Planned work includes Google Cloud Platform availability, managed maintenance, deeper OpenTelemetry support, improved configuration and tuning, cloud-object-storage data movement, and more native replication into ClickHouse.

WAL archiving in PostgreSQL copies completed write-ahead log segments to object storage for point-in-time recovery, but an archiving backlog can fill the WAL disk and trigger a database PANIC if writes outpace uploads. ClickHouse Managed Postgres addresses this risk with an automated data-plane mechanism that checks pending WAL segments every 15 seconds and progressively limits client write bandwidth through cgroup v2 I/O controls, while leaving reads and the archival, checkpointing, and logging paths uncapped so recovery can proceed at full speed. In a test using a deliberately rate-limited archiver, write caps activated as the backlog crossed 100, 500, and 1,000 segments, reducing pgbench throughput but keeping disk use below 33%; after archival throughput was restored, the uncapped archiver drained the 53 GiB peak backlog and the throttle automatically lifted once the queue cleared. The approach is intended to buy time during archival failures without requiring control-plane access or manual intervention, although its effectiveness varies with workload characteristics because some WAL generation may continue through uncapped internal processes.

ClickHouse Cloud has enhanced its autoscaling mechanism to provide more immediate resource management by integrating a reactive system alongside its existing periodic recommendation service. Originally, the autoscaling system relied on a fixed schedule to assess and adjust resource allocations, which sometimes resulted in suboptimal performance during unexpected workload spikes. To address this, ClickHouse Cloud implemented a fast-path solution using controller-runtime, a library often associated with Kubernetes, to react to real-time signals such as out-of-memory events. By leveraging ClickHouse for real-time analytics, the system can now quickly determine which services need scaling adjustments, ensuring faster response times and optimizing resource usage without creating bespoke message systems. The use of controller-runtime allows for efficient deduplication, backoff, and concurrency management, thus enabling a more responsive autoscaling process that remains robust and scalable. The new approach enhances ClickHouse Cloud's ability to adjust resources dynamically, aligning provisioning more closely with current workloads and improving the overall user experience.

Mercado Libre, a leading Latin American commerce and fintech platform, has significantly enhanced its observability capabilities by migrating its O11y events platform to ClickHouse Cloud, resulting in a 50x increase in query performance and up to 89% data compression. This migration enables the company to efficiently handle massive data volumes, scaling from 7 million spans per minute to 400 million, and provides high-cardinality filtering for identifiers like payment and user IDs. The platform, built on OpenTelemetry, offers end-to-end visibility of business flows, allowing for rapid troubleshooting and improved customer experience. At the Open House SF 2026 event, technical leader Daniel Da Rosa and software expert Francislei Reis highlighted the importance of tracing for comprehensive request views and process optimization. They shared insights into the architecture shift that moved heavy data operations into ClickHouse, leveraging materialized views and a dedicated lookup table to solve high-cardinality filtering challenges. The migration lessons emphasized the importance of profiling, benchmarking, iterative index design, and segmenting storage and compute clusters. Looking forward, Mercado Libre plans to enhance its observability platform with AI telemetry, proactive insights, and semantic queries, aiming for a future where observability transcends traditional dashboards.

ClickBench is an extensive benchmarking platform initially developed for testing ClickHouse and later expanded into a comprehensive open benchmark for analytical databases, encompassing around 100 database systems. This platform allows users to conduct performance and correctness comparisons using preloaded datasets of 100 million records across diverse database systems, ranging from relational to unconventional ones. The infrastructure, built on a foundation of automation and flexibility, uses shell scripts for database installation and benchmarking, with adjustments made for cloud SaaS databases. Despite challenges in maintaining and updating these scripts, ClickBench has evolved to support a consistent interface across systems, enabling comparative analysis and experimentation with different queries and datasets. The initiative also includes a ClickBench Playground for interactive engagement, although hosting numerous databases presents logistical and security challenges, mitigated by employing virtualization and resource allocation strategies. The platform serves as a valuable tool for database collectors and enthusiasts, facilitating the exploration of database behaviors and compliance, while also supporting the ongoing development and testing of ClickHouse.

Joining ClickHouse to lead the new ClickHouse Labs research team, the author, a former Carnegie Mellon University professor, aims to advance database technology by closely collaborating with ClickHouse engineers and industry partners. The author has tracked the development of ClickHouse since its inception in 2016, noting its pioneering features like vectorized query execution using SIMD, which were previously only available in closed-source systems. ClickHouse Labs will focus on scientific research and practical innovations, leveraging the strengths of both ClickHouse and PostgreSQL to explore a range of database management system (DBMS) challenges, including integration with emerging AI technologies. The team aspires to emulate the impact of influential industry research organizations such as IBM Research and Microsoft Research, and plans to accelerate the validation and production of existing ideas while investigating new hardware, algorithms, and strategies for DBMS development. The author envisions ClickHouse's robust relational model as a strong foundation for evolving alongside data-intensive workloads and is committed to training future database researchers as part of this initiative.

ClickHouse has launched ClickHouse Labs, a new research group led by prominent database researcher Andy Pavlo, to advance foundational database technology and openly share its findings with the community. As Vice President of Database Research, Pavlo will spearhead efforts to explore innovative data processing methods, efficient data structures, and high-performance computing to address challenges posed by AI on data-intensive workloads. ClickHouse Labs aims to push the boundaries of database systems for ClickHouse and PostgreSQL, with a focus on real-world applications in real-time analytics, data warehousing, and AI. The initiative is inspired by successful industry research models like those at Google, which have historically turned foundational research into groundbreaking technology. By leveraging ClickHouse's strong open-source culture, the research group plans to develop and test new ideas against demanding workloads, working closely with ClickHouse engineers and contributing to the broader database ecosystem. Pavlo's background includes leading the Carnegie Mellon Database Research Group and co-founding OtterTune, cementing his reputation as a leader in autonomous databases and large-scale data analytics.