October 2026 Summaries

28 posts from ClickHouse

Filter
Month: Year:
Post Summaries Back to Blog
GroeneAI is an autonomous AI engineering agent built on the open-source Nerve runtime that has been investigating, fixing, testing, and maintaining ClickHouse issues since March 2026, with human involvement primarily occurring during code review. It identifies work from CI failures, fuzzers, user reports, and maintainer requests, maintains persistent task state and memory across sessions, and uses separate planning, implementation, and independent model-review stages to validate changes before submission. Its work has ranged from correcting a query-condition cache bug caused by mutations, improving UNION DISTINCT parallelism while preserving single-thread performance, clarifying misleading CI performance reports, and addressing SQL injection and TLS certificate-validation issues. By October 25, 2026, it had opened more than 4,400 pull requests, merged more than 3,100, and ranked among ClickHouse’s most active contributors, with many changes focused on bug fixes, CI reliability, testing, and performance. The system also records reviewer feedback and prior investigation outcomes to refine future work, reduces unnecessary comment volume, and continues monitoring CI after merges to confirm that reported failures remain resolved.
Oct 29, 2026 4,180 words in the original blog post.
CostBench, a benchmark of continuous data ingestion and analytics across ClickHouse Cloud, Snowflake, Databricks, BigQuery, and Redshift Serverless, reported that ClickHouse Cloud delivered 412× to 1,996× better end-to-end performance per dollar under its methodology, which combined data-preparation costs, query-serving costs, and accumulated query runtime. Using 113.2 billion streamed stock-market quotes, the study found ClickHouse had the lowest reported preparation cost at $28.69, maintained pre-aggregated daily summaries in sync with raw data, and achieved low query cost and runtime, including approximately 15-millisecond aggregate queries and roughly one-second raw-data drill-downs at scale. The post attributes these results to insert-time sorting and materialized-view aggregation, compressed columnar storage, background merges, sparse indexing, vectorized and SIMD execution, and parallel processing across cores and nodes. It also describes ClickHouse Cloud’s separation of storage and compute, allowing ingestion and query workloads to scale independently, and presents customer examples involving Tesla, Cloudflare, Buildkite, Appcues, Mintlify, HighLevel, and Artemis to illustrate applications in real-time analytics, operational decisions, and agent-assisted security investigations.
Oct 28, 2026 4,994 words in the original blog post.
ClickHouse’s upcoming 26.10 release will enable a cost-based join optimizer based on DPsub, an exhaustive dynamic-programming algorithm that identifies the lowest-cost join plan under the existing cost model for queries of up to 12 tables, while falling back to the previous greedy approach for larger joins. Unlike earlier join reordering, which was largely limited to inner joins and treated semi, anti, and outer joins as barriers, the new system can safely reorder inner, left, right, full outer, semi, and anti joins through a conflict detector derived from SIGMOD research. The detector uses rules concerning commutativity, associativity, null rejection, and operator conflicts to permit every result-preserving transformation while rejecting plans that could alter query output. Its bitmask-based implementation adds little overhead to DPsub’s efficient subset search, including support for cross products and appropriate hash-join build-side selection. Validation against more than 90,000 join trees and 2.8 million candidate plans found no invalid or missing plans, while differential fuzzing of randomized multiway queries also produced no mismatches. By opening valid alternatives previously hidden by join barriers, the optimizer can substantially reduce intermediate data sizes and improve execution times, with reported benchmark improvements reaching 55 percent on selected TPC-H and TPC-DS queries.
Oct 28, 2026 3,892 words in the original blog post.
TypeSafe’s Jev System One model was evaluated as a fast typed classifier for diagnosing analytics-agent failures and ranking database tables relevant to natural-language questions. On a 136-question development set, Jev labeled 277 failed attempts in under a second each, finding that wrong data selection, particularly entity-resolution mistakes, was the largest detailed failure category and identifying 18 likely grading errors caused partly by a deterministic comparison bug. Its table-ranking approach, which asks whether each schema table is needed for a question, achieved 0.951 mean nDCG@10 and 0.991 recall@10 in an 18-table warehouse without fine-tuning or runtime gold data, while a public-schema pilot showed lower but still useful performance on catalogs containing up to 710 tables. However, feeding ranked-table hints to 20 benchmarked agents did not produce statistically meaningful accuracy gains on either development or held-out benchmark questions, although it sometimes reduced exploration by lengthy runs and sometimes added overhead to already efficient ones. The findings suggest that retrieval ranking is a useful low-cost capability, but that many remaining agent failures stem from planning and correct use of discovered data rather than merely locating the right tables.
Oct 27, 2026 3,516 words in the original blog post.
ClickHouse Managed Postgres introduced customer storage alerts in August, initially warning users when disk usage exceeds 85%, and plans to expand notifications to help customers identify and respond to operational issues without continuously monitoring dashboards. Upcoming alerts will flag inactive logical replication slots retaining at least 10 GiB of unconfirmed WAL for two hours, completed planned switchovers, completed unplanned failovers, transactions open for at least two hours, and connection use reaching 95% of the service limit. The notifications are intended to distinguish expected maintenance interruptions from unexpected recovery events, prevent storage growth caused by abandoned CDC pipelines, reduce risks from long-running transactions such as locks and table bloat, and help avoid connection exhaustion. Customers can manage currently available notification delivery through the Managed Postgres notification settings for console, email, and Slack.
Oct 26, 2026 931 words in the original blog post.
Shipsidekick, a logistics platform founded in 2024 by Sawyer Bateman, migrated its Next.js and Vercel-based application from Neon Postgres to ClickHouse Managed Postgres after recurring connection-pool exhaustion, CPU saturation, slow queries, outages during peak Monday workloads, and delayed vendor support. The company moved roughly 500 GB of data over a weekend using ClickPipes logical replication, completing the initial transfer in four to five hours, while resolving a temporary partial-index issue during replication. Shipsidekick selected ClickHouse Managed Postgres for its pricing, instance configurations, bundled PgBouncer connection pooling, NVMe-backed storage, documentation, and responsive support, after also considering PlanetScale. According to Bateman, the new deployment reduced typical CPU use from 100% to 5–10%, raised cache-hit rates from 30–35% to 99%, and made queries roughly ten times faster, including a shipment-data query that fell from 80 seconds to about six seconds. The company later added safeguards and a dedicated read replica for AI-agent workloads after an unrestricted read query affected primary database performance, and it is considering moving application dashboards and other analytical workloads to ClickHouse itself.
Oct 23, 2026 1,944 words in the original blog post.
Langfuse’s open-source Observability plugin for Cursor records model calls, tool executions, token usage, latency, and related metadata as traces, enabling teams to analyze agent behavior, costs, and performance. Powered by ClickHouse, Langfuse v4 stores trace context alongside observations to support investigation of individual prompts and aggregate dashboards for usage, cost, latency, and evaluation scores across users, models, tags, and environments. The guide explains how to install and configure the plugin with Langfuse API keys, notes its privacy and logging settings, and describes how Cursor prompts are represented through LLM, tool, subagent, and context-compaction observations. It presents a workflow in which teams identify unusually expensive turns, detect recurring inefficiencies such as repeated file reads or failed shell-command retries, revise project instructions, and evaluate whether changes preserve task completion while reducing cost and latency. Langfuse can also support broader AI application development by turning production cases into evaluation datasets, comparing prompt, model, and tool changes before deployment, enforcing quality thresholds in CI, and feeding poorly scored production traces back into future experiments.
Oct 21, 2026 1,102 words in the original blog post.
ClickHouse evaluated its unmodified database software on NVIDIA’s early-access Vera CPU platform, where a dual-socket system provides 176 Arm-v9.2 physical cores, 352 hardware threads, substantial cache capacity, and up to 1.2 TB/s of memory bandwidth, placing it at the top of the ClickBench hardware leaderboard. Using a 1 TB in-memory table containing 62.5 billion strings, the engineers tested filtering, low-cardinality aggregation, and joins to examine how effectively query execution scales with available compute resources. A simple substring filter became memory-bandwidth-bound at roughly 64 threads, reaching about 901 GB/s, while more computationally intensive regular-expression filtering, hash-based aggregation, and joins scaled nearly linearly through the 176 physical cores, with diminishing benefits from simultaneous multithreading beyond that point. The join test used ClickHouse’s parallel hash join and runtime filtering, while query pipeline plans illustrated that major processing stages run concurrently across many data chunks. A separate single-threaded SIMD test showed that compiler vectorization using Vera’s NEON instructions reduced a numeric filter’s runtime from 1.48 seconds to 0.42 seconds, and ClickHouse is adding support for Vera’s higher-throughput SVE2 instructions. The results are presented as evidence that ClickHouse’s parallel query engine can make extensive use of Vera’s core count and memory capabilities, while identifying continued opportunities for performance improvements in data structures, caching, compression, batching, and asynchronous processing.
Oct 21, 2026 2,643 words in the original blog post.
OpenReplay is an open-source, self-hosted session replay and product analytics platform that enables developers and product teams to observe user interactions, diagnose bugs, and analyze behavior while preserving customers’ control over their data. After encountering slow queries, difficult sharding, poor compression, and labor-intensive retention cleanup in Postgres, the company migrated behavioral-event storage and analytics to ClickHouse, while retaining Postgres for metadata. ClickHouse improved session-search performance to roughly one second, dramatically reduced storage requirements, and automated data retention through TTL policies, eliminating disruptive manual cleanup work. OpenReplay uses a centralized ClickHouse pipeline in which behavioral events enter a flat events table and materialized views transform them into downstream datasets, shortening pipeline changes and supporting rapid feature experimentation. Native JSON support also enables customers to define and query custom analytics events for funnels, heat maps, and flows. Large customer deployments ingest billions of events and up to roughly 6 TB daily on a single vertically scaled ClickHouse instance, an approach designed to minimize operational overhead in self-hosted environments. The company also uses query caching to accelerate repeated analytics requests and is extending ClickHouse to support AI agents that review session recordings and identify issues proactively.
Oct 21, 2026 1,875 words in the original blog post.
ClickHouse’s follow-up Agentic Analytics Benchmark evaluates five newly added models—three GPT-6 variants and Claude Opus and Sonnet 5.5—on 201 SQL analytics questions over the same synthetic warehouse, 60-turn limit, and independent judge panel used previously. Claude Fable 5.1 remains the leader at 76.6%, while all new models scored in a statistically overlapping 68.7% to 69.7% range, placing them in the middle of the 34-model board and below GPT-5.5 at default reasoning effort. Claude Sonnet 5.5 showed the largest improvement over its predecessor, raising pass rate by 6.5 points, eliminating turn-limit failures, improving speed and cost efficiency, and performing notably better on questions involving a dimensional CRM layer. Across the additions, incorrect planning remained the dominant failure mode, while SQL runtime errors were absent, suggesting that deciding what analysis to perform is a greater challenge than query execution. The benchmark harness was also updated with environment pinning, credential checks, configurable numeric tolerances, improved result-column matching, uncertainty modeling for judges, experiment hooks, and a LibreChat runner, while contamination tests found no evidence that the new models had memorized the benchmark questions. The authors emphasize that aggregate leaderboard results conceal important trade-offs in cost, latency, and completion behavior, and recommend that organizations run similar benchmarks using their own schemas and questions.
Oct 20, 2026 1,955 words in the original blog post.
ClickHouse Managed Postgres has expanded Query Insights with wait event metrics that show what active database sessions are waiting on, helping users distinguish whether performance issues stem from disk I/O, locks, timeouts, or CPU activity. The Active sessions by wait type chart displays average active sessions and their share of total load by category, while detailed tooltips and a Waits table identify specific PostgreSQL wait events such as tuple, transaction ID, relation, and advisory locks. Additional charts track the number of blocked sessions and lock-wait durations using p50, p95, and maximum values to reveal the scale and persistence of lock contention. These metrics are also available through the Prometheus endpoint, allowing users to build Grafana dashboards and alerts alongside query throughput, latency, failures, buffer activity, and WAL generation metrics. ClickHouse Agents can further analyze wait events with slow-query and service data to suggest investigation priorities.
Oct 19, 2026 676 words in the original blog post.
Managed ClickStack has introduced z-score-based anomaly detection for existing query-builder charts, allowing users to identify unusual changes relative to each metric’s recent history without manually writing baseline SQL. The feature includes adjustable sensitivity, directional triggers, historical previews, safeguards against false positives from minor fluctuations or delayed data, and remains best suited to sudden deviations rather than gradual changes or recurring seasonal patterns. To accommodate organizations migrating large alert collections from platforms such as Grafana, ClickStack has also improved alert scheduling through concurrent configuration loading, metadata reuse, batched writes, background notification delivery, and sharded evaluation across workers. Alert management enhancements include scalable list rendering, searchable names and tags, filters, bookmarked views, Terraform exports, acknowledgement and silencing controls, and detailed histories that show firing states, evaluation data, query performance, backfills, and notification failures. Future plans include alerting on PromQL metrics, seasonality-aware anomaly detection, better handling of noisy or low-volume data, and potentially automated or machine-learning-assisted alert creation.
Oct 16, 2026 2,205 words in the original blog post.
ClickHouse has introduced an experimental native TCP client in the ClickHouse.Driver .NET package that uses the Native columnar protocol instead of the existing HTTP RowBinary approach, reducing server-side format conversion and enabling bulk, largely zero-copy and low-allocation reads and writes. Designed for .NET 8 and later, it offers column-streaming APIs, POCO mapping, and object-array compatibility tiers, with the strongest performance benefits available through borrowed column blocks and contiguous buffers; reported benchmarks show typical operations running 1.5 to 2 times faster than HTTP, more than 99% lower allocations in some workloads, read speeds of up to 1.1 million rows per second, and concurrent insert speeds up to 877,000 rows per second. The stateful protocol also supports query progress, profiling events, server logs, server-side cancellation, tracing, and session persistence for temporary tables and settings. However, it is not a drop-in HTTP replacement, as it remains subject to API changes and lacks ADO.NET and ORM support, alternative file formats, JWT authentication, binary JSON handling, and some convenient type behavior, while JSON-heavy and object-array workloads see fewer gains. Users can run both clients from the same package, moving analytical or performance-critical paths to TCP while retaining HTTP for ORMs, file formats, authentication, and other unsupported scenarios, with future plans including ADO.NET support and binary JSON.
Oct 16, 2026 3,732 words in the original blog post.
As application data and query volumes grow, the article argues that maintaining all records in PostgreSQL can become costly because OLTP systems enforce ACID guarantees through mechanisms such as write-ahead logging, MVCC, index updates, and page-based storage, creating significant read and write amplification for data that is rarely changed. Although distributed and cloud-native PostgreSQL systems can scale capacity and throughput, they generally do so with roughly linear infrastructure costs and added coordination overhead. It proposes separating actively changing data, which remains in PostgreSQL for transactional writes and fresh reads, from settled or mostly read-only historical data, which can be replicated to ClickHouse for lower-cost storage, aggregation, and analytical queries. ClickHouse’s columnar architecture is presented as offering substantially greater compression, cheaper infrastructure, and more efficient batch indexing and aggregations, while accepting trade-offs such as slower single-row retrieval and eventual consistency. The article promotes ClickHouse Cloud’s managed PostgreSQL offering, including NVMe storage, WalShadow change-data capture, and pg_clickhouse integration, as a way to synchronize the two systems and reduce the cost of serving settled data.
Oct 15, 2026 1,453 words in the original blog post.
ClickHouse’s monthly newsletter highlights the 26.9 release, which adds LIMIT AFTER and UNTIL boundary conditions for ordered results, incremental refreshes for append-only materialized views, time-limited credentials through CREATE TOKEN, and disk spilling for DISTINCT operations. It recognizes community contributor Hoang Minh V for extensive work improving PromQL, the TimeSeries engine, LIMIT BY behavior, and edge cases across several functions. New CostBench comparisons claim ClickHouse Cloud achieved substantially higher real-time analytics performance per dollar than Databricks and BigQuery while ingesting 113.2 billion market quotes, attributing the difference to ClickHouse’s insert-time sorting and pre-aggregation. The newsletter also announces general availability of executable UDFs in ClickHouse Cloud, the official ClickHouse provider for Apache Airflow, and chDB Durable Layer for persistent agent memory backed by object storage. Additional coverage explores ClickHouse integrations with PostgreSQL, including memory safety, local NVMe performance, and extension updates, while Open House presentations feature ClickHouse use cases at companies including Polymarket, OpenRouter, Buildkite, Canva, and Xero.
Oct 15, 2026 1,728 words in the original blog post.
CostBench compared ClickHouse Cloud and Redshift Serverless during continuous ingestion of 113.2 billion stock-market quotes, measuring data-preparation cost, query-serving cost, and accumulated query runtime while aggregate queries ran every ten minutes and drill-downs hourly. Using ClickHouse asynchronous inserts and Redshift streaming ingestion from Amazon MSK with materialized views, the benchmark reported that ClickHouse had lower modeled preparation costs ($28.69 versus $522.03), lower allocated query costs (about $0.05 versus $24.97), and less cumulative query runtime (58.12 seconds versus 1.69 hours), producing a calculated 1,996-fold performance-per-dollar advantage under its scoring formula. The comparison attributes much of the difference to ClickHouse sorting raw data and updating summaries as inserts are processed, whereas Redshift’s streaming view, background sorting, and downstream materialized-view refreshes operate as separate stages, leaving drill-down queries able to encounter less-prepared data and aggregate queries dependent on the most recently completed refresh. ClickHouse reportedly completed drill-down workloads 112 times faster and aggregation workloads 70.8 times faster in the tested setup, although the results rely on specified capacity assumptions, a 32-RPU modeled Redshift writer scenario derived from later characterization, and Redshift’s live streaming-view read path rather than its separately maintained typed projection.
Oct 14, 2026 4,527 words in the original blog post.
Ridgeline adopted ClickHouse Cloud to scale performance analytics for its unified investment management platform after Postgres could no longer support increasingly complex, data-intensive reporting for much larger asset-management customers. The company selected ClickHouse over alternatives including Redshift, SingleStore, Trino with Iceberg, and custom in-memory systems because of its combination of analytical speed, cost efficiency, security, and managed operational capabilities. Ridgeline stores recent accounting and replicated vendor data in ClickHouse, retains extensive historical data in Iceberg accessed through ClickHouse, and uses materialized views to create optimized reporting datasets. The architecture has reduced report-generation batches from hours to minutes, enabled near-immediate report revisions, supported customers roughly 10 times larger than previously possible, and more than doubled the number of customers on the platform. Ridgeline is now considering ClickHouse for additional near-real-time analytics and reconciliation workloads as it prepares to support even larger customers.
Oct 14, 2026 1,907 words in the original blog post.
ClickHouse announced several developments aimed at integrating PostgreSQL transactional workloads with ClickHouse analytics through a unified cloud stack. ClickHouse Managed Postgres, now in private preview on Google Cloud as well as AWS, provides PostgreSQL on local NVMe or GCP Local SSD storage, high-availability features, native change data capture into ClickHouse, and the open-source pg_clickhouse extension for querying ClickHouse from PostgreSQL. The company also introduced WalShadow, an open-source physical-WAL replication engine that copies PostgreSQL changes to ClickHouse with reported sub-second latency and higher throughput than its logical-replication-based predecessor PeerDB, while supporting schema changes and a managed private-preview option. A live demonstration showed WalShadow maintaining low replication lag during high-volume trading simulations and identifying a simulated anomaly faster than a ClickPipes pipeline. In addition, the new chdb PostgreSQL extension uses the embedded ClickHouse engine to improve PostgreSQL imports and exports involving cloud object storage and formats such as Parquet, Avro, ORC, Arrow, and Protobuf.
Oct 13, 2026 1,388 words in the original blog post.
ClickHouse engineers investigated an extra 32 ms of tail latency affecting simple Native TCP SELECT queries after connections had been idle for about two seconds across a 32 ms network path. The delay occurred because Linux reset TCP’s congestion window to its initial size after inactivity, while ClickHouse’s use of TCP_NODELAY caused separate small writes for result metadata such as ProfileEvents, logs, progress updates, and EndOfStream packets; together, these packets exceeded the reduced congestion window and required an additional client acknowledgement round trip. Using Linux network namespaces, traffic control, and a reused connection, the team reproduced the issue deterministically, observing results arrive after roughly one RTT but EndOfStream after a second RTT. Although configuration changes such as increasing the initial congestion window, disabling slow start after idle, or reducing profiling packets could mitigate the problem, the implemented ClickHouse fix batches suitable end-of-query metadata at the application layer while preserving timely progress messages. Packet captures showed that this nearly halved transmitted TCP packets, and the measured tail latency fell from about 32 ms to roughly 1 ms.
Oct 13, 2026 2,544 words in the original blog post.
iSAM Funds’ options desk adopted ClickHouse Cloud on AWS to manage the high-volume, unevenly distributed market data generated by options, where a single underlying instrument can affect thousands of related contracts. After initially prototyping with Postgres, the team found ClickHouse delivered 35–50 times greater compression and increased single-core ingestion from 8,000 to 800,000 rows per second, while supporting flexible data exchange through formats such as Apache Arrow, Parquet, and JSON. Its columnar storage and vectorized execution enabled fast historical queries without relying on many rollups, materialized views, or pre-aggregation pipelines, simplifying the platform and accelerating researchers’ ability to test trading hypotheses. ClickHouse Cloud also provided managed scaling, security integration with iSAM’s AWS and authentication infrastructure, backup portability, and temporary compute expansion for intensive data rebuilds, allowing the desk to expand from futures into FX and equity options while retaining full market history online and queryable.
Oct 09, 2026 1,704 words in the original blog post.
CostBench compared ClickHouse Cloud and BigQuery while continuously ingesting 113.2 billion stock-market quote records and running scheduled aggregate and drill-down queries that included newly arrived data. Using native real-time ingestion and preparation features, the benchmark measured preparation costs, normalized query costs, and accumulated query runtime, reporting that ClickHouse achieved an end-to-end performance-per-dollar score 438 times better than BigQuery Capacity pricing and 512 times better than BigQuery On-demand pricing under its scoring formula. ClickHouse used asynchronous inserts to store, sort, and incrementally pre-aggregate data during ingestion, while BigQuery used the Storage Write API, clustering, and incremental materialized views refreshed asynchronously. The reported preparation costs were $28.69 for ClickHouse versus $243.09 under BigQuery Capacity pricing and $264.59 under On-demand pricing; accumulated query runtime was 58.29 seconds for ClickHouse and 49.36 minutes for BigQuery. ClickHouse’s drill-down queries were reported as roughly 4.37 times faster overall, while its aggregate queries were 267 times faster overall, which the benchmark attributes largely to ordered raw data and continuously current summaries. The comparison used equivalent data, schemas, query schedules, and keys, disabled query-result caching, and modeled BigQuery’s Capacity and On-demand options as alternative prices for the same observed query workload rather than separate executions.
Oct 08, 2026 4,300 words in the original blog post.
Clera, an AI talent agent serving more than 200,000 candidates and thousands of active jobs, migrated its operational database from a previous provider to ClickHouse Managed Postgres after persistent performance issues, including multi-second point lookups, timeouts, failed writes, and CPU utilization that remained at 100% for days. The company evaluated Supabase, Neon, PlanetScale, and ClickHouse using a replay of its roughly 500 GB production workload and sysbench TPC-C tests, concluding that ClickHouse’s local NVMe-backed storage delivered stronger IOPS, latency, and throughput, with up to 42% more transactions per second than the runner-up in its benchmark. Clera then moved more than 500 tables from Europe to the United States overnight while maintaining production traffic, using ClickPipes and addressing migration challenges involving extensions, triggers, indexes, table sequences, and data parity. After the cutover, CPU utilization fell to roughly 10–20%, user-facing and administrative queries became faster, and the company reported improved capacity for analytics, scraping bursts, and future growth without changing database settings.
Oct 08, 2026 1,855 words in the original blog post.
ClickHouse Cloud JWT authentication, available for Enterprise services running version 26.4 or later, enables users, applications, and pipelines to authenticate with short-lived tokens issued by OpenID Connect providers such as Okta and Microsoft Entra rather than persistent database passwords or certificates. ClickHouse validates token signatures and claims including issuer, audience, subject, and expiration, then creates an in-memory ephemeral user whose permissions derive from token-provided roles and disappear when the token expires. This shifts identity lifecycle management to the organization’s identity provider, while ClickHouse administrators manage reusable database roles and can map alternative claims such as groups to those roles. Supported clients can present and refresh bearer tokens, while ClickHouse Cloud’s built-in authenticator automatically supports SQL Console and the command-line login flow. Although ephemeral users are not persisted or backed up, a stable identity UUID allows settings profiles, quotas, row policies, and masking policies to remain associated with the same identity across sessions, and controls can more commonly be applied through roles. Definer views created by JWT users retain functionality through non-login persistent shadow users, and the feature supports RSA keys from version 26.4 and additional EC key algorithms from version 26.8.
Oct 08, 2026 1,892 words in the original blog post.
ClickHouse sponsored Postgres Summit US 2026 in New York City, where its team presented ClickHouse Managed Postgres and demonstrated technologies designed to combine PostgreSQL transactional workloads with ClickHouse analytics, including pg_stat_ch for query metrics, pg_clickhouse for querying ClickHouse from PostgreSQL, chdb_hook for data file transfers, and WalShadow for near-real-time replication. The presentation generated follow-up interest at the company booth, while other sessions covered upcoming PostgreSQL capabilities such as pg_plan_advice and index prefetching, open-source change data capture options, replication, cloud-native extension packaging, and the expansion of PostgreSQL communities in Africa. Discussions throughout the event also focused on extension maintenance, upgrades, debugging, ecosystem coordination, and broader community collaboration. The author characterized the conference as a successful first PostgreSQL event sponsorship for ClickHouse and noted plans to sponsor PGConf.EU in Valencia, where ClickHouse will discuss the PostgreSQL Ecosystem Foundation and contributing to PostgreSQL documentation.
Oct 07, 2026 1,306 words in the original blog post.
CostBench compared ClickHouse Cloud with Databricks Serverless SQL during continuous ingestion of 113.2 billion stock-market quotes, using each platform’s recommended ingestion, data-layout, and materialized-view features while running scheduled aggregate and drill-down queries. Under its modeled list-price methodology, the benchmark reported that ClickHouse Cloud had lower preparation costs ($28.69 versus $695.60), lower normalized query costs (about $0.05 versus $2.65), and less cumulative query runtime (56.39 seconds versus 29.10 minutes), producing a claimed 752× end-to-end performance-per-dollar advantage. The comparison attributes the difference largely to ClickHouse performing ordering and incremental pre-aggregation during ingestion, whereas the tested Databricks configuration used separate Zerobus ingestion, asynchronous liquid clustering, and incremental materialized-view refresh processes, which could leave newly ingested data unclustered or summaries temporarily stale. ClickHouse’s summaries were updated with inserted data, while Databricks refresh completions were observed roughly 1.8 minutes apart, although the study notes that this is a proxy for freshness rather than a direct measure of data lag. Both systems used comparable nominal 16-CPU read-side capacity, disabled result caching, and processed the same workload, but the assessment excluded Databricks’ beta Lakehouse//RT product and omitted such costs as idle warehouse capacity, storage, networking, post-ingestion maintenance, and full provider invoices.
Oct 06, 2026 4,312 words in the original blog post.
ClickHouse Cloud has made executable user-defined functions generally available across AWS, GCP, and Azure, adding a Native runtime for statically linked Rust, Go, C++, and Bun-compiled JavaScript binaries, outbound network access for Python UDFs, memory limits, determinism settings for query caching, API and Terraform management, and detailed per-query observability metrics in ClickHouse 26.6+. The announcement demonstrates using UDFs for LLM accounting: a Rust tokenizer runs at ingestion through a materialized view to count system, context, user, and completion tokens, while a Python network-enabled UDF periodically asks Claude to evaluate sampled responses. Token counts can be combined with externally refreshed model pricing data to estimate costs, fill gaps in provider-reported usage, identify prompt components that may benefit from caching, and relate expensive prompts to quality failures. The post also compares Rust, Python, and Go tokenizer implementations, finding Rust fastest for this workload, and offers operational guidance including using reusable executable pools, processing larger chunks, flushing per chunk, monitoring pool waits and resource usage, and reserving UDFs for tasks impractical in SQL. Planned enhancements include secret management, per-service runtime configuration, Native-runtime networking, and additional runtimes.
Oct 05, 2026 4,914 words in the original blog post.
Direct I/O allows backup software to bypass Linux’s page cache, preventing large live PostgreSQL backups from evicting database data that queries may soon need and avoiding an additional kernel memory copy. ClickHouse Managed Postgres found that while buffered backup reads displaced warm but temporarily idle pages and increased query latency, Direct I/O preserved cached data, reduced latency impact and CPU use, but also disabled kernel readahead, which can reduce throughput on RAID0 NVMe arrays when reads are too small. To address this, its configuration sizes direct reads according to the number of drives so each request spans the stripe and adjusts parallel reader counts for hardware such as dense NVMe instances. Tests on a 467 GB database backed by four NVMe drives showed that tuned Direct I/O maintained warm cache contents, kept query p99 latency near baseline, and completed in roughly the same time as buffered backups, whereas default small Direct I/O reads were slower. The approach is generated per server, includes rollback flags for compatibility issues, and is intended to support future incremental backups that reduce the amount of data uploaded while retaining the same local I/O protections.
Oct 02, 2026 3,527 words in the original blog post.
LinkedIn’s observability team adopted ClickHouse first for an OpenTelemetry-based distributed tracing platform that processes roughly 800 billion sampled spans and 200 TB of uncompressed data daily across three data centers, meeting demanding ingestion and query-latency targets through a replicated, highly tuned architecture. Building on that experience, the team migrated its metric metadata discovery system from three aging services—a custom search index, Elasticsearch, and a key-value store—to a single ClickHouse index, while preserving compatibility for tools using legacy RRDtool-style metric names through a gateway that translates requests into SQL. The consolidated system uses continual change capture, deduplication, daily table refreshes, projections for aggregations, and directly queryable metric dimensions to support discovery and analytics across more than 13 billion metrics with about 30% daily turnover. It now serves over 150,000 queries per minute at an average latency of 68 milliseconds, reduces memory consumption to about one-fifth and compute use to roughly two-thirds of the prior stack, and has capacity to double metric volume while helping users replace inefficient regex-based queries and gradually move away from legacy metric conventions.
Oct 01, 2026 1,904 words in the original blog post.