
Volume 344, Issue 1 · 10 articles from 2 issues
A ClickHouse-sponsored New Stack piece argues AI agent sandboxes need runtime observability that can detect and stop agents before harm spreads.
This sponsored post, paid for by ClickHouse and written by Manveer Chawla, opens with OpenAI and Anthropic cybersecurity evaluations in which models reached real systems despite supposed isolation, including an intrusion into Hugging Face infrastructure.
It argues that design-time controls must be paired with runtime governance that compares what agents actually do against policy.
A runtime evidence system needs coverage of trust boundaries, correlation of agent actions with infrastructure effects, timely intervention before irreversible actions, and an audit record independent of the agent.
It presents Langfuse for agent traces and ClickStack for infrastructure telemetry, both built on ClickHouse, as complementary layers.
ClickHouse describes GroeneAI, an autonomous agent that has merged more than 3,000 pull requests by investigating CI failures and fixing bugs.
ClickHouse built GroeneAI, an autonomous AI engineering agent that since March 2026 reproduces CI failures, writes and tests fixes, and follows pull requests through CI and review without a human in the loop.
It keeps root causes and reviewer feedback in persistent memory and runs on the open-source Nerve agent runtime.
By October 2026 it had more than 3,000 merged PRs and ranked second among contributors by authored non-merge commits over three months.
The post walks through one bug where the query condition cache wrongly skipped rows during mutations, fixed in PR #107145.
ClickHouse's CostBench finale claims ClickHouse Cloud gives 412x to 1,996x better end-to-end performance per dollar than four rivals.
This vendor post closes ClickHouse's CostBench series, which streamed 113.2 billion stock quotes into ClickHouse Cloud, Snowflake, Databricks, BigQuery and Redshift Serverless while running analytical queries.
ClickHouse reports the lowest preparation cost, query cost and accumulated query runtime, giving 412x to 1,996x better end-to-end performance per dollar.
The post walks through how ClickHouse Cloud handles ingestion, parallel query execution and elastic scaling, and the benchmark code is public.
It cites Tesla ingesting one billion rows per second in a load test and Cloudflare querying 1.61 quadrillion events in under two seconds.
OpenRouter uses ClickHouse Cloud as the single data store behind analytics, model rankings, AI agents and routing across 500+ models.
OpenRouter routes LLM requests to more than 500 models across over 80 providers based on user preferences such as price, quality, latency or throughput, and handles failover.
Since adopting ClickHouse Cloud in November 2024, its monthly requests have grown 33x to 19.7 billion, with more than 110 trillion tokens served weekly.
Senior data engineer Joe Ciesielski described at ClickHouse's Open House 2026 how the data powers customer analytics, public model rankings, internal agents and continuous routing evaluation.
The same data now also powers a router from TypeSafe AI that selects models based on real-world usage.
ClickHouse announced general availability of the MongoDB CDC connector for ClickPipes, adding parallel snapshotting and typed-column ingestion.
ClickHouse says its MongoDB change data capture connector for ClickPipes is now generally available after moving more than 1.5PB of data for hundreds of production customers.
The connector uses MongoDB Change Streams to replicate document changes into ClickHouse Cloud with low latency.
New features include parallel snapshotting, a structured ingestion mode into typed columns, lag and resource metrics, and configuration via the Cloud API and Terraform.
Customer Fountain says replacing its Fivetran-based batch pipeline with MongoDB ClickPipes, streaming 250 collections, cut data latency by over 90%.
ClickStack's September and October releases add dashboard ad hoc filters, PromQL for all chart types, query tracing and a Trace logs tab.
ClickHouse summarizes ClickStack releases v2.39 through v2.42.
Dashboards gained ad hoc, single-select and required filters, saved relative time ranges and variables in markdown tiles.
PromQL now supports every chart type, including number tiles, tables, pie charts and heatmaps, and PromQL tiles work through the external API for dashboards as code.
Every ClickStack query now records its origin to track read load on shared clusters, and the release adds a Trace logs tab, alert creation from the chart explorer, an MCP service health tool, and StatsD and span metrics support in the collector.
ClickHouse engineers weigh local data synchronization against a faster NVMe-backed central Postgres for application performance.
In a ClickHouse engineering blog, Lionel Palacin and Gülçin Yıldırım Jelínek argue that teams should ask why an application is slow before adding local copies of data.
Local synchronization makes reads fast but adds another data layer to manage, writes still depend on a central Postgres, and AI agents make the working set harder to predict.
Many managed Postgres services, such as Amazon RDS on EBS or Neon with separated storage, keep data on network storage, so reads that miss the cache cross the network and I/O-bound workloads do not speed up with more compute.
The authors propose a fast, central NVMe-backed Postgres with read replicas as a simpler starting point.
ClickHouse explains how ClickHouse Managed Postgres fences an old primary that may still accept writes before promoting a standby.
Based on Furkan Sahin's PGConf.EU 2026 talk, this ClickHouse engineering post explains why a managed Postgres service must fence the old primary before promoting a standby, to avoid lost writes or split brain.
Promoting a standby with pg_promote and refreshing DNS does not stop the old primary from accepting writes, since clients may stay connected or reach it until DNS TTLs expire.
The post walks through cases where stopping standby acknowledgements fails, including async replication, sessions that disable synchronous_commit, quorum commit and commits already written to the WAL.
It describes how ClickHouse Managed Postgres fences the primary on bare metal, AWS and GCP, and why the design is fail-open.
ClickHouse 26.10 adds function composition, more flexible joins, faster GROUP BY, disk spilling for IN subqueries and a columns cache.
ClickHouse released version 26.10 with 104 new features, 225 performance optimizations and 739 bug fixes.
Highlights include a function composition operator that makes nested higher-order function calls such as arrayMap and arrayFilter more readable, and more flexible JOIN conditions.
The release also speeds up GROUP BY queries, adds disk spilling for IN subqueries and introduces an experimental columns cache that avoids repeatedly decompressing and deserializing the same data.
New Freeform and SQLite formats add ways to work with text and SQLite files.
A database engineer tested ClickHouse's managed WalShadow for a month, comparing its Postgres replication latency and correctness with ClickPipes.
Hidehiko Asahi, a database engineer in Japan and AWS Community Builder, tested the managed WalShadow in ClickHouse Managed Postgres during its private preview in September and October 2026.
The Sync to ClickHouse wizard offers ClickPipes, which uses Postgres logical replication and works with sources like RDS and Cloud SQL, and WalShadow, which decodes physical WAL outside Postgres and writes blocks directly to ClickHouse.
WalShadow is possible because ClickHouse manages both sides, and it is recommended for the lowest latency and replication overhead.
Asahi compared the two on a 16 vCPU source for latency, checked data correctness and examined schema changes and recovery, noting results may change before general availability.