ClickHouse own the week with Silk, WAL-RUS, and a new stable release

ClickHouse goes deep on infrastructure

It was an unusually infrastructure-heavy week from ClickHouse, with three major engineering blog posts that signal where the company is investing its core systems.

The biggest story is Silk, a new stackful-fiber library and scheduler built for ClickHouse. Silk is NUMA-aware, uses io_uring as its I/O ground truth, and promises zero heap allocation on the steady-state hot path. The first integration target is the distributed cache, but the ambition is clear: replace the thread-per-core model with cooperative multitasking for better resource utilization on modern hardware. This is the kind of foundational systems work that can reshape performance characteristics years down the line.

Equally significant is WAL-RUS, a complete Rust rewrite of WAL-G for Postgres backups. The ClickHouse Cloud team needed better performance, lower memory overhead, and a more maintainable codebase for the critical path of WAL archival in their managed Postgres service. The blog post is refreshingly honest about the trade-offs and the decision to start from scratch rather than incrementally patch the Go original.

The v26.6 stable release shipped alongside all this, bringing the usual collection of backward-incompatible changes (including the removal of the long-experimental allow_experimental_query_deduplication) and new features documented in the changelog. The LTS track also saw updates with v26.3.16.16, v26.3.15.4, and v26.3.14.49, while v26.5.3.52-stable also landed.

Benchmarking, agents, and the Postgres bridge

On the data-adjacent side, ClickHouse published a deep CostBench analysis comparing end-to-end cost-performance with Snowflake for real-time analytics — a must-read if you're evaluating platforms. The post argues that read-only benchmarks miss the full picture, and CostBench measures continuous ingest, data maintenance, and query execution together.

The agentic trend continues: ClickHouse is now a native connection in Notion Custom Agents, and the Managed ClickStack MCP Server brings observability investigation capabilities to AI workflows. There's also a thoughtful essay on building for agents that draws parallels from the decades-long evolution of the mouse to the current state of agent UX.

For Postgres users, the pg_clickhouse v0.3.2 update brings Postgres 19 compatibility, TLS support, regex functions, and memory improvements — all without requiring a restart or reload. The What's New in ClickStack post covers dashboard improvements like table linking and scoped filters.

Dagster fixes the asset catalog and publishes a platform guide

Dagster's 1.13.11 is a small but important release: the asset catalog page now renders as a single virtualized list, meaning workspaces with many asset groups or code locations will no longer freeze the UI. The release also adds an "insights" option to DbtProjectComponent's include_metadata field for Dagster+ Insights tracking from YAML config, and new GraphQL fields for run selection.

The new guide on operationalizing data orchestration is a comprehensive look at best practices for DevOps, infrastructure, and code locations — practical reading for any data platform engineer.

Security and community

Finally, a guest post on hunting for vulnerabilities with AI shows how an experienced engineer used GitHub Copilot, Claude Opus, and Gemini to find real vulnerabilities in the ClickHouse C++ codebase. It's a fascinating look at the evolving role of AI in security research, and the results speak for themselves.