ClickStack builds dashboards from prompts, ClickHouse 26.6 ships, and Dagster tunes the runs feed

AI dashboards and workflows in ClickStack

The biggest story this week is ClickStack's AI dashboard and workflow generation. Instead of manually building dashboards by wiring queries to charts, you now describe what you're trying to understand in natural language, and ClickStack's AI Notebook explores the telemetry, validates assumptions, and produces a connected dashboard — complete with the reasoning behind each panel. The same MCP toolchain powers both ClickStack's built-in agent and external agents like Claude, Cursor, and Codex, so teams can create and refine dashboards from whichever tool they prefer.

This launch follows a week of strong ClickHouse product news. The ClickHouse 26.6 release is packed: hypothetical skip indexes let you test index effectiveness without building them, cascading refreshable materialized views chain refreshes automatically, and continuous queries enter experimental territory. Meanwhile, the engineering team published a detailed walkthrough of how they made ClickStack 5x faster for observability workloads by redesigning primary keys, adding text indexes, and rewriting queries — a useful reference for anyone running ClickHouse at scale.

ClickHouse Agents goes for Postgres, and PgBouncer gets a rethink

The AI agent story extends beyond ClickStack. ClickHouse Agents is now available for Managed Postgres, letting you explore Postgres data in plain English, query both Postgres and ClickHouse together, and build agents that monitor, tune, or even help migrate databases — all read-only by default. On the infrastructure side, a blog post on how ClickHouse scales PgBouncer in Managed Postgres explains the architecture behind running a fleet of pooler processes proportional to available cores, a pattern that avoids the single-threaded bottleneck of vanilla PgBouncer. And the "Quadrillion Rows" post chronicles LogHouse's growth to 1.59 quadrillion rows across 30+ regions and three clouds — a case study in extreme-scale ClickHouse deployment.

Dagster and the rest

Dagster 1.13.12 improves the runs feed for large asset graphs: instead of fetching the full selection at once, the feed now loads a bounded preview and pulls the complete lists on demand. The dagster-aws integration also adds a dashboard_refresh_interval parameter to PipesEMRServerlessClient so Spark UI URLs aren't invalidated mid-run. A community showcase post highlights creative projects built on Dagster, worth a look for inspiration.

On the dbt side, dbt-core v1.11.12 stops spurious deprecation warnings for state model configs — a small but welcome fix. And dbt-core v2.0.0-alpha.3 adds automatic version pointer views for versioned models, so dim_customers automatically points to the latest dim_customers_v2. Finally, Tinybird Local now runs natively on arm64, eliminating emulation overhead on Apple Silicon, and lightweight deletes enter beta via the new delete endpoint.