dbt v2 lands as the ecosystem rushes to meet it

dbt v2 arrives, and the ecosystem reorganizes around it

The headline of the week was dbt v2, and dbt Labs spent the week explaining what it changes. The v2 release itself collapses the long-running split between two parallel codebases into one shared adapter layer and drops the giant manifest.json from the docs experience. The most tangible payoff is developer feedback speed: a rebuilt SQL engine that lints up to 250x faster surfaces errors while you type instead of waiting for a run to fail, and a built-in DuckDB adapter means you can run dbt end to end against a local, offline warehouse with a small profile change. That adapter layer has company — Columnar's production ADBC drivers, now used by dbt v2, are the story of the type-system work that makes zero-copy transfers possible.

Adoption paths got attention too. Apache Iceberg is suddenly an afternoon project thanks to a new dbt catalog configuration, and for teams on a larger footprint, dbt Mesh running on Iceberg demos open data infrastructure composition, while dbt Labs and Fivetran outline moving workloads onto a data lake model by model rather than ripping and replacing. State management got a dedicated walkthrough of what dbt State does for compute and freshness.

AI moves inside the models

dbt's AI story this week was less about copilots and more about putting models in the pipeline. Google Cloud's Alicia Williams showed how to call BigQuery's AI.GENERATE from plain SQL, with typed output schemas, incremental blocks, token caps, and dbt tests on model output — no Python service required. Under Armour took a similar instinct further, materializing Snowflake AI agents as dbt models so each one passes review and CI like any table. And dbt Wizard got a design rationale: an agent built to show its work rather than a general coding agent pointed at a project.

The most interesting artifact was an experiment. Serving an engineered summary through the dbt MCP server cut AI skill cost 48.5% and runtime 43.2% versus pulling raw vendor transcripts, though it lost decisions and job titles along the way. On the debugging side, an MCP-connected agent diagnoses a broken dashboard in six minutes by tracing the bad number back through the project.

dbt also launched something genuinely new: dbt Charts, an open-source YAML language for dashboards that version and deploy alongside your models in the same repo.

Query engines get smarter about plans, merges, and cost

ClickHouse's cost-based join optimizer now exhaustively reorders inner, outer, semi, and anti joins to find the cheapest valid plan — a meaningful widening of what used to be inner-join-only territory. The company also published the Agentic Analytics Benchmark, whose sobering finding is that retrieval hints leave accuracy flat across 20 models, and expanded customer notifications for Managed Postgres to cover replication slots, failovers, and connection limits. In a different register, ClickHouse introduced GroeneAI, an autonomous contributor behind more than 3,000 merged pull requests.

Tinybird answered a practical operational question: querying system.merges to see which long-running merge or mutation is holding up a deployment. The same update cleaned up system.columns leakage, fixed stale JSONPaths breaking S3 backfills after a Data Source redeploy, and stopped tb push from dropping a Token's row-level filter.

Dagster hardens the operational edges

Dagster's two releases were mostly about things going wrong quietly. Run action menus that stayed open let repeated re-execute clicks launch duplicate runs, and pages crashed outright when IndexedDB was unavailable — both fixed. Dagster also added protobuf 7 support and retries for code locations that fail to start, a welcome knob for cloud agents, plus a preview run_job() for PipesDatabricksClient and a dagster-dbt[dbt] extra for dbt v2.