dbt v2 goes GA as analytics engineering turns its engine over to Rust

The dbt engine gets rewritten in Rust

The week's dominant story is dbt going v2 everywhere. dbt v2 reached general availability as a single Rust-based engine now powering all of dbt, distributed as the superset dbt and the Apache 2.0 subset dbt-oss. Parsing a 10,000-model project runs up to 10x faster than v1, with BigQuery, Databricks, DuckDB, Redshift, and Snowflake adapters GA and ClickHouse and Spark in beta — the adapter detail was laid out in the Summit announcement, which also sketched the company's open lakehouse vision for customer-owned storage and compute.

Alongside it, dbt State went GA across Snowflake, BigQuery, Databricks, and Redshift. Instead of running on a schedule, State reads model SQL and warehouse metadata each run to decide whether to build, skip, clone, or defer each node — no selection syntax or manifest scripts. Pricing is consumption-based on daily active target tables, with reuse after the first in a day free, per-model lag_tolerance freshness, and Cost Insights. The launch also brought Fivetran Context Layer, dbt Charts, the dbt Wizard family, and Lake Compute: Charts as version-controlled YAML for governed BI, Wizard CLI and Desktop, and Lake Compute as a DuckDB-based single-node engine querying Apache Iceberg tables.

dbt-core's own week tracked that rewrite. The 2.0 line introduced agent skills installation from installed packages and project skill-paths gated on a new ai_provider flag, native Databricks metric view materializations, and an adbc_clickhouse driver with connection-level settings, while CLI branding was renamed from Fusion/dbt-core to dbt (proprietary) and dbt-oss across --version, --help, and warning output. The rest was quieter but not trivial: a BigQuery incremental fix so 404/403 responses from ADBC.GetObjects no longer cause tables to be written with REPLACE TABLE instead of MERGE, plus a dbt-oss PyPI banner and, on the v1 line, a Dimension protocol fix and metricflow floor bump.

ClickHouse arrives in the dbt ecosystem

ClickHouse's week was mostly a steady stream of stable and LTS patch releases, but the interesting work was integration. The ClickHouse adapter for dbt v2 entered public beta, powered by the Rust engine, with ClickHouse joining the dbt platform in private beta for both open-source and Cloud deployments. On the orchestration side, the official ClickHouse provider for Apache Airflow shipped with standard SQL operators, bulk inserts, and shared setup across self-managed Airflow and Astronomer. The most technically ambitious piece was a new TimeSeries engine pitched as a drop-in Prometheus replacement, letting PromQL queries run against metrics stored in ClickHouse Cloud and joined with logs and traces. Two supporting pieces round it out: replica-aware routing in public beta to pin requests to the replica holding temporary tables and named sessions, and the August ClickStack update with dashboard variables, chart formulas, PromQL support, and LLM observability.

Agents get better handles on query tooling

Two smaller releases point the same direction: making it easier for automated clients to drive data systems. Tinybird's MCP server now lets callers choose the response format — CSVWithNames, JSON, or Markdown, with the current default preserved — and pass explain=true to inspect a query plan before executing it. On the orchestration side, Dagster added deployment capacity alert policies in Dagster+ targeting queued and in-progress runs over rolling windows, switched Serverless fast deploys to Docker images where PEX cannot run, and allowed .yml extensions for component definition files.