ClickHouse pushes into Postgres replication and AI SQL as dbt wires up its adapter

ClickHouse's platform expansion

The headline of the week is ClickHouse 26.8 LTS, which brings background queries, pipelined SQL, new text tokenizers, and faster Parquet, aggregation, and join queries — the kind of release that reshapes what you can build on the engine. That arrived alongside a run of stable and LTS point releases across the 26.x line, from the 26.3 LTS branch through the 26.6, 26.7, and 26.8 stable branches — routine churn, but useful if you're pinned to one of them.

The more interesting story is how far ClickHouse reached into adjacent territory. WalShadow replicates Postgres data directly from physical WAL into ClickHouse at roughly 200 ms latency and 289,000 rows per second, and the chdb Postgres extension brings high-performance imports and exports across cloud storage and data formats using the embedded ClickHouse engine. ClickHouse also expanded its Managed Postgres offering to Google Cloud with NVMe-backed storage, native CDC into ClickHouse, and a unified query layer via pg_clickhouse.

On the analytics side, On-Demand Compute lets you scale individual queries with extra workers without disrupting production, and the new AI Functions bring classification, generation, translation, embeddings, and semantic search into SQL with cost controls. Integration work rounded things out: MCP Toolbox turns agent text into ClickHouse vectors with no embedding service to maintain, Parquet files load into MySQL through ClickHouse, and ClickHouse joined as a launch partner for the Data agent in ChatGPT Work. Benchmarks got their own treatment too, with CostBench's end-to-end results and a breakdown of the real-time performance-per-dollar gap versus Snowflake.

dbt-core's ClickHouse adapter matures

dbt-core spent the week hardening its ClickHouse story. The 2.0.0-rc line added unit test support and expanded materialized view functionality, including on_schema_change and in-place refresh-schedule updates, with Databricks views now respecting full refresh when view_update_via_alter is enabled. The dev builds introduced a new adbc_clickhouse driver that configures connection-level settings via profile custom_settings, defaults the database to the profile schema, handles literal '?' in SQL, and identifies dbt in the HTTP User-Agent.

Maintenance branches weren't idle: v1.12.4 fixed built-in generic test eviction after a Fusion manifest load and added otel_trace_id() and otel_span_id() Jinja functions, while v1.11.15 began validating that resolved package installation paths stay within the configured packages-install-path.

Smaller fixes worth knowing

Dagster's release made run dequeuing significantly faster when concurrency pools are in use, since pool state is now read once per dequeue pass rather than repeatedly; it also gave the tick timeline time-window filtering, paging, and a live-updating view. Over at Tinybird, the Forward CLI can now retry eligible import and S3 or GCS sync jobs stuck in error or cancelled states without dropping to the API.