ClickHouse goes on offense while dbt rebuilds its ClickHouse adapter

The 26.8 LTS anchors ClickHouse's big week

ClickHouse's headline engineering story is release 26.8, the new LTS cut, which lands background queries, pipelined SQL execution, new text tokenizers, broader data-lake integrations, and faster Parquet, aggregation, and join paths. That foundation was accompanied by a genuinely wide platform push: On-Demand Compute lets you attach extra workers to a single expensive query without starving production, and Managed Postgres on Google Cloud extends the NVMe-backed, CDC-into-ClickHouse offering beyond its original cloud. ClickHouse also became a launch partner for the Data agent in ChatGPT Work, wiring ClickHouse Cloud into natural-language reporting. Maintenance releases trickled out across the supported lines — 26.7.7.92 and 26.6.5.120 on stable, 26.3.33.24 and 26.8.3.105 on LTS — the routine churn that keeps long-lived deployments patched.

Two ways to get Postgres and Parquet data in

ClickHouse's ingestion story got two distinct answers this week. WalShadow replicates Postgres straight from physical WAL into ClickHouse at roughly 200 ms latency and 289k rows/sec in benchmarks — a logical-replication-free path for teams who want near-real-time mirrors. On the batch side, the chdb Postgres extension and a guide to loading Parquet into MySQL via ClickHouse show the embedded engine being used as a portable import/export layer rather than a destination. For teams weighing the whole stack, the first end-to-end CostBench results put ClickHouse Cloud at 412–1,996× better performance per dollar under continuous load, with a companion analysis of where that gap versus Snowflake comes from.

dbt's ClickHouse adapter reaches parity

The dbt-core 2.0 pre-release track is converging on ClickHouse. The adbc_clickhouse driver now handles connection-level settings via profile custom_settings, defaults the database to the profile schema, tolerates literal ? in SQL, and identifies dbt in the HTTP User-Agent, with dev.39 and the earlier dev.37 filling in v1-parity defaults; test-PyPI builds (dev.39, dev.38, dev.37) mirror them for early adopters. More consequentially, rc.2 brings unit tests and full materialized-view support to ClickHouse — table compatibility with MVs, on_schema_change, and in-place refresh-schedule updates — closing a long-standing feature gap. On the stable side, 1.12.4 fixes built-in generic test eviction after a Fusion manifest load and adds otel_trace_id()/otel_span_id() Jinja functions, while 1.11.15 hardens package install paths.

Orchestration and streaming glue

Elsewhere, Dagster 1.13.22 speeds up run dequeuing when concurrency pools are saturated, adds time-window filtering and live paging to the tick timeline, and drops the code-location requirement from its MCP launch tools. Tinybird's Forward CLI 4.6.17 adds tb job retry for failed imports and cloud-storage syncs, though it also renames tinybird.rate_limits_log to cluster_control_log — a breaking change for anyone querying the old Service Data Source. Finally, ClickHouse's own AI Functions write-up (SQL-native classification, generation, embeddings, and semantic search) and a tutorial on turning agent text into ClickHouse vectors with Google's MCP Toolbox suggest where the vector-and-agent workload is heading.