ClickHouse's 26.9 feature drop anchors a week of query-engine refinements

ClickHouse 26.9 lands, and the maintenance trains follow

The week's headline is ClickHouse 26.9, which introduced conditional LIMIT boundaries, incremental refreshes for append-only materialized views, disk spilling for DISTINCT, time-limited tokens, and faster min/max/count queries. Those are the kinds of changes that touch actual query plans: spilling DISTINCT keeps large aggregations from dying on memory limits, append-only materialized views can now refresh incrementally instead of rebuilding, and time-limited tokens give teams a way to hand out short-lived credentials without a separate auth layer.

With the new line cut, the rest of the product's week was the usual multi-branch maintenance rhythm. The stable channel moved through its early 26.9 patches and then a second round of 26.9 fixes, a third and a fourth, while the LTS branches got parallel attention: the 26.8 LTS line picked up its first September patch, followed by more 26.8 LTS fixes, a further 26.8 maintenance cut and another. Older supported lines weren't forgotten — the 26.7 stable branch was patched and again later in the week, and 26.3 LTS received its own fix. If you run ClickHouse in production, the practical takeaway is that 26.9 is now the branch to evaluate and the 26.8 LTS is where the conservative upgrades landed.

Tinybird moves scheduled Copy Pipes off workspace compute

Tinybird's change this week is small in surface area but meaningful for anyone pushing data between datasources on a schedule: paid workspaces can now run scheduled Copy Pipes on isolated on-demand compute by adding a single flag to the Copy node. Scheduled executions stop competing with workspace compute, which matters when a recurring copy job is heavy enough to skew the latency of interactive queries. Manual runs still need the on-demand flag passed explicitly on the CLI.

dbt-core brings Unity Catalog reads to Lake Compute

The dev line of dbt-core continues to fill out Lake Compute, this time with Unity Catalog reads against AWS Databricks using PAT or OAuth M2M credentials, plus declaration-aware validation and parity for dbt debug attach. The same development build also fixed modules.re functions silently dropping keyword arguments such as flags, maxsplit, and count, and stopped Databricks tag diffs from reporting unchanged columns and keys. Note the migration caveat: existing local Unity attach options under config.lakecompute have to move to config.duckdb, and Unity access is read-only — model writes aren't supported.

Dagster sharpens its error messages

Dagster's week was mostly quality-of-life: projects whose root module isn't importable now raise an error explaining that the package is probably not installed rather than a bare ModuleNotFoundError, and the ECS agent gained cross-account service discovery so it can register code server tasks in a Cloud Map namespace living in a different AWS account, reconciling on a configurable interval.