dbt opens the door to Exasol as ClickHouse turns SQL into an API layer

dbt-core widens its adapter surface

The biggest story of the week was dbt-core's march toward 2.0, where the headline is a brand-new warehouse: Phase 1 Exasol adapter support arrived with connection and auth handling (TLS, timeout profile options), an agate/seed type conversion layer including HASHTYPE, reserved-keyword auto-quoting, and the full macro bundle covering table, view, incremental (append, merge, delete+insert, microbatch), snapshots with hard deletes, model contracts, grants, and persist_docs. The same release line brought Databricks query tags that populate automatically from metadata and can be configured at the profile or model level, plus parse-safe views exposing dbt.models and dbt.columns that never leak compile-only columns at parse time. On the security side, dbt closed a symlink and hard-link extraction vulnerability in package tarballs, and the dev line picked up Databricks persist_constraints recognition so Fusion can persist legacy constraints and unnamed primary-key expressions from schema.yml meta, along with a fix that retains all remote table properties when detecting drift.

ClickHouse's 26.8 SQL surface, explained

ClickHouse's engineering blog spent the week documenting what 26.8 makes possible rather than what it fixes. Two posts stand out: building a streaming HTTP API inside ClickHouse using named handlers, typed parameters, pagination, framing formats, and access control — effectively turning the database into its own service layer — and pipelined SQL, which lets multi-stage queries read as a linear sequence of transformations instead of nested subqueries. Elsewhere, the MySQL CDC connector for ClickPipes reached general availability with parallel snapshots, safer production defaults, better observability, and infrastructure-as-code support for replicating MySQL and MariaDB into ClickHouse Cloud, while Managed Postgres added scheduled upgrade windows for Scale and Enterprise customers.

Two posts worth flagging for anyone evaluating analytical agents: ClickHouse released an open agentic analytics benchmark built from 201 real production warehouse questions and 28 models scored on correctness, cost, and speed, with the harness published so teams can rerun it on their own data. And for Postgres users, a survey of four new system views in PostgreSQL 19 that surface lock contention, recovery state, autovacuum priorities, and dynamic shared memory allocations. Meanwhile the ClickHouse server itself cut nine LTS and stable releases across the 26.3, 26.6, 26.7, and 26.8 lines — routine long-term-support churn rather than new capability.

Platform plumbing: Tinybird, Dagster

Tinybird promoted the JSON data type out of private beta and on by default, so any workspace can create JSON columns and query semi-structured event data without requesting a flag — the kind of change that quietly unblocks a lot of dynamic schema work. Organization pages (Overview, Observability, Logs, Cluster) now also load for members who don't belong to a workspace. Dagster's 1.13.21 patch fixed a backfill that could hang forever when all runs finished but partitions still lacked materialization status — they now complete as failed instead of staying in progress — and, in Dagster+, extended single-action dynamic partition wipe-and-delete to multi-partitioned assets, while locking YAML-defined alert policies against UI edits.