ClickHouse 26.5 arrives alongside Delta Lake integration and dbt's AI agent push

ClickHouse 26.5: backward-incompatible defaults and a new kernel for Delta Lake

The biggest story of the week is ClickHouse 26.5, a major release that ships several backward-incompatible changes. The defaults for date_time_input_format and cast_string_to_date_time_mode have switched from basic to best_effort, which will parse a wider range of datetime strings by default — teams relying on the stricter old behavior will need to set compatibility or the explicit basic mode. Tuple element names can no longer be null (it conflicts with Nullable subcolumn names), and new settings — dynamic_disk_allow_from_env, dynamic_disk_allow_from_zk, and dynamic_disk_allow_include — now disable dynamic disk operations by default for security hardening.

Alongside the release, the ClickHouse blog published a deep technical post on integrating the Rust Delta Kernel to replace the native Delta Lake implementation. The move reduces maintenance overhead while unlocking Delta Lake writes, schema evolution, time travel, and partition pruning. For teams working with Delta Lake tables, this is a meaningful step toward first-class interoperability.

Also notable: clickhousectl v0.2.0 added Postgres and ClickPipes support, SQL over HTTP, and a standalone Rust client library for the ClickHouse Cloud API — useful for teams managing ClickHouse infrastructure from the command line.

Dagster and dbt: performance tooling and AI agent tooling

Dagster published a detailed report on migrating to Astral's ty type checker in their monorepo. The performance gains were dramatic, but the more surprising finding was that ty caught real runtime bugs that Pyright missed. The post is worth reading for any team evaluating Python type checkers at scale, especially for data engineering codebases.

Dagster also shipped 1.13.6 / 0.29.6 with a practical scheduling improvement: schedule, sensor, and asset daemon ticks now dispatch instigators round-robin across code locations, preventing a single code location with many instigators from delaying launches elsewhere. The release also adds dg api run launch for CLI-based run launches against the Dagster+ API.

dbt Labs had a busy week on the AI front. The dbt Developer Agent entered preview — an AI-native developer that understands dbt projects — and the company published a detailed guide on Agent Skills for shipping smarter agents in production. The broader theme is captured in a post about AI-ready data in practice, which argues that most organizations stop at cleaning and structuring data when they need to add the layer of semantic context — where metrics come from, who owns them, and what they mean. This is a direct pitch for the dbt Semantic Layer and MCP server as the infrastructure for agentic data workflows.

On the maintenance side, dbt-core shipped v1.11.11 and v1.10.22, both restoring the dbt-adapters upper bound to <2.0 after a temporary cap was lifted.

Tinybird: PREWHERE in Forward and quarantine auto-cleanup

Tinybird's latest changelog brings PREWHERE clause support to the Forward engine across all clusters, matching the Classic behavior. Quarantine tables now get a 30-day auto-cleanup TTL, which should help teams manage storage without manual intervention. The TypeScript SDK also saw incremental improvements with multi-argument support.

Native random sampling in ClickHouse

The ClickHouse blog published a practical guide on native random sampling with SAMPLE BY, covering how to run fast approximate queries on large datasets with minimal accuracy trade-offs. For teams using ClickHouse for exploratory analytics, this is a worthwhile read to understand when SAMPLE BY is the right tool versus a full scan.