ClickHouse’s conference week reshapes query execution and managed services

Multi-stage queries reshape large-scale analytics

The biggest engineering story this week is ClickHouse’s multi-stage distributed query execution, unveiled at the company’s Open House conference. Multi-stage distributed execution repartitions intermediate data between query stages, removing the bottlenecks that plague large joins and high-cardinality aggregations in distributed settings. Early TPC-H results show join-heavy queries running up to 3.4× faster, while aggregation scaling remains near-linear at 7.4× faster on eight nodes than one. This is a fundamental architectural change for ClickHouse Cloud, not a tuning knob — teams running complex analytical queries across many nodes will see the most impact.

Managed Postgres and serverless observability enter beta

ClickHouse’s conference also brought two significant managed service announcements. Postgres managed by ClickHouse entered public beta, offering a fully managed, NVMe-backed Postgres service with native CDC into ClickHouse and a unified query layer via the pg_clickhouse extension. On the observability front, ClickStack Cloud launched as a serverless platform where teams send OpenTelemetry data to a managed endpoint and immediately explore logs, metrics, and traces without operating any infrastructure. The observability story deepened with AI Notebooks and an MCP server, both now available in beta, alongside CostBench, an open benchmark for comparing cloud data warehouse cost-performance across vendors.

Dagster tightens the governance and Snowflake loop

While ClickHouse dominated with conference news, Dagster shipped version 1.13.7 with several practical improvements. The build_schedule_from_partitioned_job function now accepts an owners parameter, making schedule ownership explicit at definition time. The Fivetran component gained optional column-level metadata fetching for synced tables — a small but meaningful step for lineage tracking. Two accompanying blog posts framed Dagster’s positioning: one explained how Dagster Compass powers self-service analytics at Brooklyn Data by layering governance and business context on top of Snowflake, while another made the case that Snowflake runs data, Dagster runs everything else, covering how the orchestrator handles transformation, lineage, automation, and cost visibility across the broader platform.

dbt on agent infrastructure

A single post from dbt Labs this week generated conversation: what data infrastructure agents actually need. The argument is that AI agents fail not because of weak models but because they run on infrastructure designed for batch training and human analytics — a claim that resonated with teams already experimenting with agentic workflows. It’s a short read that frames a growing tension between OLAP-oriented warehouses and the real-time, decision-making demands of autonomous systems.