Tinybird tightens the permission model as ClickHouse benchmarks agents

Permissions and sharing in Tinybird

Tinybird spent the week drawing lines around who can touch what. Viewers can still list and read Workspace secrets, but creating, updating or deleting them now returns a 403 — and the same tightening extends to Classic Releases, where only admins can create, promote or delete. It's the kind of change that shows up as a failed CI job for anyone who handed a service account the viewer role and assumed it could manage variables.

The friendlier half of the story is collaboration. Explorations, which used to live entirely in their creator's private space, can now be shared with the whole Workspace from the Share control beside the title, with the familiar Favorites, Private and Workspace tabs in the sidebar to sort your own work from your team's. Only the owner can change the sharing, so the permission model stays consistent: read is broad, write stays narrow.

dbt State goes free, and the Champions program turns six months old

dbt State's trial no longer leads to a paywall. A free plan now picks up where the trial ends, with a monthly 1,000 DATT allowance that resets on the first of the month and asks for no card or contract. Self-hosted users who switch dbt State on get the same deal; anyone already on the dbt platform gets 2,000 DATT a month through March 2027. For teams that have been weighing whether to try dbt State for orchestration and warehouse cost tuning, the decision is now essentially free to make.

Alongside it, dbt Labs looked back at the first six months of its Champions program — 30 Champions, 20 Big Projects, and a pile of community work spanning dbt Pipelines, finance modeling and open-source tooling. Less a product announcement than a check-in with the people doing the unpaid evangelism, but a useful map of where community energy is going.

ClickHouse: agent benchmarks, NVIDIA silicon, and Postgres observability

The ClickHouse blog kept up a steady pace. Round two of the Agentic Analytics Benchmark added GPT-6 Astra, Luna and Sol, plus Claude Opus 5.5 and Claude Sonnet 5.5 — all five land within a point of each other mid-board, with Sonnet 5.5 moving most on turns, cost and truncated runs. If model choice for text-to-SQL felt settled, this suggests the leaderboard is compressing rather than diverging. On the hardware side, ClickHouse on NVIDIA's Vera CPU topped ClickBench, with micro-benchmarks for filters, aggregation and joins showing near-linear scaling across 176 cores. And for anyone running Postgres on the side, Query Insights now surfaces wait events: breakdowns of what active sessions are blocked on, lock-wait and blocked-session charts, and the same metrics exported to Prometheus.

There's also a practical walkthrough for tracing Cursor agents with Langfuse on ClickHouse, covering per-turn traces, spotting expensive model and tool calls, and confirming with evaluations that a fix actually helps.

Meanwhile ClickHouse's release train rolled out another week of stable and LTS patches across the 26.3, 26.7, 26.8, 26.9 and 26.10 lines — routine version churn rather than a story, though anyone pinned to an LTS branch has fresh builds to pick up, including the latest 26.3 LTS and 26.8 LTS cuts.