Dagster 1.13.22 - Faster Concurrency Pool Dequeue

Dagster published core 1.13.22 and library packages 0.29.22 on 11 September 2026. On a deployment that gates run starts with concurrency pools, dequeuing is faster when many runs sit queued. Pool state is read once per dequeue pass.

The full release notes and downloads are on the GitHub release page.

The dequeue pass selects queued runs and starts them when a concurrency pool has room. Previous passes read pool state repeatedly while many runs were queued. 1.13.22 reads that state once and then finishes the pass. The release notes call the result significantly faster. They include no timings and no sample queue depth. Measure queue age on your own backlog if you need a number.

No config key ships with the change. The faster pass applies when concurrency pools are in use. A deployment that already uses them picks up the single read when dequeue runs on 1.13.22. There is nothing to enable.

One read per pass is cheaper on a deep queue. The pass then runs from one snapshot of pool occupancy. The notes say nothing about a lock or a refresh inside that pass. After the upgrade, check pools with a low limit and a high arrival rate. Queue age on those pools matters as much as how fast one run leaves the queue.

Two items in this release apply to Dagster+.

The MCP server tools launch_job_run and launch_asset_run no longer require a code location and a repository name. Both values are looked up from the job being launched. A caller that knows the job, and does not know which code location loaded it, can omit those arguments. Callers that still pass a code location and a repository name stay within what this note describes. The server supplies the values when they are left out.

The bug fix is limited to Dagster+ as well. Querying asset metrics by asset key no longer fails with a server error. The note gives no exception name, no status code, and no asset key shape. An open source deployment that never issues that query is outside the fix. A Dagster+ workspace that charts asset metrics by key should run the query again once 1.13.22 is the version serving it.

The tick timeline in the UI can filter to a time window, page through history, and hold a live updating view. Storage format, daemon flags, and API fields are absent from the note. A sensor or schedule with a long tick history can be narrowed to a window and walked a page at a time. The view can refresh as new ticks land. People who read ticks from logs, or from an external metrics store, see no change from this item.

dagster-dbt on library line 0.29.22 adds support for dbt-core 1.12. The notes credit jrbasso. Coverage inside the 1.12 line is not spelled out, and the note gives no new minimum pin. dbt-core 1.12 is a supported pairing with this Dagster release. Keep the pin you already resolve until you choose to move.

dagster-looker now requires python-liquid>=2.2.1. The previous constraint was python-liquid<2. Any python-liquid release below 2.2.1 fails the new pin. That includes the 1 series the old bound allowed, and it includes 2.0 and 2.1. A lockfile that still selects Liquid 1 fails to install dagster-looker 0.29.22. Render a representative Looker asset with python-liquid 2.2.1 or newer in a branch environment before the library reaches production. The notes do not catalog template syntax that differs across that jump.

dagster-prefect is new and marked preview. A Dagster asset can launch Prefect deployments and background tasks through Dagster Pipes. The note includes partition mapping and cancellation forwarding. It names no minimum Prefect version, no install extra, and no config keys. Preview means partition mapping and cancellation forwarding can still change shape. Keep the integration in a separate code location until that contract settles.

  • Install core 1.13.22 with library packages at 0.29.22. The release name is 1.13.22 (core) / 0.29.22 (libraries).
  • Pin python-liquid to 2.2.1 or newer before upgrading dagster-looker. Liquid 1, Liquid 2.0, and Liquid 2.1 all miss the new floor.
  • Treat dagster-prefect as preview. Partition mapping and cancellation forwarding ship inside that preview.
  • The dequeue change has no flag. It is active where concurrency pools already gate runs.
  • Dagster+ MCP clients can omit code location and repository name on launch_job_run and launch_asset_run. The server looks both up from the job being launched.