Dagster 1.13.12 - Kubernetes Refresh And dbt Event Fixes

Dagster 1.13.12 was published on July 2, 2026, for core 1.13.12 and libraries 0.29.12. The most operator visible fix is a retry path for code locations left in an error state after transient gRPC failures during Kubernetes rolling deployments. For data teams running asset automation, dbt ingestion, or EMR Serverless jobs, this patch is mainly about control plane recovery and event correctness.

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

The main platform fix is in code location refresh. Before this release, a transient gRPC failure during refresh could leave a code location stuck in an error state. The release notes call out Kubernetes rolling deployments as a common trigger, which is exactly where this kind of failure is most visible to operators.

Dagster now has the gRPC server watcher periodically retry refresh for locations whose load error is a DagsterUserCodeUnreachableError. That matters because rolling a user code deployment should not require manual repair of the location state after a brief network or process handoff failure.

This is not a broad change to deployment semantics. It is a retry path for one concrete unreachable code error. That narrower scope is useful. It reduces stale error states without pretending that every user code load failure is transient.

Dagster 1.13.12 also fixes a partitioned asset check problem in automation. A partitioned asset check with its own automation condition could raise a CheckError and halt the entire automation tick when it was requested on a tick that did not materialize the asset it targets.

For asset graph operators, the important part is the blast radius. The old behavior could turn one check request into a full tick stop. The fix keeps the automation tick from being halted by that specific mismatch between a check request and a target asset materialization.

This is the kind of patch that matters most in large Dagster installations where checks, partitions, and automation conditions are used together. The release notes do not describe a schema migration or a new config flag for this item. It is a runtime correctness fix for the scheduler path.

The runs feed gets a performance fix for runs that target large numbers of assets and asset checks. Instead of fetching the full selection immediately, the feed now fetches a bounded preview and loads the full lists on demand.

Although this is marked as a UI change, it has operational value. Large asset selections are common in backfills, broad materialization runs, and maintenance jobs. A runs feed that blocks on large selection payloads makes incident triage slower, even if the underlying run execution is healthy.

The pool display fix is smaller but still practical. Long pool names are now truncated, and the run pools dialog wraps and scrolls when many pools exist. That helps teams that encode environment, queue, tenant, or workload class in pool names. It does not change scheduling behavior, but it removes a display failure around capacity inspection.

The dagster-dbt fix is specific and worth noting. Dagster now avoids a KeyError on failures when converting dbt run results to Dagster events for runs that omit that key, including dbt Fusion runs.

That is an event conversion problem, not a dbt execution problem. If Dagster is the system of record for lineage, asset health, and run observability, losing the conversion path because one key is absent is still a pipeline visibility failure.

For AWS users, PipesEMRServerlessClient now accepts a dashboard_refresh_interval parameter, and the default was increased so Spark UI URLs are no longer invalidated too early during a run. This is most relevant while debugging active Spark workloads through the EMR Serverless dashboard. It keeps the inspection window useful for longer running jobs without forcing every caller to carry a local workaround.