Dagster 1.13.28 - Startup retries and dependency fixes


Dagster 1.13.28, published on October 30, 2026, adds optional retries for code locations that fail to start, including image pull failures and startup timeouts. That gives platform operators a recovery option for a failure that occurs before pipeline execution. This is a stable release, pairing core 1.13.28 with libraries 0.29.28.

The full release notes are on the GitHub release page.

Agents in dagster-cloud can now retry a code location after startup fails. The setting is code_location_start_retry, exposed in Helm as workspace.codeLocationStartRetry. It is opt in, so upgrading alone does not enable this behavior.

The useful distinction is where recovery happens. An image pull failure or startup timeout can prevent a code location from becoming available before any job gets a chance to run. This setting addresses that startup failure. Teams that currently intervene after such failures have a specific agent setting to evaluate.

Retries can help when the cause is temporary. A persistent image or configuration problem still needs a fix. The notes do not specify retry timing or limits, so those details need checking before this becomes part of an operational recovery procedure.

For teams that install dbt packages themselves, project preparation can now skip dbt deps. Set install_deps=False through the supported interfaces on DbtProject, DagsterDbtProjectPreparer, or the DbtProjectComponent project arguments.

A build process can own package installation, while Dagster prepares the project without repeating that command. The tradeoff is that the external installation step must supply the packages the project needs. Use the option where that responsibility already belongs to the build process.

There is also a concrete storage compatibility fix. Accessing UPathIOManager.storage_options could raise an AttributeError with universal-pathlib version 0.3.0 or newer. The affected IO managers include PolarsDeltaIOManager.

If that combination is in a pipeline environment, the exception has a named fix in this release. An upgrade check should exercise the affected IO manager with the dependency versions actually installed, including access to storage_options.

The Dagster Plus Insights metrics “Runs in progress” and “Runs queued” now capture brief spikes that previously went unrecorded. Bursts of very short runs are the example given in the notes. Requests for either metric now use Max when the request omits an aggregation.

That changes how operators should read an alert after upgrading. A threshold can fire on a short burst that the previous metric behavior missed. More alerts do not, by themselves, establish that the workload has become busier. Compare the burst duration and the requested aggregation when investigating the difference.

The “Run concurrency” Insights metric is also deprecated in favor of “Runs in progress”. Operators using it in dashboards or alerts should review that dependency. The release notes announce the deprecation without giving a removal date.

The Snowflake integration raises two dependency minimums. dagster-snowflake now requires snowflake-snowpark-python>=1.37.0 and snowflake-connector-python>=3.14.0. The stated reason is an import failure in older Snowpark releases when setuptools is version 82 or newer.

Environments with older versions pinned need those constraints reviewed before updating the integration. Check the resolved packages in the built runtime and verify that the Snowflake integration imports there. A successful dependency resolution alone does not exercise the import failure described in the notes.

These minimums are the concrete packaging change to account for here. The new code location retry behavior remains a separate configuration decision because its setting is optional.