Hey Hussein, -1 for apache-airflow-providers-common-ai 1.1.0rc2 - I think we should exclude it from this vote and cut rc3 with a fix. Proposing no change for microsoft-azure 15.3.0 and openlineage 2.21.0.
The reason: common-ai declares "pydantic-ai-slim>=2.36.0" and "opensandbox>=1.1.0" with no upper bounds, so a fresh install today pulls pydantic-ai 2.62 and opensandbox 1.2.0, and the 1.1.0rc2 code is broken with both: * pydantic-ai 2.60 changed how a streamed model request is recorded by durable execution (bare ModelResponse instead of the response+events wrapper). AirflowDurability fails to decode it, so durable runs using streaming break - both live and on replay. * opensandbox 1.2 allows egress rules with ports and no target. When the enforced policy does not match the requested one, building the error message raises TypeError on that None and hides the real reason the sandbox was refused. The same code is in the already released 1.0.0, so users installing it now hit this as well - shipping 1.1.0 as-is would ship it a second time. I found it via the CI environment upgrade PR, where the common-ai tests started failing after the dependency bump. The fix (compatible with both the old and new versions, keeping the current lower bounds) is part of that PR: https://github.com/apache/airflow/pull/75270 Once it is merged, I think good idea is to cut common-ai 1.1.0rc3 including it? J. --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
