hi @DGolubets - on some level this is expected to me, i.e. your deployment's runtime is likely missing polars as a dependency, so your code failed at import time before a flow run logger could be setup.

is your expectation that this line

ModuleNotFoundError: No module named 'polars'

is surfaced by the worker in the flow run logs? I think that's a very reasonable enhancement request

Hi @zzstoatzz

Thanks for quick reply.
I wrongly assumed that Prefect collects the logs from pods stdout.
If it send them from Python, then this behavior is understandable indeed.

Though UX would definitely be better if Prefect could handle this case.

Though UX would definitely be better if Prefect could handle this case.

Agreed! I will mark this as an enhancement

In cases where the flow author does not have direct access to the infrastructure, this becomes a point of friction. For example, the flow author might be a data scientist who does not have direct access to their organization's kubernetes cluster. In this case, the data scientist would have to rely on their cluster owner to find and share the stack trace.

My understanding is that these are errors that happen in the Runner. Errors that happen while provisioning infrastructure are captured in worker logs, and errors that happen in the context of a flow are captured in flow run logs, but between the infra being provisioned and the flow starting, logs are not captured.

In addition to import errors I believe this would also happen with syntax errors. Suppose you create a deployment using git storage. After creating the deployment, you commit and push a syntax error. Starting a flow run will then retrieve the commit with a syntax error and crash, but not report logs.