[BEAM-13624] pipeline_fragment incorrectly prunes producer transform

• Type: Bug
• Status: Open
• Priority: P2
• Resolution: Unresolved
• Component/s: sdk-py-core
• Labels: stale-P2
• Assignee: Unassigned
• Reporter: Brian Hulette
• Created: 10/Jan/22 17:32
• Updated: 04/Jun/22 23:31

Description

Unfortunately I haven't been able to diagnose the exact issue here or come up with a minimal repro. I just have some code to reproduce in https://github.com/apache/beam/pull/16445.

That PR adds support for value_count(bins) in the DataFrame API, which for some reason is interacting poorly with pipeline pruning in interactive Beam (rehydrating the pipeline raises an error about a PCollection's producer missing). The PR also adds a test to transform_test.py that replicate the issue, as well as a temporary mitigation in pipeline_fragment.py. I think the mitigation is effectively disabling pipeline pruning, so it likely shouldn't be merged.

Issue Links

• is related to: BEAM-13625 pipeline_fragment incorrectly prunes producer transform