Fn Runner Watermark issue #34484
Fn Runner Watermark issue #34484
Conversation
|
Checks are failing. Will not request review until checks are succeeding. If you'd like to override that behavior, comment |
|
Assigning reviewers. If you would like to opt out of this review, comment R: @claudevdm for label python. Available commands:
The PR bot will only process comments in the main thread (not review comments). |
| @@ -388,7 +388,8 @@ def create_stages( | |||
| translations.lift_combiners, | |||
| translations.expand_sdf, | |||
| translations.expand_gbk, | |||
| translations.sink_flattens, | |||
| translations.fix_flatten_coders, | |||
There was a problem hiding this comment.
The fix_flatten_coders was added to solve yaml unit tests.
It also mimics the translations.standard_optimize_phases() used by Portable Runner
Without it, theYamlMappingTest::test_basic yields
apache_beam.testing.util.BeamAssertException: Failed assert: [
Row(label='11a', isogeny='a'), Row(label='37a', isogeny='a'), Row(label='389a', isogeny='a')] == [BeamSchema_ccf257cb_1966_410e_8157_00cd826e7392(label='11a', isogeny='a'), BeamSchema_ccf257cb_1966_410e_8157_00cd826e7392(label='37a', isogeny='a'), BeamSchema_ccf257cb_1966_410e_8157_00cd826e7392(label='389a', isogeny='a')],
unexpected elements [BeamSchema_ccf257cb_1966_410e_8157_00cd826e7392(label='11a', isogeny='a'), BeamSchema_ccf257cb_1966_410e_8157_00cd826e7392(label='37a', isogeny='a'), BeamSchema_ccf257cb_1966_410e_8157_00cd826e7392(label='389a', isogeny='a')],
missing elements [Row(label='11a', isogeny='a'), Row(label='37a', isogeny='a'), Row(label='389a', isogeny='a')
]
There was a problem hiding this comment.
It looks like the PreCommit YAML is still failing, can you please take a look>
There was a problem hiding this comment.
I think this maybe doesn't work because fix_flatten_coders assumes that the flattens will eventually be dealt with by sink_flattens. That may be what is causing the failures?
| @@ -388,7 +388,8 @@ def create_stages( | |||
| translations.lift_combiners, | |||
| translations.expand_sdf, | |||
| translations.expand_gbk, | |||
| translations.sink_flattens, | |||
| translations.fix_flatten_coders, | |||
| # translations.sink_flattens, | |||
There was a problem hiding this comment.
To be clear, this PR isn't fixing the underlying issue, rather it is disabling the optimization?
Can we add a comment referencing why this is disabled, with reference to the bug?
Also I am not sure if the tradeoff between disabling this optimization is worth fixing this specific edge-case, maybe @damccorm has thoughts?
There was a problem hiding this comment.
I think it would be better to try to fix the underlying issue. This has the potential to introduce new encoding/decoding problems which it may be possible to generally avoid
Proposing a solution to #26190 .
It appears the Flatten was not setting a watermark, which caused following steps not to execute.
The issue was returning errors on beam versions 2.39 onwards, and it potentially produced unstable results before 2.39.
There are meaningful TODOs mentioned:
# TODO(robertwb): Possibly fuse multi-input flattens into one of the stages.# TODO(pabloem): Consider case where there are various producers