[
https://issues.apache.org/jira/browse/BEAM-9487?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17361945#comment-17361945
]
Ahmet Altay commented on BEAM-9487:
-----------------------------------
> I was thinking that would be captured in CHANGES.md whenever it was updated.
Great idea.
> However, we could set --allow_unsafe_triggers to True for the time being.
+1 to this, because I think we will break some customer using automation. It is
hard for some customer to add new flags quickly because they have complicated
setups. (airflow based setups usually have this complication for beam jobs.)
> Is there a particularly release where we'd want to set the default back to
> False?
We typically allow for 2 releases after introducing the new flag. So if the
flag will be first available in 2.31, we can switch the default in 2.33. What
do you think?
> As for PR14780, I don't think it will break any pipelines, as my
> understanding is that it is simply shifting the point of failure and making
> it easier to see, but considering we had some internal tests that broke from
> it, it might break some tests. Should we also disable it by default for the
> time being?
It make sense to switch to default for the time being. If we caught issues in
our tests, it will very likely cause some user pipelines to break. (And plenty
of warnings in CHANGES.md, and in the code warning about upcoming changes.)
> GBKs on unbounded pcolls with global windows and no triggers should fail
> ------------------------------------------------------------------------
>
> Key: BEAM-9487
> URL: https://issues.apache.org/jira/browse/BEAM-9487
> Project: Beam
> Issue Type: Bug
> Components: sdk-py-core
> Reporter: Udi Meiri
> Assignee: Zachary Houfek
> Priority: P2
> Labels: EaseOfUse, starter
> Time Spent: 13.5h
> Remaining Estimate: 0h
>
> This, according to "4.2.2.1 GroupByKey and unbounded PCollections" in
> https://beam.apache.org/documentation/programming-guide/.
> bq. If you do apply GroupByKey or CoGroupByKey to a group of unbounded
> PCollections without setting either a non-global windowing strategy, a
> trigger strategy, or both for each collection, Beam generates an
> IllegalStateException error at pipeline construction time.
> Example where this doesn't happen in Python SDK:
> https://stackoverflow.com/questions/60623246/merge-pcollection-with-apache-beam
> I also believe that this unit test should fail, since test_stream is
> unbounded, uses global window, and has no triggers.
> {code}
> def test_global_window_gbk_fail(self):
> with TestPipeline() as p:
> test_stream = TestStream()
> _ = p | test_stream | GroupByKey()
> {code}
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
