cassandra-commits mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Sylvain Lebresne (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (CASSANDRA-12277) Extend testing infrastructure to handle expected intermittent flaky tests - see ReplicationAwareTokenAllocatorTest.testNewCluster
Date Tue, 26 Jul 2016 08:08:20 GMT

    [ https://issues.apache.org/jira/browse/CASSANDRA-12277?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15393397#comment-15393397
] 

Sylvain Lebresne commented on CASSANDRA-12277:
----------------------------------------------

bq. If we, collectively, don't have the discipline not to abuse something like this, we have
bigger problems.

We're all sometimes lazy, and it's not specific to this project, so I think not making it
too easy to do the wrong thing (here, blindly ignoring a rare flaky test as long as it doesn't
fail too often) is just smart project management. Also, test is the one area where I think
we've been historically pretty bad at discipline, so while I'm all for improving our ways,
I'm going to object on assuming discipline until I consider we've collectively and consistently
shown discipline for a reasonable length of time.

bq. If we leave a precise comment in the code, we still have a non-green test-board and the
cognitive burden of filtering out "known flaky" failures when checking test results.

You misunderstood what I meant. I'm not saying we let the test flake, I'm saying we manually
modify the test to "run this thing N times, if M pass we're good" (and clearly explain why
it's ok to do so), but without bothering adding infrastructure for it.

But let me also be clear I'm not suggesting we'd manually modify tests that way on any regular
basis, that would be stupid.

Basically, I feel we're conflating 2 things here. It seems to me {{ReplicationAwareTokenAllocatorTest.testNewCluster}}
is (potentially) a rare case where we *understand* why the test is flaky, but where 1) fixing
the flakiness is not worth it and 2) we decide that we understand the flakiness well enough
that we can trust the test to provide value _even_ if we ignore a few failed runs.

But I can't really see that being anything else than a very very rare situation (I'm not even
entirely saying I'm fine with that one). So I don't think we should base our infrastructure
for handling flaky tests on those assumptions. In general, a flaky test is a bug (probably
of the test but still), and we should identify the reason of the flakiness and fix it. I'm
fine marking tests with {{@flaky}} temporarily (when our educated guess is that it's probably
a test problem), while we find time to fix it properly so it doesn't clutter the test result
board. But I'm not convinced we should *replace* that by a different annotation that considers
a flaky test is fine as long as it doesn't fail too often, which is what I understand from
the description of this ticket.


> Extend testing infrastructure to handle expected intermittent flaky tests - see ReplicationAwareTokenAllocatorTest.testNewCluster
> ---------------------------------------------------------------------------------------------------------------------------------
>
>                 Key: CASSANDRA-12277
>                 URL: https://issues.apache.org/jira/browse/CASSANDRA-12277
>             Project: Cassandra
>          Issue Type: Bug
>            Reporter: Joshua McKenzie
>            Assignee: Branimir Lambov
>            Priority: Minor
>              Labels: test
>
> From an offline discussion:
> bq. The ReplicationAwareTokenAllocatorTest.testNewCluster failure is a flake -- randomness
will sometimes (on the order of 1/100) cause it to fail. Extending the ranges to avoid these
flakes goes too far and makes the test meaningless.
> bq. How about instead of @flaky/@Ignore which currently indicates a test that intermittently
fails but we do not expect it to, we instead use @tries, or @runs, or some annotation that
indicates "run this thing N times, if M pass we're good". This would allow us to keep the
current "we don't care about these test results (in context of green test board) because intermittent
failures are not expected and the test quality needs shoring up" from "we expect this test
to fail sometimes in this particular way."



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

Mime
View raw message