The issue isn't that we miss teardown, it's that WaitUntilFinish doesn't wait until all teardown calls have finished. However this is nearly as bad as not calling teardown at all, and can result in flaky tests.
I think we should make an effort to get this fix into 2.4.0, but I also don't think we should block 2.4.0 on it. The reason being is that this is a long-standing issue (it's been a bug since Teardown was introduced) and it's not a production data-correctness issue; in the past we've stated that we will not block releases for bugs like this, however we will make an attempt to include the bugfixes if possible. Reuven On Tue, Feb 27, 2018 at 1:42 PM Romain Manni-Bucau <rmannibu...@gmail.com> wrote: > Not sure when it appeared but running any test you miss the teardown which > is: > > 1. A part you cant test today cause of that > 2. Leaking and can impact other tests depending the impl and setup > 3. Prevent any correct concurrency plannification of tests (even using the > random of the direct runner you can plan it but not with teardown leaks > leading to numerous unexpected threads) > > Can work for simple cases like the IO in beam but this week (since > yesterday ;)) already 2 users reported me broken executions due to that and > I was only able to say the PR was ready since weeks :(. Since code fix is > here I think it should make it as a minimum to make the directrunner usable > for tests. Cleanup of the code - dont think there is any - can come later > if needed but normally all was addressed anyway and doesnt impact code > quality. > > Le 27 févr. 2018 22:34, "Reuven Lax" <re...@google.com> a écrit : > >> Can you explain "unusable?" We have hundreds of direct-runner tests that >> appear to still be running just fine. Is this a new regression, or old >> behavior? >> >> >> On Tue, Feb 27, 2018 at 1:30 PM Romain Manni-Bucau <rmannibu...@gmail.com> >> wrote: >> >>> Updated 3409 as blocker since it makes direct runner - and therefore any >>> test - unusable at all (leaks+unexpected wait times + retry strategy >>> required+no real alternative even for dofn now dofntester is deprecated). >>> >>> Le 27 févr. 2018 21:45, "Chamikara Jayalath" <chamik...@google.com> a >>> écrit : >>> >>>> Looks like previous URL was broken. Created >>>> https://s.apache.org/beam-2.4.0-burndown. >>>> >>>> - Cham >>>> >>>> On Tue, Feb 27, 2018 at 12:22 PM Robert Bradshaw <rober...@google.com> >>>> wrote: >>>> >>>>> I'm planning on cutting the 2.4.0 release branch soon (tomorrow?). I >>>>> see 13 >>>>> open issues on JIRA [1], none of which are labeled as blockers. If >>>>> there >>>>> are any that cannot be bumped to the next release, let me know soon. >>>>> >>>>> - Robert >>>>> >>>>> >>>>> [1] >>>>> >>>>> https://issues.apache.org/jira/browse/BEAM-3749?jql=project%20%3D%20BEAM%20AND%20status%20in%20(Open%2C%20%22In%20Progress%22%2C%20Reopened)%20AND%20fixVersion%20%3D%202.4.0 >>>>> >>>>
