This is an archive of the discontinued LLVM Phabricator instance.

[clangd][tests] Bump timeouts in TUSchedulerTests to 60 secs
ClosedPublic

Authored by kadircet on Aug 31 2023, 11:42 PM.

Details

Summary

There are some slow/congested bots, that can't go idle in 10 secs, see https://github.com/llvm/llvm-project/issues/64964

Diff Detail

Event Timeline

kadircet created this revision.Aug 31 2023, 11:42 PM
Herald added a project: Restricted Project. · View Herald TranscriptAug 31 2023, 11:42 PM
kadircet requested review of this revision.Aug 31 2023, 11:42 PM
sammccall accepted this revision.Sep 1 2023, 3:13 AM
This revision is now accepted and ready to land.Sep 1 2023, 3:13 AM
This revision was landed with ongoing or failed builds.Sep 1 2023, 3:21 AM
This revision was automatically updated to reflect the committed changes.

not sure if this is the right way to fix these tests. The problem is if a device is constraiend, this will further slow down the device and create more backlogs. Can we allow a way to skip these tests based on a flag/environment variable?

not sure if this is the right way to fix these tests. The problem is if a device is constraiend, this will further slow down the device and create more backlogs. Can we allow a way to skip these tests based on a flag/environment variable?

The affected tests aren't fundamentally slower or less important than others (I expect many tests are very slow at the same time on such systems).
So skipping them doesn't seem appropriate.

The reason these deadlines exist in these tests is they coordinate multiple threads, and we want a deadlock bug (which you'd also see on healthy buildbots) to show up as a test failure rather than as the test running forever, which is harder to debug.
It's nice if it does so quickly, but if the buildbots are too slow, a longer deadline works fine for this purpose.

It will "further slow down the device" in the sense that running more tests slows down the device compared to running fewer tests, but again that's true for all tests, and we don't have reason to believe these ones are unreasonably expensive. (If so, we should understand why)

I believe @kadircet is looking into the possibility of of reducing/disabling test sharding of clangd tests on bots that don't have enough resources to gain throughput from it.