Open jktjkt opened 2 months ago
I think this is the test that I already saw failing a few times but it is so rare I have not even attempted to fix it and am hoping for a better use-case. There is nothing fishy in the test, a notification should be replayed and is not in this failing case. I do not even see any data race so perhaps there is a corner case with the timestamp of the stored notification file and the start_time
used in the subscription? There was one fixed a long time ago. I will try to look for something when I have some spare time.
Well, the subscriptions were seriously buggy but it did not manifest until I attempted (and succeeded) to make the tests more efficient. I have tried to fix it all so let me know if you still encounter some issues.
We just got this failure in the CI using this morning's latest
devel
of everything:Two runs succeeded, one run failed. The failing one was using TSAN which imposes a "slightly slower effective CPU speed". Perhaps there's some sleeping code in the test that's too optimistic?