nix: set enableParallelChecking = false#272438
Conversation
tests/gc-auto.sh.test and others fail nondeterministically. This is extremely frustrating for doing development on `staging` where every PR has to rebuild `nix`.
RaitoBezarius
left a comment
There was a problem hiding this comment.
I agree with the fix, I think it's important to nail down the source of nondeterminism so that the Nix core team can work on this on their pace and fix it maybe.
|
Failed build: |
|
Note: I do not think this is a concurrency bug in Nix -- I think it's likely a concurrency bug in the test apparatus. I'm fine with passing along this info to upstream, but in general I think that having our tests be deterministic is more important than helping upstream troubleshoot concurrency bugs in their test suite. If there were any sign that this was a concurrency bug in Nix itself (there is not) that would be a different story. If a test is specifically designed to test concurrency (like the Frankly our |
|
I assume the time difference is acceptable and not way to the moon. |
|
I feel we should still open an issue upstream. |
|
Thanks for the ping!
We use distinct local stores in each test ( You may still merge this if you want to test my theory. If it succeeds consistently, then that might suggest the existence of some shared state not covered by
Please do. |
|
@amjoseph-nixpkgs poke - were you able to see the test flakyness with all Nix versions, or only 2.3? IIRC, there's been some stability improvements in the test suite, and some got backported to 2.3.17 (by @Ericson2314). Would help to know if this still applies to the master branch, or if 2.3 is only missing some stability backports (and in that case, |
tests/gc-auto.sh.testand others fail nondeterministically. This is extremely frustrating for doing development onstagingwhere every PR has to rebuildnix. This is not worth whatever minor speedup we get from the parallel checks.