Heisenbug: tests sometimes fail audio buffer validation

[bmcfee]

Description

I've noticed that some of our CI tests seem to randomly fail with audio buffer validation errors, only to pass correctly when restarted. I don't have any explanation for this, and I doubt it's due to anything in librosa.

It seems to have started after the switch to soundfile as the primary decoder, and it may be indicative of a bug deep in the guts of libsndfile, if not in the python bindings. Any help diagnosing this would be much appreciated!

[lostanlen]

Is there a bruteforce procedure to trigger the Heisenbug with (1-ε) probability ?

[bmcfee]

Is there a bruteforce procedure to trigger the Heisenbug with (1-ε) probability ?

I haven't found any reproducible trigger for it, no. I can't get it to happen on my local installation either.

[lostanlen]

(7 weeks later)
is this still happening?

[bmcfee]

I haven't seen it lately, but I also haven't been running tests very often.

[lostanlen]

Would be nice to keep an eye on this in the wake of #1064

[bmcfee]

Would be nice to keep an eye on this in the wake of #1064

Indeed. FWIW i haven't seen it happen lately, and certainly not since rewriting the test fixtures. I think it's probably clear, but let's hold off on closing this out until we're ready to push 0.8 (ie after having merged a bunch of PRs and run lots of tests).

[bmcfee]

This came back at us in #1120 -- a reboot of the build image fixed it. So it's definitely still active.

[bmcfee]

Aha! I think I've figured this one out.

Some of our safety checks (eg length checks or error cases) that use dummy input audio (because the content doesn't matter) use np.empty to construct the buffer. If we get unlucky with the uninitialized data, it can parse as nan/inf and fail the valid_audio checks.

I'll create a PR to remove np.empty from all test cases shortly.