Fix #5495. Combine two tests to prevent race cond #5656
Conversation
|
The nice thing is is that if my theory is correct, this is a test-only bug, and not really a bug in matplotlib itself. Longer term, we could add some sort of hook to the testing infrastructure so that filenames can't be reused or so that writes are atomic. |
|
I think your hypothesis is correct for the theoretical race condition, but I don't see how this is any better. Wouldn't the results get saved out to the same named file? So if the first figure had a regression in it, the second figure could clobber that. |
|
Yes, the results get saved out to the same file, but they are compared one-at-a-time. So both are tested against the same baseline image in turn. It does mean that the first one to fail is the one that's left on disk for inspection later, and there may be other failures "beyond" that, but that's no worse than the usual situation with asserts in unit tests. We have a number of tests that do this already without problem. |
|
ah, ok. That makes sense. Just waiting for Travis to greenlight this. |
Fix #5495. Combine two tests to prevent race cond
Fix #5495. Combine two tests to prevent race cond
|
back ported as dfadd4d |
There are two tests in
test_cycles.pythat use the same baseline image. If running the tests in multiprocess mode, it's theoretically possible that both tests would write to the output file at the same time, corrupting the file and hence the backtrace when reading it back in. I think combining these into a single test with the same image file will accomplish what is intended (sharing the same baseline data) but never run in parallel.