Closed andrewclegg closed 10 years ago
They are almost all due to sys.modules
changing size while clean_warning_registry
iterates over it.
And we previously had no failures, with the same release of sklearn.
I think this is a race condition which only shows itself under heavy load. We can probably get around it by patching clean_warning_registry
to iterate over a shallow copy of sys.modules
, if it rears up again.
All tests pass when running on its own, so I haven't introduced any nasty regressions. About to test on fully-loaded server.
Works fine under load too. Pull request sent (https://github.com/scikit-learn/scikit-learn/pull/3196). Will keep this open for tracking, til the pull's accepted. Snake Charmer will continue to install off my fork til then.
https://gist.github.com/andrewclegg/e3060f4e49e04ee66dac
Planning to upgrade to a newer build and see where that gets me.
Raise scikit-learn ticket if this doesn't help.