You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
TestPfiClusteringOnDenseFeatures has, I believe introduced a week ago in PR #1832, has been responsible for some seemingly spurious test failures that are blocking builds. See, e.g., this build here from @Zruty0 for his PR #1920 , where PFI (which is totally unrelated to @Zruty0 's change) is blocking checking it in.
My best guess as to why this is happening is that for this clustering test the number of threads was not set to 1. (That obviously will cause issues, but whether it's the only issue is more than I know right now.) Sometimes you can kinda get away with that (e.g., sometimes results from logistic regression can resemble each other in different configuration settings since it is solving a convex problem), but this is not true of clustering. For that reason, the test TestPfiClusteringOnDenseFeatures has been responsible for multiple spurious test failures since its introduction last week.
We will have a PR to set the threads to 1, at least as a first try to get the situation somewhat under control. Note that this will not be necessary in FastTree since it is engineered in such a way that it gets the same result no matter how many or few threads are used.
TestPfiClusteringOnDenseFeatures has, I believe introduced a week ago in PR #1832, has been responsible for some seemingly spurious test failures that are blocking builds. See, e.g., this build here from @Zruty0 for his PR #1920 , where PFI (which is totally unrelated to @Zruty0 's change) is blocking checking it in.
My best guess as to why this is happening is that for this clustering test the number of threads was not set to 1. (That obviously will cause issues, but whether it's the only issue is more than I know right now.) Sometimes you can kinda get away with that (e.g., sometimes results from logistic regression can resemble each other in different configuration settings since it is solving a convex problem), but this is not true of clustering. For that reason, the test
TestPfiClusteringOnDenseFeatures
has been responsible for multiple spurious test failures since its introduction last week.We will have a PR to set the threads to 1, at least as a first try to get the situation somewhat under control. Note that this will not be necessary in FastTree since it is engineered in such a way that it gets the same result no matter how many or few threads are used.
/cc @rogancarr @shmoradims @artidoro
The text was updated successfully, but these errors were encountered: