MRG, MAINT: Fix PySurfer-related crash

[larsoner]

Let's see if larsoner/mayavi@c2a8a23 fixes our Azure failures. Based on the traceback it might.

Will require several restarts of the Azure runs because this is a bit of a random bug.

[larsoner]

• First test
• Second test
• Third test
• Fourth test
• Fifth test
• Sixth test failed

[codecov]

Codecov Report

Merging #7295 into master will decrease coverage by 0.74%.
The diff coverage is 100%.

@@            Coverage Diff             @@
##           master    #7295      +/-   ##
==========================================
- Coverage   89.82%   89.07%   -0.75%     
==========================================
  Files         447      447              
  Lines       80694    80974     +280     
  Branches    12876    13032     +156     
==========================================
- Hits        72482    72131     -351     
- Misses       5385     6050     +665     
+ Partials     2827     2793      -34

[larsoner]

Trying again with a new approach that changes PySurfer's __del__

• First test failed

[larsoner]

Neither worked :(

Next I'll see if it's a PyVista or Mayavi problem by removing each of those.

PyVista:

• first test
• second test failed

[larsoner]

Pushed a commit to try to clean things up explicitly with gc.collect in case mayavi/pyface/PySurfer/whoever has a bug:

• first test
• second test
• third test
• fourth test
• fifth test
• sixth test (unrelated download failure)
• seventh test
• eighth test
• ninth test
• tenth test

[larsoner]

It looks like this might fix our sporadic failures. I'll go ahead and merge if we get through 10 iterations without error, okay @GuillaumeFavelier @agramfort ?

[GuillaumeFavelier]

It's okay for me

[agramfort]

yes... although the diff is puzzling...

[larsoner]

yes... although the diff is puzzling...

Briefly, from what I understand:

• PySurfer does some mlab.close when objects get .close or __del__'ed (i.e., during garbage collection)
• We can't be sure of when the GC happens (and triggers the __del__)
• When creating two brain objects, it can be the case that the end up using the same Mayavi figure because of how Mayavi keeps track of its figures and how we create them in PySurfer (this is probably a bug with PySurfer's use of Mayavi)

Thus you can end up in some test creating a Brain instance with a Mayavi figure, then Python does GC and calls the __del__ from a previous Brain that shared that figure, killing the resources, and when your shiny new Brain instance goes to draw somewhere down the line (Mayavi or TVTK) it tries to draw some objects that have been killed by mlab.close.

This PR works around this corner case stuff by forcing GC. Eventually we should probably fix PySurfer, too, but even once we do, these GC calls shouldn't hurt anything.

[agramfort]

thanks for the explanations

…

[larsoner]

This behavior might also now be fixed by nipy/PySurfer#285, but I plan to merge this PR to MNE after test #10 so that we can have stable CIs again while we wait for that to be merged (and released, etc.)