Fix LayoutGrid leaks #25853
Fix LayoutGrid leaks #25853
Conversation
While the above comment claims that these variables are useful, they are not added to the solver nor added to any constraints, so they cannot contain any useful information. They are only really used for the `__repr__`, which can be changed to the actually-used Variables instead.
Internally, the parent is only used in `__init__`, so doesn't need to be saved as an attribute. Externally, the constrained layout code needs the parent only for subfigure margins, but we have the parent gridspec available directly to use. The `LayoutGrid` stores its children in a NumPy array, and saving the parent causes a reference cycle. Apparently having one reference in a C-level NumPy array causes Python to not realize it can garbage collect the cycle, causing all `LayoutGrid` and its children to leak.
There was a problem hiding this comment.
Nice investigation! I verified that this works locally.
I'm wondering if leaks like this are worth adding to the test suite to make sure the reference cycles don't get reintroduced? We could depend on objgraph for this, or do some of our own verification asserting LayoutGrid isn't in gc.get_objects()
|
So was it the second paragraph that was the memory leak or the first paragraph also? Thanks! |
|
The second paragraph is the leak; the first is just motivation for making the change this way instead of trying to break the cycle after layout manually. |
…853-on-v3.7.x Backport PR #25853 on branch v3.7.x (Fix LayoutGrid leaks)
|
Thanks for fixing this - sorry I didn't have time to investigate, but suspect I'd never have figured this out. I'm happy with the fix, however, this does seem to me like an upstream bug? Is it really a thing that object |
PR summary
Internally, the
LayoutGridparent is only used in__init__, so doesn't need to be saved as an attribute. Externally, the constrained layout code needs the parent only for subfigure margins, but we have the parent gridspec available directly to use.The
LayoutGridstores its children in a NumPy array, and saving the parent causes a reference cycle. Apparently having one reference in a C-level NumPy array causes Python to not realize it can garbage collect the cycle, causing allLayoutGridand its children to leak.Removing the
parentattribute removes the cycle and fixes the leak.Additionally, I've removed extraneous
Variables that are not used for anything other than__repr__, but they are not actually filled with anything real there. In the example from #25819, this saves 16Variables per iteration.Fixes #25819
PR checklist