This seems reasonable.

Unfortunately, the figure doesn't survive a pickle - i.e. with the new cache, the objects are not serializable. I expect the figure used to have something that detached the cache, and you could likely do the same thing here? But I'm not an expert on (nor a fan of) pickling figures ;-)

Thanks for the brief explanation@jklymak.
I managed to pickle theweakref by adding the__setstate__ in Text. Using the same example from the issue, the improvement is conserved. with this change,test_pickle.py passes locally.
However, some of my local tests are failing, even in the main branch. I pushed the changes and will use the CI to check if there are more tests failing. (Sorry about that, I know it isn't a good practice, but it will help me while I find out what broke in my environment.)

Many tests fail locally for me as well. If you are pretty sure it's not due to your pr it's perfectly fine to push tothe more controlled test environment.

Thanks for the comments@tacaswell, I hope I addressed your comments with the recent changes.

I wonder(?) if it may make sense to keep the old global cache layer as well, so that if you're e.g. drawing 10 axes each with the same tick labels, the labels on the latter axes still benefit from the work done on the first axes. Of course, that would require a bit more rearchitecting of the caching...

@anntzer what is the status here after#22271? Does this just need a rebase?

No, because then there was#22323 which further got rid of the explicit_cached in favor of a lru_cache. There may(?) still be some small benefit to caching layout info per-instance, although restoring that (which was mostly removed in#22271, which places the cache a bit earlier in the call tree) would require some more work, not a simple rebase (and I don't know if timings would show a large further improvement).

Moved to 3.8 as the rebase is non trivial and there have been other changes to this caching layer so we need to evaluate if we still need this.