GeezerVison

This looks like a reasonable change.

Consider whether OrderedDict should be used in the cache where the order changes. We've shown that next(iter(d)) is quadratic while od.popitem(False) is linear. For the cache sizes in the re module, this might not matter or might be outweighed by OrderedDict's slower __getitem__ and __setitem__.

It is fine to go ahead with this PR since it is improves on what we have now.

It was discussed in #76519, performance difference is negligible. On other hand, OrderedDict adds a dependence. And it would be much worse if it is implemented in Python.

I added some data in the issue.

Side discussion: I was looking at the code in tomllib and its cache has maxsize=None. I wondered whether its use pattern is substantially different or whether the policy should be the same as for re. Perhaps the tomllib module should be bounded, or perhaps the re module didn't really need a limit. Another alternative is to set a limit that is very high so that most users never hit the limit and never trigger an eviction. After all, the objects being cached are very small, so setting a size limit is arguably a premature optimization. The code would be much simpler and a little faster without a size limit.

Side discussion: I was looking at the code in tomllib and its cache has maxsize=None. I wondered whether its use pattern is substantially different or whether the policy should be the same as for re

Hi there, it's tomllib author here! Tomllib's cache does not have an explicit bound, but is implicitly bound to 2880 items. This stems from the fact the the function's input always has to go through a regex before the lru_cached function is called.

The number 2880 comes from 24 hours * 60 minutes * 2 (offset direction).

@hukkin I suggest adding a comment to that effect.

Thank you Raymond for looking at this yet one time.

As for the limit, I think it is needed here, because RE patterns can be created on-fly, depending on the user data, and some programs may use millions of different patterns, but every pattern is only used once.