cache extract_lambda_source

[tybug]

Addresses the performance part of #3963. Have not profiled memory. It should also be significantly reduced as we're calling ast.parse less now, but this will not address any underlying memory leak, if one exists.

[jobh]

👍

Some general comments, for consideration.

• The value of maxsize seems sensible enough, but there is a risk of the lambdas being memory heavy by pinning big objects (through their closures). We could consider a WeakKeyDictionary, or even invasively storing the source on the lambda itself.
• repr(f) /"{self!r}" can potentially also be expensive, so much so that we have special checks for especially bad ones (see repr_call; I don't know if that is likely to apply to functions though). Elsewhere, we defer the repr by storing the format string plus arguments, not interpolating the string until it's requested. Is that a possibility here? I'm not familiar with the observability interface, so I don't know.

[edit to clarify: By "here", I mean at the call site in stateful.py]

[Zac-HD]

• The value of maxsize seems sensible enough, but there is a risk of the lambdas being memory heavy by pinning big objects (through their closures). We could consider a WeakKeyDictionary, or even invasively storing the source on the lambda itself.

I like the idea of using WeakKeyDictionary; it'd mean we get an unlimited-size cache with negligible memory overhead.

• repr(f) /"{self!r}" can potentially also be expensive, so much so that we have special checks for especially bad ones (see repr_call; I don't know if that is likely to apply to functions though). Elsewhere, we defer the repr by storing the format string plus arguments, not interpolating the string until it's requested. Is that a possibility [in stateful.py]?

The contrasting concern is that if the arguments are mutated, the deferred reprs end up actively misleading. Given the nature of stateful tests, I think we should probably continue to repr in advance until we get a complaint or see a profile which shows that repr performance is a problem here in practice.

[Zac-HD]

Coverage issue is tracked in #3968, I'll aim to get that over the weekend if nobody beats me to it.

[tybug]

Good call with WeakKeyDictionary, thanks!

[Zac-HD]

🚀