Joachim B Haga

Even if not supported, the current behaviour is inconsistent - the lock can be acquired multiple times, but only released once: ``` >>> locker = pals.Locker('test', url) >>> lock1, lock2...

...mind, I sense a whiff of overengineering here.

If you dislike the new `pin` semantics, a fairly straightforward alternative is to allow "sticky" pinning of unknown keys by storing them as `not_set`. I.e., it's the _key_ that's pinned,...

Thanks for the review @tybug! > * old footgun: if a key is evicted before you pin it, and you later re-cache that key, you think it's pinned but the...

Note, I added a minor change to always compare `sort_key` instead of `score`, and to remove a bad comment and associated no-cover pragma.

> Note, I added a minor change to always compare `sort_key` instead of `score`, and to remove a bad comment and associated no-cover pragma. It turned out this was a...

> I'm not really sure what the differences are between the non-flaky failing cases and similar passing cases (`check-py310-cover`). The setup process looked pretty identical besides additional info about their...

> The consequence of using pytest.warns in internal tests is that the exceptions are never raised, so we don't see the consequences of mis-handling them. Maybe we should always use...

FYI @tybug The element `st.dictionaries` performance (with parameters as above) went down a lot with 3f82f86b3, I don't know if that's known/expected?

I think this is probably the conjecture data cache in `internal/conjecture/engine.py`. It is bounded but large, at `CACHE_SIZE = 10000` examples. There is a memory/performance trade-off here, and the benefits...