Yury Selivanov
Yury Selivanov
> Use our trampolines as the immutable versions, make the baseline ones stable. Isn't this the best option?
We'll need to chat about this... I'm not sure how to make this not too complicated to configure
I'd like @deepbuzin to also look at this
I'm seeing this as well -- Anthropic API seems to be subtly broken and responds with a malformed XML instead of the expected JSON. This is **sonnet 4.5** routed through...
I can ponder on this during the core sprint (and think how this will play with uvloop).
@alexhartl have you considered just making `_scheduled_tasks` a regular set instead of using `weakref.WeakSet()`? What's the downside of that? IMO it doesn't matter much where a strong reference is stored...
> @1st1 But that would keep the awkward design where all loops share the global "all tasks" set, which is not how it's ever used. (I believe this design was...
> I'd prefer to avoid holding strong references in global scope to reduce the potential for memory leaks. I.e. the loop owns these references and we can be sure that...
https://github.com/geldata/gel-python/pull/705
I'll repeat benchmarks, but this can be closed now, see the merged PR