gh-91887: Store strong references to pending tasks

[alexhartl]

This adds a _pending_tasks set to BaseEventLoop. On Task creation, a (strong) reference to the task is added to this set in _register_task. When a task completes, the respective reference is removed from _pending_tasks in _unregister_task. See the discussion at #91887.

• Issue: gh-91887

[ghost]

All commit authors signed the Contributor License Agreement.

[1st1]

I can ponder on this during the core sprint (and think how this will play with uvloop).

[freakboy3742]

@alexhartl I'm at the CPython core team sprint, so I've taken the liberty of merging with main so I can discuss this with @1st1 and others. There were some conflicts introduced as a result of #120974.

[alexhartl]

Thank you for picking this up @freakboy3742 ! Yes, the last time I checked, the asyncio C code appeared to be in the middle of some restructuring. Let me know if I can help with anything.

[bedevere-app[bot]]

A Python core developer has requested some changes be made to your pull request before we can consider merging it. If you could please address their requests along with any other requests in other reviews from core developers that would be appreciated.

Once you have made the requested changes, please leave a comment on this pull request containing the phrase I have made the requested changes; please review again. I will then notify any core developers who have left a review that you're ready for them to take another look at this pull request.

[1st1]

@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 -- in the event loop or in the asyncio module. Either way a proper cleanup is required, but using a regular set would be a trivial change and everything in the ecosystem would just work.

[gvanrossum]

@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 -- in the event loop or in the asyncio module. Either way a proper cleanup is required, but using a regular set would be a trivial change and everything in the ecosystem would just work.

@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 just a historical accident; I've never found an explanation.)

[1st1]

@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 just a historical accident; I've never found an explanation.)

Yeah, but what's awkward about it? There's a low level API that manages the all tasks set and other event loops (at least uvloop) already use that API. Adding yet another tracking API for running tasks is harder than switching the tracking API we already have to just use a regular set. The additional tracking will only introduce additional, albeit minuscule, overhead. At least this is how I see this. I do think it would be quite elegant to make the existing APIs asyncio._register_task and asyncio._unregister_task have some additional functionality.

Adding new API (like what this PR is doing) means that other loops will have to always implement it (or be broken). Which I obviously can do for uvloop, quite easily, but it will grow the API surface which is already huge.

Lastly, a minor point: event loop doesn't have a lot to do with tasks. The loop is mostly concerned with running callbacks. Task is a self-contained primitive that just schedules callbacks to the event loop. So it rubs me the wrong way to introduce tracking to the loop for Tasks, I believe a global threadlocal mapping is a better solution, which is what "all tasks" can be.

[alexhartl]

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 everything is cleaned up at latest when the loop is destructed.

Also, the modifications that @kumaraditya303 did lately were a lot about performance improvements. Registering all tasks in a dict, and unregistering in a done callback is quite likely to negate these improvements to some extent, no matter where this dict is stored. Having this set and weakset separate might give us the opportunity to make this behaviour optional, so that the user is able to retain the performance improvements.

Adding new API (like what this PR is doing) means that other loops will have to always implement it (or be broken). Which I obviously can do for uvloop, quite easily, but it will grow the API surface which is already huge.

Yes, true.

Lastly, a minor point: event loop doesn't have a lot to do with tasks. The loop is mostly concerned with running callbacks. Task is a self-contained primitive that just schedules callbacks to the event loop. So it rubs me the wrong way to introduce tracking to the loop for Tasks, I believe a global threadlocal mapping is a better solution, which is what "all tasks" can be.

Technically, that is true. But from a user's perspective it's the loop that basically represents the state of asyncio. I think it would make a lot of sense if tasks are owned by this loop.

[1st1]

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 everything is cleaned up at latest when the loop is destructed.

IMO this isn't a strong argument. Actual applications rarely have more than one event loop during the whole program run (be it a short program or a web server). Now, if there's a bug not cleaning up tasks properly, then the bug is better to be actually fixed, otherwise the fact that event loop clears its tasks won't help at all.

Also, the modifications that @kumaraditya303 did lately were a lot about performance improvements.

What are those modifications? Link?

Registering all tasks in a dict, and unregistering in a done callback is quite likely to negate these improvements to some extent, no matter where this dict is stored.

This is such a minor micro-performance thing in a context of the relatively heavy cost of 'await' and the Task abstraction itself that IMO it's not worth talking about it. Adding/removing a Task from a dict will not be detectable even in a micro-benchmark, it will be below noise floor.

Technically, that is true. But from a user's perspective it's the loop that basically represents the state of asyncio. I think it would make a lot of sense if tasks are owned by this loop.

I see the logic here, but I still think that adding this API to the loop does not make sense in light of other APIs to track tasks that already exist in asyncio: _register_task(task), _enter_task(task).

Also I find adding _-leading APIs to the event loop inelegant.

So bottom line, I'm -1 on merging this PR as is. Please let's work together to re-align it and make it better fit with the existing asyncio APIs.

---

I suggest we keep in place the part of this PR that deterministically calls _unregister_task. _scheduled_tasks becomes a set(). That's it.

[alexhartl]

Sure. Are there other opinions? Otherwise, I can create a PR on this suggestion.