fix: Make sure the Connection.disconnect gets called before the loop is closed

[bellini666]

Fix #328

[carltongibson]

Ah,a PR good... (didn't see this before commenting on the ticket just now 🙂) Sorry I was reading on my phone and only now realise the separate, related reports. 🤦

Thanks for the report. Looks ≈right but let me have a play.

How might we test this? 🤔

[bellini666]

Hi @carltongibson ,

I tested it in a live project. Without this patch everytime I would instantiate the layer object and call group_send inside a django view (not a channel consumer) I would get the error I reported in the issue. That because the pool gets deleted inside the wrapped loop.close, which schedules .disconnect() to be called by the loop, but since it is being closed there it doesn't get a chance to run.

Wi that patch the error don't happen because the connection gets closed in that redis.close() call (which is the correct behaviour, since we don't have more than one connection in that pool. It was created specifically for that conn)

If you are looking for a unit test, I can try to write one. Don't know how to properly mock this to ensure that the coroutine was never run, but I can give it a try.

Note that this fixes the PubSub layer. The core one has a similar issue (#327), which I did not find a good solution for it. The only difference there is that the disconnect does get called, but it is somehow run by a different loop (probably because the loop that created the pool is not the same that called the close_pools function, but I don't know exactly why)

[bellini666]

obs. I'm running the beta versions of channels and channels_redis, and while this is not merged I monkey-patched it to force the same behaviour:

_original_ensure_redis = RedisSingleShardConnection._ensure_redis


def _ensure_redis(self):
    # FIXME: This can be removed once this issue is fixed and released:
    # https://github.com/django/channels_redis/issues/328
    need_patch_redis = self._redis is None
    retval = _original_ensure_redis(self)

    if need_patch_redis:
        self._redis.close = functools.partial(
            self._redis.close,
            close_connection_pool=True,
        )

    return retval


RedisSingleShardConnection._ensure_redis = _ensure_redis

It is running in production in a project that relies a lot in the PubSub behaviour for 3 days now and everything is running fine! :)

[carltongibson]

Thanks for the follow up @bellini666

As I say sounds correct. Let me have a play. 👍

[carltongibson]

Hi @bellini666 — thanks for this.

I added a test that triggers the shutdown warning at least.

🎁