Fix ConnectionPool state after async cancellation

[grynko]

Summary

The PR fixes ConnectionPool state after exceptions/async cancellations.

Root cause analysis

ConnectionPool removes only connections which are:

1. closed
2. idle
3. expired

At the same time, some implementations of AsyncConnectionInterface, e.g. AsyncHTTPConnection don't make actual connection on construction, but are left in semi-valid state with _connection = None. Thus, is_closed()/is_idle/has_expired() return False before actual connection is established or failed during the call to AsyncHTTPConnection.handle_async_request.

Let's take a look at ConnectionPool.handle_async_request method:

try:
    ...
        with self._optional_thread_lock:
            # [POINT 1] the `AsyncHTTPConnection` may be created to handle the request:
            closing = self._assign_requests_to_connections()
        await self._close_connections(closing)

        connection = await pool_request.wait_for_connection(timeout=timeout)

        try:
            # [POINT 2]  the `AsyncHTTPConnection` may create an actual connection:
            response = await connection.handle_async_request(
                pool_request.request
            )
        ...

Please, pay attention, that if the request task is being cancelled between [POINT 1] and [POINT 2], the Schrödinger connection (that is neither opened, nor closed) stays in the pool forever. In the long run (that depends on the load and pool max size), all new requests are refused by the pool, raising PoolTimeout for any request.

Please, see the test_connection_pool_cancellation_during_waiting_for_connection() for a reproducible example.

Solution

Introduce new method of AsyncConnectionInterface - is_connected(). It's not just not is_closed(), but returns False for such semi-constructed connections, allowing the ConnectionPool to collect them.

Checklist

• I understand that this PR may be closed in case there was no previous discussion. (This doesn't apply to typos!)
• I've added a test for each change that was introduced, and I tried as much as possible to make a single atomic change.
• I've updated the documentation accordingly.

[drozdvadym]

Thanks for working on this! We’ve encountered similar issues with connections getting stuck in the pool after async cancellations, so it’s great to see this being addressed. This fix will be helpful—I'm looking forward to it getting merged!