Infinite hang on leaving conn.transaction() block

Hi, I am experiencing hangs at aexit of Transaction context manager, with traceback as follows:

...
  File "/usr/local/lib/python3.7/dist-packages/asyncpg/transaction.py", line 89, in __aexit__
    await self.__rollback()
  File "/usr/local/lib/python3.7/dist-packages/asyncpg/transaction.py", line 198, in __rollback
    await self._connection.execute(query)
  File "/code/callmeback/detail/log_helpers/postgre_conn.py", line 53, in execute
    return await super(LoggingConn, self).execute(query, *args, timeout=timeout)
  File "/usr/local/lib/python3.7/dist-packages/asyncpg/connection.py", line 273, in execute
    return await self._protocol.query(query, timeout)

Default command timeout is set for connection:

>>> conn = frame.f_locals['self']
>>> conn._config
ConnectionConfiguration(command_timeout=10.0, statement_cache_size=100, max_cached_statement_lifetime=300, max_cacheable_statement_size=15360)
>>> conn._protocol._get_timeout(None)
10.0

So I expect self._protocol.query() to throw TimeoutError of some kind, but related coroutine hanged for three days with no progress.

I will happily provide you with any additional information, if I can. My app has aiomonitor onboard, so if any info reachable that way will help you - please let me know.

I tried to fiddle with conn._protocol, but almost nothing of its state is reachable from aioconsole since it's Cython.

My suspection is that exception that leads to rollback is coming from in-transaction query that somehow leads connection in such a state that rollback() cannot be processed normally. I will provide you with information about exception once that problem triggers next time.

• asyncpg version: 0.18.3
• PostgreSQL version: 10.7
• Do you use a PostgreSQL SaaS? If so, which? Can you reproduce
  the issue with a local PostgreSQL install?:
• Python version: 3.7
• Platform: Ubuntu 16.04
• Do you use pgbouncer?: Yes (session-level pooling)
• Did you install asyncpg with pip?: Yes
• If you built asyncpg locally, which version of Cython did you use?: --
• Can the issue be reproduced under both asyncio and
  uvloop?: Yes, I am using uvloop

Start with asyncpg/transaction.py at aexit and rollback, then follow connection.py execute() into _protocol.query(). Reproduce the hang with PostgreSQL 10.7, session-level pgbouncer pooling, and uvloop if possible; done means the rollback path either completes or surfaces the configured timeout instead of hanging.