Feature Request: Retry on failed acquires (timeouterror/connectionerror)

• asyncpg version: 0.21.0
• PostgreSQL version: any
• Do you use a PostgreSQL SaaS? If so, which? Can you reproduce
  the issue with a local PostgreSQL install?: -- dont use it
• Python version: any
• Platform: any
• Do you use pgbouncer?: no
• Did you install asyncpg with pip?: yes
• If you built asyncpg locally, which version of Cython did you use?: not built locally
• Can the issue be reproduced under both asyncio and
  uvloop?: yes

Would you accept a contribution to asyncpg's Pool in which we can retry failed acquires (timeouterrors/connectionresetbypeer/...) which are caused by network issues?
The default behaviour wouldn't retry, as it's happening right now.

This is useful for data processing pipelines with multiple acquires in it's flow (distinct dbs): Being able to configure acquire_max_retries when creating the pool would help keeping the code simpler and not have to wrap all our 'acquires' with retries. The retry decorator logic on acquire can become non-trivial because of the contextmanager auto-releasing the connection after successful usage OR other errors (unrelated to asyncpg) that happen after acquiring a connection+exiting the context (this caused some pain for us).

If you are willing to accept this feature, I'm happy to implement it on my own and submit a PR.

Just to give you an idea, this is what we had to implement on our project in order to have pool acquires with retries:

async def retry(coro, exceptions=(Exception,), max_retries=3, retry_delay_seconds=0):
    for attempt in range(max_retries):
        try:
            return await coro()
        except exceptions:
            logger.exception(f"{coro} attempt failed. Attempt {attempt + 1}/{max_retries}")
            if attempt + 1 == max_retries:
                raise
        await asyncio.sleep(retry_delay_seconds)


class PoolWithRetries:
    def __init__(self, pool: asyncpg.pool.Pool, max_retries=3, retry_delay_seconds=0):
        self.pool = pool
        self.max_retries = max_retries
        self.retry_delay_seconds = retry_delay_seconds

    @asynccontextmanager
    async def acquire(self, *args, **kwargs):
        # Reimplementation of PoolAcquireContext with retries
        # Only supports `async with this.acquire() as con` syntax
        con = await retry(
            partial(self.pool.acquire, *args, **kwargs),
            exceptions=(asyncio.TimeoutError, ConnectionError),
            max_retries=self.max_retries,
            retry_delay_seconds=self.retry_delay_seconds,
        )
        try:
            yield con
        finally:
            await self.pool.release(con)

    async def release(self, *args, **kwargs):
        await self.pool.release(*args, **kwargs)

    async def close(self):
        await self.pool.close()

The implementation becomes much simpler if we can do it inside asyncpg.

Commencez par le point d’entrée asyncpg Pool.acquire et le comportement de PoolAcquireContext décrit dans l’issue. Définissez comment les nouvelles tentatives d’acquire configurables doivent gérer les erreurs de délai d’expiration et de connexion sans réessayer les erreurs qui ne sont pas liées ni perturber la libération du context manager, puis ajoutez une couverture pour le comportement de nouvelle tentative accepté avant de considérer la fonctionnalité comme complète.