Feature request: Support for asynchronous command processing

libpq has support for something called asynchronous command processing (https://www.postgresql.org/docs/current/libpq-async.html).

Thea idea is fairly simple: you can queue up multiple queries on one connection without waiting for the previous result and then wait for all the results at the end. This avoids waiting for network round trips between queries, which at least in theory should have the same performance benefits as running multiple queries with one network round trip.

It would be useful if asyncpg could support the same concept. The Python API should be fairly straightforward. You just start multiple coroutines with queries and then await all of them together at the end. Something like this:

q1 = conn.fetch('SELECT $1', 1)
q2 = conn.fetch('SELECT $1', 2)
q3 = conn.fetch('SELECT $1', 3)
results = await asyncio.gather(q1, q2, q3) # (currently this raises InterfaceError)

This could have significant performance benefits in situations where you make multiple fast independent queries, especially if there is a large network round-trip delay.

AFAIK implementing this is definitely possible in theory. In practice, the complexity depends a lot on how the internals of asyncpg are structured, which I am not familiar with.

EDIT: This request is motivated by the same goal as https://github.com/MagicStack/asyncpg/issues/839, but my proposal is to specifically use asynchronous command processing without pipelining. If my understanding of the extended query protocol is correct, it should be possible to issue multiple queries without waiting for results between them but with a sync point after each query, which helps avoid the complex error handling rules of pipeline mode. (If my understanding of extended query protocol is not correct, this may be impossible. Feel free to correct me.)

Comienza con la documentación sobre el procesamiento asíncrono de comandos de libpq enlazada en el issue; después inspecciona la API de conexión de asyncpg alrededor de conn.fetch y el comportamiento actual de InterfaceError cuando se reúnen coroutines. Compara el comportamiento propuesto sin pipelining con el extended query protocol y determina si la API y la semántica de sincronización solicitadas son viables; el issue no proporciona ningún punto de entrada de archivo o de prueba.