sql/pgwire: responses to commands should be coalesced

[knz]

Noticed while comparing pg and cockroach network traces, e.g. in #4538 (comment)

The responses to a combined command packet are sent as separate/individual packets by cockroach, whereas pg keeps the responses combined. The extra packets mean more network traffic / overhead.

Responses should be coalesced for performance.

[petermattis]

Strange. Looks to me like we only flush the write buffer immediately before trying to read another command from the client.

[petermattis]

Ah, it looks like postgres is trying to read the next command from the client before flushing the response to the previous command. We could certainly do that as well, though it is mildly irritating to do from Go because the I/O is blocking. What we could do is check to see if the read buffer has another command buffered before flushing the response to the previous command.

[bdarnell]

This is what Nagle's algorithm is for. Go's net package turns TCP_NODELAY on by default; maybe we should turn it off? Although it would be good to be able to flush the outgoing buffer before we go into a blocking read.

[knz]

( While looking at #4036 )