From: Tom Lane <tgl@sss.pgh.pa.us>
To: Richard Michael <rmichael@edgeofthenet.org>
cc: pgsql-performance@postgresql.org
Subject: Re: Slow ext'd query via client native implementation vs. libpq &
 simple psql
In-reply-to: 
 <CABR0jERv5L7U0eYyNxv3ZES9HbJj9r2KPK3_=oPMBXTbTC+oXw@mail.gmail.com>
References: 
 <CABR0jERv5L7U0eYyNxv3ZES9HbJj9r2KPK3_=oPMBXTbTC+oXw@mail.gmail.com>
Comments: In-reply-to Richard Michael <rmichael@edgeofthenet.org>
	message dated "Wed, 11 Mar 2020 15:31:48 -0400"
MIME-Version: 1.0
Content-Type: text/plain; charset="us-ascii"
Content-ID: <6261.1583967256.1@sss.pgh.pa.us>
Date: Wed, 11 Mar 2020 18:54:16 -0400
Message-ID: <6262.1583967256@sss.pgh.pa.us>
Precedence: bulk

Richard Michael <rmichael@edgeofthenet.org> writes:
> On PG 12.2, I am analyzing a performance problem when using a client
> (Elixir/postgrex) querying via the extended query protocol.  I am comparing
> with
> psql and a C program.  Logs for all three follow this short explanation.

Hmm, your auto-explain log entries show all three queries having
server-side execution times of a couple dozen msec.  It seems like
the excess time must be involved in transmitting the data to the
client.  So I guess I'd be looking at whether the client is really
slow at absorbing data for some reason.

IIRC, auto-explain's 'actual time' for the top-level query node does not
count the time to format data or transmit it to the client.  Still, we
have an upper bound of ~80 msec for that to happen with the libpq client,
and there's no obvious reason why it'd be different for the other client.

[ thinks for a bit... ]  You might double check that those two clients
are using the same client_encoding setting.  400ms doing encoding
conversion seems excessive, but there aren't that many other possibilities
for the I/O time to be different.

			regards, tom lane