Date: Fri, 9 Oct 2020 11:46:25 -0700
From: Andres Freund <andres@anarazel.de>
To: Peter Eisentraut <peter.eisentraut@2ndquadrant.com>
Cc: pgsql-hackers <pgsql-hackers@postgresql.org>
Subject: Re: dynamic result sets support in extended query protocol
Message-ID: <20201009184625.4hfgpcxzd6jfredb@alap3.anarazel.de>
References: <6e747f98-835f-2e05-cde5-86ee444a7140@2ndquadrant.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <6e747f98-835f-2e05-cde5-86ee444a7140@2ndquadrant.com>
Precedence: bulk

Hi,

On 2020-10-08 09:46:38 +0200, Peter Eisentraut wrote:
> New would be that the server would now also respond with a new message, say,
> 
> S: DynamicResultInfo

> Now, if the client has seen DynamicResultInfo earlier, it should now go into
> a new subsequence to get the remaining result sets, like this (naming
> obviously to be refined):

Hm. Isn't this going to be a lot more latency sensitive than we'd like?
This would basically require at least one additional roundtrip for
everything that *potentially* could return multiple result sets, even if
no additional results are returned, right? And it'd add at least one
additional roundtrip for every result set that's actually sent.

Is there really a good reason for forcing the client to issue
NextResult, Describe, Execute for each of the dynamic result sets? It's
not like there's really a case for allowing the clients to skip them,
right?  Why aren't we sending something more like

S: CommandPartiallyComplete
S: RowDescription
S: DataRow...
S: CommandPartiallyComplete
S: RowDescription
S: DataRow...
...
S: CommandComplete
C: Sync

gated by a _pq_ parameter, of course.


> I think this would all have to use the unnamed portal, but perhaps there
> could be other uses with named portals.  Some details to be worked out.

Which'd avoid this too, but:

> One thing that's missing in this sequence is a way to specify the desired
> output format (text/binary) for each result set.

Is a good point. I personally think avoiding the back and forth is more
important though. But if we could address both at the same time...


> (I suspect what would be more useful in practice is to designate
> output formats per data type.)

Yea, that'd be *really* useful. It sucks that we basically require
multiple round trips to make realistic use of the binary data for the
few types where it's a huge win (e.g. bytea).

Greetings,

Andres Freund