Re: WIP: Faster Expression Processing v4

From: Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us>
To: Andres Freund <andres(at)anarazel(dot)de>
Cc: pgsql-hackers(at)postgresql(dot)org, Heikki Linnakangas <hlinnaka(at)iki(dot)fi>, Tomas Vondra <tomas(dot)vondra(at)2ndquadrant(dot)com>
Subject: Re: WIP: Faster Expression Processing v4
Date: 2017-03-15 22:48:28
Message-ID: 10199.1489618108@sss.pgh.pa.us
Views: Raw Message | Whole Thread | Download mbox
Thread:
Lists: pgsql-hackers

Andres Freund <andres(at)anarazel(dot)de> writes:
> On 2017-03-15 18:16:57 -0400, Tom Lane wrote:
>> [ scratches head... ] What deforming logic do you think I'm proposing
>> removing?

> I thought you were suggesting that we don't do the get_last_attnums (and
> inlined version in the isSimpleVar case) at execution time anymore,
> instead relying on logic in the planner to know how much to deform ahead
> of time. Then we'd do slot_getsomeattrs in the appropriate places. But
> I understood you suggesting to do so only in scan nodes - which doesn't
> seem sufficient, due to the use of materialized / minimal tuples in
> other types of nodes. Did I misunderstand?

We would need to do it anywhere that we might be projecting from a
materialized tuple, I suppose. From the planner's standpoint it would be
about as easy to do this for all plan nodes as only selected ones.

Anyway the core point here seems to be that skipping ExecProject misses a
bet because the underlying tuple doesn't get disassembled. A quick-hack
way of seeing if this helps might be to do slot_getallattrs in the places
where we skip ExecProject. I'm not sure we'd want that as a permanent
solution, because it's possible that it extracts columns not actually
needed, but it should at least move the hotspot in your example case.

regards, tom lane

In response to

Responses

Browse pgsql-hackers by date

  From Date Subject
Next Message Julien Rouhaud 2017-03-15 22:55:42 Re: pg_stat_wal_write statistics view
Previous Message Simon Riggs 2017-03-15 22:27:59 Re: Measuring replay lag