Horton

Re: Damage control for planner's get_actual_variable_endpoint() runaway

Tom Lane <tgl@sss.pgh.pa.us>

View thread View on postgr.es
From: Tom Lane <tgl@sss.pgh.pa.us>
To: Andres Freund <andres@anarazel.de>
Cc: Robert Haas <robertmhaas@gmail.com>, Simon Riggs <simon.riggs@enterprisedb.com>, pgsql-hackers@lists.postgresql.org, Jakub Wartak <jakub.wartak@enterprisedb.com>, Alvaro Herrera <alvherre@alvh.no-ip.org>, PostgreSQL Hackers <pgsql-hackers@postgresql.org>
Date: 2022-11-21T22:15:53Z
Lists: pgsql-hackers 
Andres Freund <andres@anarazel.de> writes:
> On 2022-11-21 16:17:56 -0500, Robert Haas wrote:
>> But ... what if they're not? Could the index contain a large number of
>> pages containing just 1 tuple each, or no tuples at all? If so, maybe
>> we can read ten bazillion index pages trying to find each heap tuple
>> and still end up in trouble.

> ISTM that if you have an index in such a poor condition that a single
> value lookup reads thousands of pages inside the index, planner
> estimates taking long is going to be the smallest of your worries...

Yeah, that sort of situation is going to make any operation on the
index slow, not only get_actual_variable_endpoint().

I think we should content ourselves with improving the demonstrated
case, which is where we're forced to do a lot of heap fetches due
to lots of not-all-visible tuples.  Whether we can spend a lot of
time scanning the index without ever finding a tuple at all seems
hypothetical.  Without more evidence of a real problem, I do not
wish to inject warts as horrid as this one into the index AM API.

			regards, tom lane

Commits

  1. YA attempt at taming worst-case behavior of get_actual_variable_range.

  2. Improve performance of get_actual_variable_range with recently-dead tuples.

  3. Use SnapshotDirty rather than an active snapshot to probe index endpoints.