From: | Tom Lane <tgl(at)sss(dot)pgh(dot)pa(dot)us> |
---|---|
To: | David Gould <daveg(at)sonic(dot)net> |
Cc: | Alexander Kuzmenkov <a(dot)kuzmenkov(at)postgrespro(dot)ru>, pgsql-hackers(at)lists(dot)postgresql(dot)org, Alina Alexeeva <alexeeva(at)adobe(dot)com>, Ullas Lakkur Raghavendra <lakkurra(at)adobe(dot)com> |
Subject: | Re: [patch] BUG #15005: ANALYZE can make pg_class.reltuples inaccurate. |
Date: | 2018-03-12 16:21:34 |
Message-ID: | 6981.1520871694@sss.pgh.pa.us |
Views: | Raw Message | Whole Thread | Download mbox | Resend email |
Thread: | |
Lists: | pgsql-hackers |
I wrote:
> Maybe this type of situation is an argument for trusting an ANALYZE-based
> estimate more than the VACUUM-based estimate. I remain uncomfortable with
> that in cases where VACUUM looked at much more of the table than ANALYZE
> did, though. Maybe we need some heuristic based on the number of pages
> actually visited by each pass?
I looked into doing something like that. It's possible, but it's fairly
invasive; there's no clean way to compare those page counts without
altering the API of acquire_sample_rows() to pass back the number of pages
it visited. That would mean a change in FDW-visible APIs. We could do
that, but I don't want to insist on it when there's nothing backing it up
except a fear that *maybe* ANALYZE's estimate will be wrong often enough
to worry about.
So at this point I'm prepared to go forward with your patch, though not
to risk back-patching it. Field experience will tell us if we need to
do more. I propose the attached cosmetic refactoring, though.
regards, tom lane
Attachment | Content-Type | Size |
---|---|---|
analyze-reltuples-change-2.patch | text/x-diff | 8.4 KB |
From | Date | Subject | |
---|---|---|---|
Next Message | Narendra Pradeep U U | 2018-03-12 16:32:52 | Ambigous Plan - Larger Table on Hash Side |
Previous Message | Pavel Stehule | 2018-03-12 16:13:40 | Re: [HACKERS] proposal: schema variables |