MIME-Version: 1.0
From: Michail Nikolaev <michail.nikolaev@gmail.com>
Date: Fri, 15 Dec 2023 20:07:29 +0100
Message-ID: <CANtu0oiLc-+7h9zfzOVy2cv2UuYk_5MUReVLnVbOay6OgD_KGg@mail.gmail.com>
Subject: Revisiting {CREATE INDEX, REINDEX} CONCURRENTLY improvements
To: PostgreSQL Hackers <pgsql-hackers@postgresql.org>, Alvaro Herrera <alvherre@2ndquadrant.com>
Content-Type: text/plain; charset="UTF-8"
Archived-At: <https://www.postgresql.org/message-id/CANtu0oiLc-%2B7h9zfzOVy2cv2UuYk_5MUReVLnVbOay6OgD_KGg%40mail.gmail.com>
Precedence: bulk

Hello, hackers!

I think about revisiting (1) ({CREATE INDEX, REINDEX} CONCURRENTLY
improvements) in some lighter way.

Yes, a serious bug was (2) caused by this optimization and now it reverted.

But what about a more safe idea in that direction:
1) add new horizon which ignores PROC_IN_SAFE_IC backends and standbys queries
2) use this horizon for settings LP_DEAD bit in indexes (excluding
indexes being built of course)

Index LP_DEAD hints are not used by standby in any way (they are just
ignored), also heap scan done by index building does not use them as
well.

But, at the same time:
1) index scans will be much faster during index creation or standby
reporting queries
2) indexes can keep them fit using different optimizations
3) less WAL due to a huge amount of full pages writes (which caused by
tons of LP_DEAD in indexes)

The patch seems more-less easy to implement.
Does it worth being implemented? Or to scary?

[1]: https://postgr.es/m/20210115133858.GA18931@alvherre.pgsql
[2]: https://postgr.es/m/17485-396609c6925b982d%40postgresql.org