Horton

Re: Flush pgstats file during checkpoints

Bertrand Drouvot <bertranddrouvot.pg@gmail.com>

View thread View on postgr.es
From: Bertrand Drouvot <bertranddrouvot.pg@gmail.com>
To: Michael Paquier <michael@paquier.xyz>
Cc: Tomas Vondra <tomas.vondra@enterprisedb.com>, Konstantin Knizhnik <knizhnik@garret.ru>, Postgres hackers <pgsql-hackers@lists.postgresql.org>
Date: 2024-07-12T13:01:19Z
Lists: pgsql-hackers 
Hi,

On Fri, Jul 12, 2024 at 12:10:26PM +0000, Bertrand Drouvot wrote:
> Need to spend more time and thoughts on 0002+.

I think there is a corner case, say:

1. shutdown checkpoint at LSN1
2. startup->reads the stat file (contains LSN1)->all good->read stat file and
remove it
3. crash (no checkpoint occured between 2. and 3.) 
4. startup (last checkpoint is still LSN1)->no stat file (as removed in 2.) 

In that case we start with empty stats.

Instead of removing the stat file, should we keep it around until the first call
to pgstat_write_statsfile()?

Regards,

-- 
Bertrand Drouvot
PostgreSQL Contributors Team
RDS Open Source Databases
Amazon Web Services: https://aws.amazon.com

Commits

  1. Add redo LSN to pgstats files

  2. Add more debugging information when failing to read pgstats files

  3. Make write of pgstats file durable at shutdown

  4. Add assertion in pgstat_write_statsfile() about processes allowed