On Thu, Jan 30, 2020 at 06:11:22PM +0100, Michael Banck wrote:
> The other question is whether it is possible to end up with a
> pg_internal.init.$PID file in a running cluster. E.g. if an instance
> crashes and gets started up again - are those cleaned up during crash
> recovery, or should pg_checksums ignore them? Right now pg_checksums
> only checks against a list of filenames and only skips on exact matches
> not prefixes so that might take a bit of work.
Indeed, with a bad timing and a crash in the middle of
write_relcache_init_file(), it could be possible to have such a file
left around in the data folder. Having a past tablespace version left
around after an upgrade is a pilot error in my opinion because
pg_upgrade generates a script to cleanup past tablespaces, no? So
your patch does not look like a good idea to me. And now that I look
at it, if we happen to leave behind a temporary file for
pg_internal.init, backups fail with the following error:
2020-01-31 13:26:18.345 JST [102076] 010_pg_basebackup.pl ERROR:
invalid segment number 0 in file "pg_internal.init.123"
So, I think that it would be better to change basebackup.c,
pg_checksums and pg_rewind so as files are excluded if there is a
prefix match with the exclude lists. Please see the attached.
--
Michael