On 7/22/22 23:36, Michael Paquier wrote:
> On Fri, Jul 22, 2022 at 04:24:54AM -0500, Zsolt Ero wrote:
>> I've opened an issue on pgbackrest, where the developer confirmed that
>> those files are not touched by pgbackrest, they are basically copied back
>> exactly as they were. The backup/restore is on the same partition as the
>> data folder, so no chance of file system changes, etc. Single master, no
>> replication.
>>
>> I'm always doing clean data dir + init_db so there is nothing there before
>> the restore. I can replicate this behaviour 100%.
>>
>> PostgreSQL version: psql (PostgreSQL) 12.11 (Ubuntu 12.11-1.pgdg18.04+1)
>>
>> OS is Ubuntu 18.04
>>
>> pgbackrest issue: https://github.com/pgbackrest/pgbackrest/issues/1815
>>
>> log details:
>>
>> could not link file "pg_wal/000000010000015600000098" to
>> "pg_wal/00000001000001570000006E": File exists
>
> FWIW, the backend code has protections to prevent *exactly* this kind
> of problems when recycling WAL segment files at checkpoints with a set
> of LWLocks taken on the control file, for one. Perhaps you have
> messed up things and you have finished in such a state that backrest
> writes to pg_wal/ concurrently with a cluster running and running a
> checkpoint, which would explain those link() calls to be failing?
During recovery pgbackrest only writes into pg_wal when invoked by
restore_command. That generally means pgbackrest should only be writing
to RECOVERYXLOG.
Zsolt, can you add --log-level-console=info to your archive_command so
we can see what parameters are being passed to pgbackrest? This logging
will appear in the postgres log.
Regards,
-David
