Re: Remove Deprecated Exclusive Backup Mode
David Steele <david@pgmasters.net>
From: David Steele <david@pgmasters.net>
To: pgsql-hackers@lists.postgresql.org
Date: 2018-12-11T22:24:15Z
Lists: pgsql-hackers
Commits
Same data as JSON:
GET /api/v1/messages/:b64id/commits
the thread's linked commits as JSON, with link sources.
API reference →
-
Remove exclusive backup mode
- 39969e2a1e4d 15.0 landed
Attachments
- remove-exclusive-backup.patch (text/plain) patch
On 11/26/18 10:13 PM, David Steele wrote: > Hackers, > > I propose we remove the deprecated exclusive backup mode of > pg_start_backup()/pg_stop_backup() for Postgres 12. > > The exclusive backup mode has a serious known issue. If Postgres > terminates ungracefully during a backup (due to hardware, kernel, > Postgres issues, etc.) then Postgres may refuse to restart. > > The reason is that the backup_label file will usually reference a > checkpoint LSN that is older than the WAL available in pg_wal. Postgres > does emit a helpful error message while PANIC'ing but that's cold > comfort to an admin who must manually intervene to get their cluster > operational again. > > The deprecated exclusive mode promises to make a difficult problem > simple but doesn't live up to that promise. That's why it was replaced > externally in 9.6 and why pg_basebackup has not used exclusive backups > since it was introduced in 9.2. > > Non-exclusive backups have been available since 9.6 and several > third-party solutions support this mode, in addition to pg_basebackup. > > The recently introduced recovery changes will break current automated > solutions so this seems like a good opportunity to make improvements on > the backup side as well. > > I'll submit a patch for the 2019-01 commitfest. Attached is the patch. I was a bit surprised by how much code went away. There was a lot of code dedicated to making sure that backup_label got renamed on shutdown, that there was not an exclusive backup running, etc. There were no tests for exclusive backup so the test changes were minimal. I did have to replace one "hot backup" in 010_logical_decoding_timelines.pl. I'm not sure why the backup was being done this way, or why the standby needs a copy of pg_replslot (which is not copied by pg_basebackup). It might be worth looking at but it seemed out of scope for this patch. Regards, -- -David david@pgmasters.net