Re: Detecting some cases of missing backup_label

On 12/18/23 10:39, Stephen Frost wrote:
> Greetings,
> 
> * Stephen Frost (sfrost@snowman.net) wrote:
>> * Andres Freund (andres@anarazel.de) wrote:
>>> I recently mentioned to Robert (and also Heikki earlier), that I think I see a
>>> way to detect an omitted backup_label in a relevant subset of the cases (it'd
>>> apply to the pg_control as well, if we moved to that).  Robert encouraged me
>>> to share the idea, even though it does not provide complete protection.
>>
>> That would certainly be nice.
>>
>>> The subset I think we can address is the following:
>>>
>>> a) An omitted backup_label would lead to corruption, i.e. without the
>>>     backup_label we won't start recovery at the right position. Obviously it'd
>>>     be better to also catch a wrong procedure when it'd not cause corruption -
>>>     perhaps my idea can be extended to handle that, with a small bit of
>>>     overhead.
>>>
>>> b) The backup has been taken from a primary. Unfortunately that probably can't
>>>     be addressed - but the vast majority of backups are taken from a primary,
>>>     so I think it's still a worthwhile protection.
>>
>> Agreed that this is a worthwhile set to try and address, even if we
>> can't address other cases.
>>
>>> Here's my approach
>>>
>>> 1) We add a XLOG_BACKUP_START WAL record when starting a base backup on a
>>>     primary, emitted just *after* the checkpoint completed
>>>
>>> 2) When replaying a base backup start record, we create a state file that
>>>     includes the corresponding LSN in the filename
>>>
>>> 3) On the primary, the state file for XLOG_BACKUP_START is *not* created at
>>>     that time. Instead the state file is created during pg_backup_stop().
>>>
>>> 4) When replaying a XLOG_BACKUP_END record, we verif that the state file
>>>     created by XLOG_BACKUP_START is present, and error out if not.  Backups
>>>     that started before the redo LSN from backup_label are ignored
>>>     (necessitates remembering that LSN, but we've been discussing that anyway).
>>>
>>>
>>> Because the backup state file on the primary is only created during
>>> pg_backup_stop(), a copy of the data directory taken between pg_backup_start()
>>> and pg_backup_stop() does *not* contain the corresponding "backup state
>>> file". Because of this, an omitted backup_label is detected if recovery does
>>> not start early enough - recovery won't encounter the XLOG_BACKUP_START record
>>> and thus would not create the state file, leading to an error in 4).
>>
>> While I see the idea here, I think, doesn't it end up being an issue if
>> things happen like this:
>>
>> pg_backup_start -> XLOG_BACKUP_START WAL written -> new checkpoint
>> happens -> pg_backup_stop -> XLOG_BACKUP_STOP WAL written -> crash
>>
>> In that scenario, we'd go back to the new checkpoint (the one *after*
>> the checkpoint that happened before we wrote XLOG_BACKUP_START), start
>> replaying, and then hit the XLOG_BACKUP_STOP and then error out, right?
>> Even though we're actually doing crash recovery and everything should be
>> fine as long as we replay all of the WAL.
> 
> Andres and I discussed this in person at PGConf.eu and the idea is that
> if we find a XLOG_BACKUP_STOP record then we can check if the state file
> was written out and if so then we can conclude that we are doing crash
> recovery and not restoring from a backup and therefore we don't error
> out.  This also implies that we don't consider PG to be recovered at the
> XLOG_BACKUP_STOP point, if the state file exists, but instead we have to
> be sure to replay all WAL that's been written.  Perhaps we even
> explicitly refuse to use restore_command in this case?
> 
> We do error out if we hit a XLOG_BACKUP_STOP and the state file
> doesn't exist, as that implies that we started replaying from a point
> after a XLOG_BACKUP_START record was written but are working from a copy
> of the data directory which didn't include the state file.
> 
> Of course, we need to actually implement and test these different cases
> to make sure it all works but I'm at least feeling better about the idea
> and wanted to share that here.

I've run this through a bunch of scenarios (in my head) with parallel 
backups and it does seem to hold up.

I think we'd need to write the state file before XLOG_BACKUP_START just 
in case. Seems better to have an extra state file rather than have one 
be missing.

As you say, we'll need to store redo for the last recovered backup in 
pg_control. I guess it would be OK to remove that when the cluster is 
promoted. As long as recovery is going on seems like it would always be 
possible to hit an XLOG_BACKUP_STOP for an even longer running backup.

I'm a little worried about what happens if a state file goes missing, 
but I guess that could be true of any file in PGDATA.

Probably we'd want to exclude *all* state files from backups, though. 
Seems like in various PITR scenarios it could be hard to determine when 
to remove them.

Regards,
-David