On Wed, Feb 10, 2016 at 7:46 AM, Magnus Hagander <magnus@hagander.net> wrote:
> Per discussionat the developer meeting in Brussels, here's a patch that
> makes some updates to the backup APIs, to support non-exclusive backups
> without using pg_basebackup. The idea is to fix at least three main issues
> that are there today -- that you cannot run concurrent backups, that the
> backup_label file is created in the data directory which makes it impossible
> to distinguish between a cluster restored from backup and one that crashed
> while a backup was running, and a cluster can get "stuck" in backup mode if
> the backup script/software crashes.
>
> To make this work, this patch:
>
> * Introduces a new argument for pg_start_backup(), for "exclusive". It
> defaults to true, which means that just calling pg_start_backup() like
> before will behave exactly like before. If it's called with exclusive='f', a
> non-exclusive backup is started, and the backup label is collected in memory
> instead of in the backup_label file.
>
> * If the client disconnects with a non-exclusive backup running, the backup
> is automatically aborted. This is the same thing that pg_basebackup does. To
> use these non-exclusive backups the backup software will have to maintain a
> persistent connection to the database -- something that should not be a
> problem for any of the modern ones out there (but would've been slightly
> trickier back in the days when we suggested shellscripts)
>
> * A new version of pg_stop_backup is created, taking an argument specifying
> if it's exclusive or not. The current version of pg_stop_backup() continues
> to work for exclusive backups, with no changes to behavior. The new
> pg_stop_backup will return a record of three columns instead of just the
> value -- the LSN (pglsn), the backup label file (text) and the tablespace
> map file (text). If used to cancel an exclusive backup, backup label file
> and tablespace map will be NULL. At this point it's up to the backup
> software to write down the files in the location of the backup.
>
>
> Attached patch currently doesn't include full documentation, since Bruce was
> going to restructure that section of the docs (also based on the
> devmeeting). I will write up the docs once that is done (I assume it will be
> soon enough, or I'll go do it regardless), but I wanted to get some review
> in on the code while waiting.
Wow, this is a great idea.
--
Robert Haas
EnterpriseDB: http://www.enterprisedb.com
The Enterprise PostgreSQL Company