"David Roland" <david(dot)roland(at)soapware(dot)com> wrote:
> changes to the data folder continue even after the PG_START_BACKUP
> command has been issued. This implies to me that the contents of
> any copy of the data folder may be unreliable. i.e. the copy may
> not reflect the state of the data folder either before the copy
> started or after the copy has finished. It may reflect the state
> of the data folder in some transient form.
Right.
> Assuming this is true, is the copy still usable for restoration?
Yes.
> If so, how does PostgreSQL get the data folder to a stable
> state? Is it by the use of the WAL files that may be created
> during the backup process and the restore.config file?
Exactly. Simplifying somewhat:
- The pg_start_backup causes the WAL position to be remembered. You
will need to start WAL replay at this point.
- Every significant change to a page is WAL-logged.
- Changes after you record the restart point may or may not be in
the base backup.
- After the base backup copying is complete, pg_stop_backup records
the WAL position. You will need to replay WAL *at least* to this
point to have a consistent database.
- WAL replay will bring every page modified after the "start" to a
valid state, whether or not any or all direct modifications to
that page made it into the base backup.
- This actually goes beyond page fix-ups to file creation and
deletion, etc.
-Kevin