[DRBD-user] DRBD below or on top of LVM
H.D.
devnull at deleted.on.request
Thu Feb 15 14:38:45 CET 2007
OK, thank you for your detailed answer again, it really helps me!
On 15.02.2007 14:34, Graham Wood wrote:
>
>> To me that reads `LVM is only capable of creating an "almost
>> consistent" snapshot'.
> Of a running MySQL database... The issue is at the MySQL level, rather
> than the LVM level. Without grabbing the read lock that document
> mentions, you can easily have transactions partially commited to disk,
> without a consistent state on disk.
>
>> Is that true?
> Restoring from a snapshot is the same (almost) as recovering from a
> power failure at the time the snapshot was taken. If (for example)
> MySQL is in the process of commiting a transaction in 2 places on the
> disk when the snapshot is taken, because it is instantaneous you might
> only have 1 of them commited at the time the snapshot is taken, and
> it'll be inconsistent.
>
> In the case of Oracle you can put the database into a hot backup mode,
> take the snapshot, and then take it out. When you do the recovery you
> can then replay the log it generated in that time to recover. This was
> intended for tape backups (since they can take a long time, and
> therefore will NOT be consistent), and actually allows you to replay to
> any time between start and finish....
>
> Looking on google for references to postgreSQL and snapshot backups, it
> looks like you've got the same functionality as provided by InnoDB on
> MySQL (and pretty much every other DB out there) - log replay
> functionality all the time.
>
> As long as the logs are snapshotted at the same time as the main
> database, on recovery it can replay the log to get back to a consistent
> state. If you've got the log files on a different device to the main
> database then you ideally want to snapshot the two at exactly the same
> time (theoretically impossible)....
>
> However, I think if you snapshot the main device, and then the log
> files, you are almost definitely going to be "OK". The only problem
> would be if the logs had rotated and lost information from when the main
> devices were snapshotted. At that point, your snapshots are going to be
> smoking anyway, so you're going to have other problems.
>
> Graham
--
Regards,
H.D.
More information about the drbd-user
mailing list