On Thu, 9 Oct 2003, Shridhar Daithankar wrote:
> nolan@celery.tssi.com wrote:
>
> >>One of my friend lost data with mysql yesterday.. The machine was taken down for
> >>disk upgrade and mysql apperantly did not commit the last insert.. OK he was
> >>using myisam but still..:-)
> >
> >
> > It sounds like that is more a problem with improper operating protocols
> > than with the underlying database.
>
> No. Problem is machine was shutdown with shutdown -h. It sends sigterm to
> everybody. A good process would flsuh the buffers to disk before finishing.
> Mysql didn't on that occasion.
>
> Transactions or not, this behaviour is unacceptable for any serious app.
>
> >
> > Would PG know enough to do a commit regardless of how the database was
> > shut down? A second question is whether doing a commit is what the user
> > or application would always want to have happen, as it could result in a
> > half-completed transaction.
>
> Do a shutdown -h on a live database machine with pg. It will gracefully shut
> itself down.
If you're running on a properly setup machine (i.e. SCSI and / or battery
backed RAID cache) you can just pull the plug and pg will restart
gracefully. It's one of the standard tests.
Try this on a SCSI based machine with a journaling file system:
pgbench -i -s 10
pgbench -c 200 -t 1000000
wait a minute
start a vacuum
pull the plug
wait a minute
plug it in (turn it on if necessary)
voila, postgresql will recover gracefully with no corruption.
Now ask the consultant to do the same for MySQL. :-)
