Comments on: SnapBack: The joys of backing up MySQL with ZFS…

By: Lars Timmann

Lars Timmann — Fri, 25 Apr 2008 09:17:03 +0000

Still troubles to get things working in 64Bit-mode...
I compiled MySQLdb with:
# CFLAGS="-m64 -B/usr/ccs/bin/amd64" LDFLAGS="-L/lib/64 -R/lib/64 -L/usr/sfw/lib/amd64 -R/usr/sfw/lib/amd64 -static-libgcc -lgcc_s_amd64" /usr/bin/amd64/python2.4 setup.py build
# CFLAGS="-m64 -B/usr/ccs/bin/amd64" LDFLAGS="-L/lib/64 -R/lib/64 -L/usr/sfw/lib/amd64 -R/usr/sfw/lib/amd64 -static-libgcc -lgcc_s_amd64" /usr/bin/amd64/python2.4 setup.py install
It seems to work but I get
# /usr/bin/amd64/python ./mysql_snapback.py mysql_snapback.cfg
Traceback (most recent call last):
File "./mysql_snapback.py", line 44, in ?
import MySQLdb
File "build/bdist.solaris-2.10-i86pc/egg/MySQLdb/__init__.py", line 19, in ?
ImportError: No module named _mysql

But I am far away from python 'til now...

By: Matt Montgomery

Matt Montgomery — Tue, 11 Mar 2008 14:37:08 +0000

Jason Arnel, See: http://dev.mysql.com/tech-r...

By:

Tue, 19 Feb 2008 10:38:17 +0000

Hi Shane,

We've experienced a small number of cases as well, where a rollback recovery was required after restoring the backup. I believe there's also some work being done in the 5.1 series code to make InnoDB more consistent on a FLUSH command. We've never experienced a non-recoverable/corrupt dataset though.

By: Remigusz Sokolowski

Remigusz Sokolowski — Mon, 11 Feb 2008 06:59:03 +0000

I think, Shane, that InnoDB would do well even without FLUSH TABLES WITH READ LOCK. Actually I have made something similar and FLUSH [..] was performed only to provide consistent MyISAM tables, InnoDB should simply recover helping itself with own transactional changelog – of course You need to be sure to have not broken D in ACID

By: Shane Milton

Shane Milton — Mon, 04 Feb 2008 05:50:59 +0000

Great script, however I'd be careful trusting the innodb snapshots even with "FLUSH TABLES WITH READ LOCK" I've run into several instances where innodb still had data in memory that didn't seem to make it to the data files. Maybe 1 in 12 snapshots innodb datafiles where corrupt ( i was using a netapp in my case ) just my $.02

By:

Mon, 21 Jan 2008 17:30:16 +0000

joe: I have not, but it'd be great if it worked. The path/name of the ZFS userspace command is configurable in the configuration file. If the syntax for the command is the same on FreeBSD it should work.

yeek: Depends on how you want to arrange your set up. The most straightforward would be to create a ZFS filesystem for each MySQL database directory under the master ZFS FS. Then create a SnapBack configuration for each FS. Alternately, you could copy the DB files you want out of the snapshot back into the writeable filesystem. We run multiple DBs on a single FS and snapshot that. We copy from the snapshot to long term disk-to-disk backup over GigE.

By: yeek

yeek — Mon, 21 Jan 2008 13:59:29 +0000

It looks to me the Snapback is actually is taking a Snapshot of all the files of MySQL database. This might work well if your instance only contains one database. However, in many cases, we actually have multiple databases in one instance, and we only want to restore one database on the instance?

How does the Snapback work in this case?

By: joe

joe — Sun, 20 Jan 2008 22:07:18 +0000

Have you tried it on FreeBSD? FreeBSD can use ZFS now as well (ZVF v6 not v9/10 as with Solaris). I hope it can be made to work!

By:

Fri, 18 Jan 2008 16:21:15 +0000

Selim: I hope it works for you. I wish I was behind the Sun/MySQL deal. Actually we're pretty excited about it. With PostgreSQL getting all the Sun-love till now, its really a phenomenal announcement. Hopefully, the DTrace probes won't have to wait till 6.0 now.

vnull: Somehow I never noticed the lack of RSS/Atom in the template. It should be fixed now. Thank you.

Jason: UFS is slightly better performing in reads. However, for writes we find ZFS to have superior performance. But, even if ZFS had 30% less performance the trade-off would be worth it for the features. We save so much money on snapshot licenses that adding extra disks to close the performance gap would still be cheaper.

By: remigiusz sokolowski

remigiusz sokolowski — Fri, 18 Jan 2008 07:29:28 +0000

That is right, that UFS performs better than ZFS in terms of using it with a database, but… - IO is most important on DSS databases - with OLTP the most important thing is to keep with proper buffering and then the IO performance seems to be less critical factor - backup is especially hard to take on OLTP databases, which must work 24×7x365