Notes from Scaling MySQL – Up or Out

Here is the quick notes from the session Scaling MySQL – Up or Out ? moderated by Kaj Arno as part of the todays keynote.

Here is the list of panelists are ordered by Alexa ranking.

1. Monty Taylor (MySQL)
2. Matt Ingerenthron (Sun)
3. John Allspaw (Flickr)
4. Farhan Mashraqi (Fotolog)
5. Domas Mituzas (Wkipedia)
6. Jeff Rotheschild (Facebook)
7. Paul Tuckfield (YouTube)

Here is the list of questions and answers from panelists:

Â	How many servers	Number of DBAs	How many web servers	Number of caching servers	Version of MySQL	Language, platform	Operating System
MySQL	1 M, 3 S	1/10	2	2	5.1.23	Perl,php and bash	Linux fedora
Sun	2 clustered, 2 individual	1.5	160+	8	5.0.21	Lots of stuff (java mostly)	Open Solaris
Flickr	166	At present 0	244	14	5.0.51	Php and some Java	Linux
Fotolog	140 databases on 37 instances	10 instances a DBA	70	40 ( 2 on each, 80 total)	4.11 and 4.4	Php, 90% Java	Solaris 10
Wikipedia	20	None, but everybody is kind of a  DBA	70+200	40 ( 2 on each, 80 total)	Â	Php, c++, python	Fedora / Ubuntu
Facebook	30000 databases, 1800 db servers	2	1200	805	5.0.44 with relay log corruption patch	Php, python, c++ and erlang	Fedora / RHEL
Youtube	I can not say	3	I can not say	I can not say	5.0.24	Python	SuSE 9

Few more misc questions …

Number of times re-architected ?

• My: 2 times – 1 time slave, 1 time memcached
• SN: site depend (many times over the year)
• FR: 2.5 (various clusters federated)
• FL: many cached replacements (about to do one change now)
• WK: Never (Spaghetti)
• FB: Every Tuesday, continual
• YT: Pretty continual, 2-3 times (replication, sharding, federation)

What happens if server fails ? what actions you will generally take ..

• FR: All of our servers are federated, pairs of servers, we can loose any one side of shard, we can loose boxes.. traffic goes to either side of shard, now it goes to one, and we will get another one (very transparent to user) 
• WK: Users shout at them on IRC then they moderate … fixed in seconds
• FB: one of 1800-1900 will always fail, just operate well, minor impact, with data going away for a while…we restore from binlog and start the server quickly, promote slave to master and number of ways
• FL: we simply mount the snapshots to different servers and get
• YT: SAN etc, very important data.. recover the server, mirrored disk …mirrored hard drive is crucial

Any recommendation of scaling technology that you wanted to bring

• FL: UltraSPARC-T1 (excellent master, multi threaded) and UltraSPARC-T2 for slave (single threaded)
• WK: good network switch
• FB: cheap switch causes problems and learned lessons, we do not use SAN, neatly partitioned, they scale independently and fail independently
• MY: cluster very sad

Server virtualization ?

• nobody uses at this time
• FB: ETL cluster, we may run more than one in the future

Anything to worry at present ?

• FB: app design is the key to use resources, data center power supply and consumption
• FL: Google has to approve for our lab power (cut app servers by 1/2 by moving from php to java)
• YT: not at all

Any reco, lessons to DBA

• better you know what the systems are, then you can
• performance, scaling taking it serious
• nothing more permanent than temp solutions (if you don’t know when you will fail, then you will )
• architect properly in start, schema, cost of serving data