Don’t prematurely optimise Just remember the 7P’s:Prior & Proper Planning Prevents Piss Poor Performance
Reference Architectures• Is there one for the Web world?• Your choices are to: • scale up • scale out• Which do you pick?
Scaling out• Buying (renting) commodity hardware• Using the cloud to expand • Or using the cloud totally: e.g. http:// heroku.com/facebook
OS• If you didn’t go opensource, you’re silly• Tuning Linux/BSD is mandatory • ﬁlesystem: xfs, ext3(4) • swap is the devil • different schedulers work better for different tasks (web, database, etc.) • NFS? You’d better tune that! (stopgap, scaling is hard)
Web server• Apache, lighthttpd, nginx• They all require conﬁguration (httpd.conf)• Simple things like maximum connections, worker MPM, usually go unconﬁgured
Language• “PHP doesn’t scale.” - Cal Henderson, when he was at Flickr.com• Languages are not meant to scale for you• Use bytecode caches (PHP, Python, etc.)• Compile away -- HipHop• Library, driver support; developer communities
Databases• are slow, period.• partition data into shards• tune that database
How do you scale easily?• Use caches• Disk-based caching (cache_lite via php- pear). RAM disks on SSDs... fast!• In-memory caching (APC, memcached)• Cloud-based caching (S3, MogileFS)
memcached• Easy to setup and use• Very fast over the network• Scales, has failover, widely supported• Centralised and shared across the site
S3• Databases are good for storing relational data, but suck for blob storage• S3 is a ﬁle & data store, running over HTTP• In theory, inﬁnitely scalable• Centralised & shared across site• Costs money, no Malaysian POP• See OpenStack’s Swift Object Store
CDN• Outsource it• Costs a lot of money• Aﬂexi is a Malaysian company making a pretty darn good CDN (resold via Exabytes?)• Out of your control but will help you scale, scale, scale
Back to the database...• Sharding • not all data lives in one place • hash records to partitions • partition alphabetically? put n-users/ shard? organise by postal code?• horizontal vs vertical partitioning
Horizontal vs Vertical Partitioning192.168.0.1 192.168.0.1 User User id int(10) Better if INSERT id int(10) username char(15) username char(15) password char(15) email char(50) 192.168.0.3 heavy and there’s password char(15) email char(50) User less frequently192.168.0.2 id int(10) 192.168.0.2 changed data username char(15) password char(15) User email char(50) UserInfo id int(10) login datetime username char(15) md5 varchar(32) password char(15) guid varchar(32) email char(50)
MySQL has engines• InnoDB (XtraDB) for transactional use• MyISAM for “data warehousing” use• Maria in time
MySQL has replication!• Simple, easy to implement (async)• Row based replication is better than statement based replication• You do not need mysql cluster (ndb)• Look at Tungsten Replicator, Galera, etc. for other topologies (e.g. many masters)
Use INDEXes• Covering index: all ﬁelds in SELECT for speciﬁc table are contained in index • EXPLAIN will say “Using index”
Monitor everything!• Benchmarking allows tracking performance over time• Nagios• MySQL (MariaDB/Percona Server) • slow query log, extended stats in slow query log, use EXPLAIN, microsecond process list, userstats v2, SHOW PROCESSLIST, etc.
Fulltext Search• Don’t use the database!• Sphinx • SphinxSE for MariaDB• Lucene
Don’t• SELECT * FROM room WHERE room_date BETWEEN ‘2011-02-25’ AND ‘2011-02-27’ • not have an INDEX on ﬁeld being operated on by range operator => full table scan• not allocate a primary key• over-normalise (3NF is ﬁne)
Keeping state• Session data in DB • PHP has ﬁles, doesn’t scale. DB +Memcached goes far• Replicate/Partition/Cache state• Cookies can be validated by checksums and timestamps (encryption consumes CPU)
General advice• Your DB servers are not your web servers and they’re not your load balancers• Write non-locking code• Don’t block loading unnecessarily• Cache partially (esp. w/dynamic pages)• Use UTC for time (replication across geographies?)• Keep everything in version control• Migrations are never recommended unless you’ve exceeded capabilities of current solutions. Beware v2 disasters.
NoSQL• MongoDB "I dont foresee StumbleUpon ever giving up on all of its MySQL instances. RDBMSs are just too useful. The plan, though, is to shrink what MySQL does over• Redis time, let MySQL do what its good at and have HBase take over where MySQL is running up against limits handling ever-growing write rates, table sizes, etc." -• hBase/Hadoop Michael Stack, hbase project chair, StumbleUpon DBA http://www.theregister.co.uk/2011/01/19/• CouchDB hbase_on_the_rise/• And the 45 other solutions out there...
A lot of web scale tech comes from...• Brad Fitzpatrick• LiveJournal infrastructure• memcached (distributed caching, hits less DB), MogileFS (distributed ﬁle system), Perlbal (reverse proxy load balancer), Gearman (remotely run code, load balanced, in parallel)• next: camlistore (http://camlistore.org/)
“Without money the site cant function. Okay, let me tellyou the difference between Facebook and everyoneelse, we dont crash EVER! If those servers are down foreven a day, our entire reputation is irreversiblydestroyed! Users are ﬁckle, Friendster has proved that.Even a few people leaving would reverberate throughthe entire userbase. The users are interconnected, that isthe whole point. College kids are online because theirfriends are online, and if one domino goes, the otherdominos go, dont you get that?” -- Mark Zuckerberg(okay, not really, Jesse Eisenberg, in The Social Network)
Resources• High Performance Web Sites (Steve Sounders)• High Performance MySQL (Jeremy Zawodny, Baron Schwartz, Peter Zaitsev, et al)• Study HyperDB (Powers wordpress.com)• http://kb.askmonty.org/