MySQL Conference – YouTube

MySQL Conference 2007 Day 4 rolled quickly into the second keynote Scaling MySQL at YouTube by Paul Tuckfield.

The introduction by Paul Tuckfield was; “What do I know about anything, I was just the DBA at PayPal, now I’m just the DBA at youTube. There are only 3 DBA’s at YouTube.”

This talk had a number of great performance points, with various caching situations. Very interesting.

Scaling MySQL at YouTube

Top Reasons for YouTube Scalability

The technology stack:

  • Python
  • Memcache
  • MySQL Replication

Caching outside the database is huge.

It a display of numbers of hits per day it was said “I can neither confirm or deny the interpretation will work here (using an Alexa graph)”. This is not the first time I’ve heard this standard “Google” response. They must get lessons by lawyers in what you can say.

Standardizing on DB boxes (but they crash almost daily)

  • 4x2ghz opteron core
  • 16G RAM
  • 12x10k scsi
  • LSI hardware raid 10
  • Replication played a big part in fixing
  • Get a reliable H/W supplier

Replication Lessons

  • You don’t worry about it when a replicas fail.
  • One thing that sucks, Innodb doesn’t recover very fast. It does that durability think, but it takes hours to finish recovering (was it going to finish)
  • How many backups can you restore. When you switch you a replica, are you sure it’s right?
  • Did you test recovery, did you test your backups.
  • replication was key to trying different H/W permutations to identify incompatible H/W (combinations of controllers/disks)
  • we got good at re-parenting/promoting replicas, really fast
  • we built up ways to clone databases as fast as possible
  • Excellent way to test tuning changes or fixes (powerful place to test things)
  • Keep “intentional lag”/Stemcell replicas – Stop SQL thread, keeps a server a few hours or a day behind. Say if you drop a table you have a online backup.
  • When upgrading, always mysqldump then reload, rather then upgrade database.
  • Don’t care about CPU’s. I want as much memory as possible, I want as many spindles as possible.
  • For YouTube 2-3 second lag is acceptable.

If you db fits in ram, great otherwise

  • Cache is king
  • Writes should be cached by raid controller (buffered really) not the OS
  • Only the db should cache reads (not raid, not Linux buffer cache)

Only DB should cache reads

  • Hit in db cache means lower caches went unused.
  • Miss in db cache can only miss in other caches since they’re smaller.
  • Caching reads is worse then useless. It’s serialized writes.
  • Avoiding serialization in reads reaps compounds benefits under high concurrency

An important lesson learned. Do no cache reads in F/S and Raid Controller.

Caching Lessons
Overcoming Mystery Serialization

  • Use O_DIRECT
  • vm.swappiness=1-5
  • if you’re >80% buys — your not doing I/O concurrently look at other figures e.g. 80% busy 8 I/O’s, next configuration 80%, only 4 I/O’s
  • Mirror in H/W strip in S/W

Scale Out

  • Writes are parallel to master, but serialized to replicas. We need true horizontal partitioning.
  • We want true independent masters
  • EMD – Even More Databases — Extreme Makeover Database
  • Slave transactions must serialize to preserve commit order (this is why replication is always way slower)
  • The oracle caching algorithm (that’s a small o) — predicting the future
  • Replication lags: one IO bound thread. You do know the future, commands are coming up serially.
  • Write a script to do reads, before updates coming up (because they are cache hits).
  • The diamond. For golive, play shards binlogs back to original master for fallback.