Martin's blog^2.0

Excelent video about youtube DBA explaining performance tuning and scalability stuff. The slides are missing so if anyone has them post a comment.

In this video you can see Paul Tuckfield, talk about how they scaled Youtube, some interesting points:

Things that made youtube scale:

• Python
• Caching (before getting to the DB)
• DB tuning (that's what the video is about

From the DB tuning my highlights are these:

• Views count goes to a MyIsam DB
• Favourites, playlists, etc goes to a DB that thens replicates to other 5, as to divide the read queries
• Slow replication is a very big problem and is difficult to keep replication fast.
• InnoDB doesn't recover easily and is not nice to switch to the replica because you are not sure if that's OK.
• You can test things in one replica and if it works pass that change to the rest of them.
• A good trick is to have a replica with some lag as a way to have a easy restorable backup.
• When migrating to 5.0 is better to dump and load instead of updating in place.
• To upgrade hardware means much more memory and many disk drives (even if they are small) instead of fast CPUs and big hard drives
• If it is on cache you are ok, if your database fits in cache, go to have a coffee
• Write Cache is OK, Read Cache is BAD in levels below the database(Linux or RAID).
• For parallel reads is better to have big chunks