Notes from Social graph and the Database

From Jeff Rothschild, VP of technology, Facebook

1. The power of connectedness
2. Impact on the database
3. Our challenge 

• The power of connectedness
  • photo tagging
  • getting an email that someone tagged a photo in Facebook
  • inter tagging between friends and it continues as a network
  • outstanding growth because of the photo tagging
  • 26B photos in archive now
  • Most trafficked photos applications than anybody else
  • events, invites impact on social graph
  • Opening up the social graph to outside and create a platform and API, so other developers can make use of it (28K applications so far)
  • Hypergrowth because of the platform (1M users in four days)
  • 200 applications that has over million users
  • circle of connectivity can make the globe (with positive feedback, performance etc)
  • the social graph links everything
  • and that means lot of work for database
• Impact on the database
  • user -> connections -> objects
  • interact with people, people has lot of data
  • 100 friends, 100 friends will have 100’s of objects/friend…10’s of thousands of possible objects
  • Horizontal partitioning supports parallel queries
  • 50K+ requests per second now
  • memcached can match speeds with the application
  • multi-threading with multi-core to take advantage with 1G cards
  • New C client which enhances run-time performance for PHP
  • New binary protocol for efficiency both ends (app and memcached)
  • Introduced memcached proxy to purge dirty records
  • MySQL replication to replicate data between data centers
  • Proxy is not a solution here due to replication
  • Changed Insert statements with MEMCACHE_DIRTY key1, keys …
  • Put more RAM in cache (95% hit rate is in cache)
  • 20M requests and only 500K requests goes to MySQL server
• Optimizations
  • memory + servers (reduce to save $$$)
  • Tuning (moderate query scope, aggregate related data)
  • Partitioning (selective caching, archiving stale data)
• Our Challenge
  • MySQL and memcached is a powerful combination
  • MySQL at the speed of memcached
  • memcache performance for simple tables
  • flash memory storage
  • dramatic improvement in I/O
  • supports higher miss rate
  • allows higher data/ram ratio
  • persistent “Cache”
    • flash for memcache
    • persistent distributed hash table could be a good fit
  • structured storage
    • not all data benefits
    • integrated cache and persistent replication
    • unmanaged distribution
  • A strong community is our challenge