https://www.usenix.org/system/files/conference/nsdi13/nsdi13-final170_update.pdf

• Summary: https://medium.com/@shagun/scaling-memcache-at-facebook-1ba77d71c082
• Problems
  • Low latency access to shared storage pool at low cost
  • K/V Interface, Scalable, eventually consistent.
  • Demand filled look-aside cache.
  • Users consume orders of magnitude more content than they create.
• Summary

  • Consistent hashing to distribute keys across servers.

    • Web servers have to routinely communicate with multiple cache servers to satisfy user req. This may cause incast congestion or a single server becoming the bottleneck.[Long tail]

  • Clients

    • serialization, compression, request-routing, error handling, req batching
    • Maintain a map of all available servers, updated via auxiliary config system.
    • Batch requests to individual partitions to fetch multiple items concurrently
    • UDP for GET requests
      • Bypass mcrouter, no TCP overhead; custom detection of dropped out of order packets via seq numbers.
    • TCP for set and delete via mcrouter

  • MCRouter

    • A proxy deployed to web servers
    • Request Coalescing via the open TCP connections to Memcached servers.

  • Favor Deletes for invalidations over updates. Deletes are Idempotent.

  • Cache invalidations via Daemons reading MySQL update log.