Meta (Facebook): TAO Distributed Graph Data Store

Facebook models its entire universe as a giant interconnected graph: users, pages, photos, comments, groups, check-ins, and friendships.

In Facebook's early days, this data was stored in sharded MySQL databases with a Memcached lookaside caching layer. While Memcached provided fast key-value lookups, it introduced severe architectural challenges at social scale:

• Lack of Graph Primitives: Memcached treats values as opaque byte blobs. Appending a new "Like" to a photo required fetching the entire serialized list from Memcached, modifying it in PHP, and writing it back, creating severe race conditions and write amplification.
• Thundering Herd & Cache Stampedes: When a viral post's cache expired, thousands of simultaneous PHP web servers queried MySQL directly, overwhelming database connection pools.
• Complex Invalidation Logic: Application code was responsible for maintaining cache consistency, leading to frequent bugs where stale cached data diverged from MySQL.

In 2013, Meta published the landmark paper: "TAO: Facebook’s Distributed Data Store for the Social Graph", replacing Memcached lookaside with a geographically distributed, graph-aware, read-through/write-through cache tier.