Fb Distributed Data Store Tao

intro

read-optimized social graph

memcache and look-aside

The distinction between look-aside and look-through caches is not whether data is fetched from the cache and memory in serial or in parallel. The distinction is whether the fetch to memory on a cache miss originates from the caller or the cache. If the fetch to memory originates from the caller on cache miss, then you’re using a look-aside cache. If the fetch to memory originates from the cache on cache miss, then you’re using a look-through cache.

1. @ FB, cache is controlled by client code

2. PHP abstraction direct r/w nodes and edges

inefficient edge lists

1. kv cache not good for edges, must fetch entire edge list to change a single edge list

2. basic list support not sufficient for concurrent incremental updates to cached lists

distributed control logic

1. control logic run on the clients which don't communication with each other

2. leases -> a general solution

3. fixed API, control logic moved into cache

read-after-write consistency