Database partitioning strategies for social network data

Abstract

In this thesis, I designed, prototyped and benchmarked two different data partitioning strategies for social network type workloads. The first strategy takes advantage of the heavy-tailed degree distributions of social networks to optimize the latency of vertex neighborhood queries. The second strategy takes advantage of the high temporal locality of workloads to improve latencies for vertex neighborhood intersection queries. Both techniques aim to shorten the tail of the latency distribution, while avoiding decreased write performance or reduced system throughput when compared to the default hash partitioning approach. The strategies presented were evaluated using synthetic workloads of my own design as well as real workloads provided by Twitter, and show promising improvements in latency at some cost in system complexity.