Executing Web Application Queries on a Partitioned Database

Partitioning data over multiple storage servers is an attractive way to increase throughput for web-like workloads. However, there is often no one partitioning that yields good performance for all queries, and it can be challenging for the web developer to determine how best to execute queries over partitioned data. This paper presents DIXIE, a SQL query planner, optimizer, and executor for databases horizontally partitioned over multiple servers. DIXIE focuses on increasing interquery parallel speedup by involving as few servers as possible in each query. One way it does this is by supporting tables with multiple copies partitioned on different columns, in order to expand the set of queries that can be satisified from a single server. DIXIE automatically transforms SQL queries to execute over a partitioned database, using a cost model and plan generator that exploit multiple table copies. We evaluate DIXIE on a database and query stream taken from Wikipedia, partitioned across ten MySQL servers. By adding one copy of a 13 MB table and using DIXIE's query optimizer, we achieve a throughput improvement of 3.2X over a single optimized partitioning of each table and 8.5X over the same data on a single server. On specific queries DIXIE with table copies increases throughput linearly with the number of servers, while the best single-table-copy partitioning achieves little scaling. For a large class of joins, which traditional wisdom suggests requires tables partitioned on the join keys, DIXIE can find higher-performance plans using other partitionings.