HYBRID DBMS OPTIMIZED FOR READ-INTENSIVE APPLICATIONS


Michael Stonebraker
Department of Electrical Engineering and Computer Science

The structure of today's most popular relational database management systems (DBMS) is not efficient for data warehouse applications. Today's database systems are "write-optimized"—they store information in a "row," meaning all fields in a given record are stored contiguously.

The DBMS "row store" has proved very successful for commercial transaction-processing applications, because a record can be inserted or deleted with a single disk operation, but it is not as appropriate for warehouse applications. In these applications, data from transaction systems is loaded periodically into a historical store, and analysts run ad-hoc queries to garner intelligence about the business. A DBMS that could store similar fields in different records together in a "column store" structure would be at least an order of magnitude faster than "row-store" database systems. Such a "column store" would produce a "read-optimized" DBMS.

Commercial vendors are loathe to build and maintain two radically different DBMS and have thereby focused on transaction processing and lived with lower efficiency in their burgeoning data warehouses. This project aims to design a hybrid system that delivers the best of both worlds: a write-optimized engine that performs updates efficiently coupled with a back-end, read-optimized engine that performs massive queries efficiently.