|
Horizontica -- A New Approach to OLTP Data Bases
Daniel Abadi, Stavros Harizopoulos, Sam Madden & Michael Stonebraker
Relational Data Bases have been the dominant mechanism for storage and processing of business data
processing (OLTP) data for more than two decades. The traditional architecture, embodied in all the major
commercial systems, is based on the principle of "one size fits all". In other words a single DBMS
code line is used for all DBMS problems. The prevalent architecture uses a row-oriented representation
of tuples, a disk-based storage system, B-trees as the indexing mechanism, small disk blocks with a
main memory cache, and a row-oriented general purpose executor and optimizer.
We have previously demonstrated that the one-size-fits-all architecture can be beaten dramatically (one to
two orders of magnitude) by a column-oriented solution in the data warehouse market and by an array
DBMS in the scientific and intelligence market. In this project, we are aiming to prove that the same dramatic
advantage can be realized in the business data processing (OLTP) market. The system we are building,
Horizontica, is in the early design phase. The ideas we are exploiting include optimization for main
memory data, built in high availability, highly simplified transaction logic, new indexing tactics,
and the ability to exploit the capabilities of an associated application server. Also included are the specification
of a small subset of SQL, appropriate for the OLTP market, and the abandonment of ODBC/JDBC as the user
interface, in favor of one oriented toward web services.
|
|
