PERFORMANCE IN A BOX
What most data warehouse appliances have in common, whether built on row- or column-store databases, is a massively parallel processing, shared-nothing architecture. MPP means that the query load is spread across many processors, or nodes, usually on commodity hardware running Linux. Shared nothing means that each node is independent, with its own memory and storage. The result is high performance without the expense of the high-powered, symmetric multiprocessor servers that typically run conventional data warehouses.
Appliances also are gaining ground because they're easier to deploy and maintain than conventional warehouses, which have to be tuned, optimized, and, lately, clustered to perform in large-scale deployments. Trading on this appeal, column-store database vendors, including ParAccel, Sybase, and Vertica, have introduced software-hardware bundles built on third-party hardware.
In the case of Teradata, which was the first to bundle hardware and software in an MPP, shared-nothing architecture (without calling it an appliance), the upstarts are competing primarily on price. Teradata responded last month with its own appliances--one for data marts and one for small warehouses--while also upgrading the performance and scalability of its core EDW product.