Oracle has struggled to find a competitive foothold against data warehousing leaders Teradata, Netezza and Greenplum. With Exadata, Oracle is marketing that they have a solution, but the details tell another story. While they’ve cut EMC out of the picture, the new Oracle and HP solution is weighed down by the same overly-complex shared-disk architecture that Oracle RAC customers know well. Oracle knows the marketing story they want to tell, but they are no closer to building a proven shared-nothing architecture like Greenplum that can economically scale from terabytes to petabytes with massively parallel query performance.

Oracle has been getting beat badly in the high-end warehousing space. They've had a number of issues working against them:
- Oracle is an OLTP warehouse at heart. It has never been good at running large parallel analytical queries. Anyone with experience trying to implement a 50TB Oracle warehouse will tell you that it takes rocket-science tuning to get it to behave.
- Oracle's scale-out story is RAC, which requires a big and expensive shared-disk infrastructure (i.e. SAN) for coordination. Problem is that SANs only have ~4GBytes/s bandwidth out, so I/O scalability is inherently capped. RAC is also a bear to configure and has been known to have some stability problems.
- Oracle is competing against leading vendors such as Teradata, Netezza and Greenplum who are masters of the massively parallel shared-nothing architecture -- the only architecture that's known to work at these scales. Oracle has the wrong architecture, and they just can't compete technically.

So what has changed with the Exadata announcement? Once you cut through the marketing, this is really about swapping out EMC storage for HP commodity gear, taking money from EMC's pocket and putting it in Oracle's. Sure, there’s a little more I/O bandwidth, and very simple queries (e.g. basic table scans) may be faster. However this is just the low-hanging fruit, and the rest of RAC is unchanged, with significant parallel query planning and execution problems, and is sorely in need of re-architecting to allow multiple servers to work together effectively to plan and execute queries with complex joins and aggregations. Expect that extensive tuning of query plans will be required to avoid it melting down on anything mildly complex.

Oracle RAC has a fundamental limitation due to its shared-disk architecture — it requires that the Oracle Database servers are separated from the storage (i.e. SAN or Exadata Storage Servers), because each database server needs to be able to see all the data in order to function (figure 1). This means that the processing is happening on an entirely different box than the one accessing the data in disk. Oracle’s ‘intelligent storage’ story (i.e. predicate pushdown) is a baby step to reduce this deficiency by pushing around less data, but doesn’t address the core issue.

By contrast Greenplum’s shared-nothing architecture means that the database lives on the storage nodes (figure 2). This allows it to fully parallelize SQL, MapReduce and R (for statistical analysis), and do state-of-the-art analysis directly against the data — which makes Oracle’s ‘intelligent storage’ look awfully dumb by comparison. Until Oracle is able to run the entire database directly against their disks (i.e. a shared-nothing architecture) they are going to be living with fundamental processing limitations.

We welcome Oracle to the game, and we agree with their fundamental premise — i.e. that the traditional Oracle approach doesn’t cut it and customers need a highly-scalable solution, built on commodity hardware, that has been architected from the bottom-up for data warehousing. And while Oracle’s new solution just puts a new face on the old problems of Oracle RAC, Greenplum is delivering on this today with real solutions to address real customer needs -- scaling from terabytes to petabytes for some of the most demanding and sophisticated customers in the world.