on August 10, 2007 at 03:38 PM
So I'm not going to claim to be Kevin Closson - because I'm not. I'm also not going to wade into a shared-nothing vs. shared-storage architecture debate. And here's why: there is no right answer.
As with anything else, it comes down to what you want to do. Look at what Kevin says in his very long-windedly (yet nicely) titled:
Nearly Free or Not, GridSQL for EnterpriseDB is Simply Better Than Real Application Clusters. It is Shared-Nothing Architecture After All! « Kevin Closson’s Oracle Blog: Platform, Storage & Clustering Topics Related to Oracle Databases
Folks, today’s applications are built on large numbers of tables and complex joins. The reason shared-nothing is nothing like RAC is because instead of only shipping functions (or tasks) and lock messages to the clustered nodes, as is the case with RAC, shared-nothing requires the shipping of data
I'm going to do my best not to make this sound like a shot - because it's not supposed to be. But today's applications ON ORACLE, are built on large numbers of tables and complex joins. Because that's what you use if you have a setup that needs that sort of thing. Yeah - a shared nothing, single-contiguous image database in the 1000GB scale is going to have a whole host of issues. Likewise, I'd love to see a RAC setup try to beat MySQL Cluster in the HLR world. Good luck. There are still plenty of setups where the system is not built on large numbers of tables and complex joins. And in many of the case where they are overly massive and overly complex, the schema can be fixed.
Consider massive web property scale out where you can partition the data at the application layer. Say, an enormous social network, for instance. In a typical social network, you may have millions of users, but most of the interaction is clustered by user, so you can split a user's data across multiple machines. It really doesn't matter all that often to Alice what's in Bob's user data, except at specified points of interaction. (read - friends lists) In the places where the data does cross user boundaries, it's in such a small subdomain of data that it can be easily split out into a smaller and simpler vertical app that does one or two smaller tasks.
In this case, you're doing shared nothing, but you don't have to do a join across all of the data, because of the overall structure and relation of the data. The cost-per-unit-of-storage in trying to scale to numbers that big would be ridiculous on RAC or any shared-storage solution. And part of the reason there is that you just don't need the contiguous system. You can divide and subdivide. So the data shipping problem isn't nearly as much of a problem.
The approach I’ve taken to the shared-nothing versus shared-disk architecture topic is one of theory versus reality. I don’t care how many people say shared-nothing is the best for one workload or the other. The point is that by measured results it is not clearly the best for one (DSS) and is certainly not fit at all for the other (OLTP).
Interesting, because I was about to make the same point in the opposite direction. First of all - no argument about shared nothing in the DSS space. It just doesn't make a whole lot of sense, for exactly the reasons listed here. But in the OLTP space, I'd beg to differ, and I'd like to use the same theory-vs-reality moniker. The "reality" listed above is based on benchmark testing. Now I'm as big a fan as anyone of benchmarks. They're great in some contexts. But an OLTP benchmark is theory when compared to systems out in the field doing real work. Many OLTP systems can quite easily be sharded to handle an enormous amount of data and transaction rates. Having 1000's of parallel systems each working on their own small slice of the puzzle seems to scale in a much more linear fashion in practice than does purchasing massive RAC-like systems.
Then there is the ability to run shared-nothing on tons of smaller pieces of commodity hardware with an built in assumption that pieces of hardware are going to fail. If you want the "reality" backing that one up - see Google.
Again, I'm not trying to claim an this-vs-that victory. I'm just trying to deflect an attempt at one.
It's not that simple. It's not cut-and-dried. There are situations where shared nothing makes a lot of sense, and there are situations where shared strorage makes sense. Nothing is built with just a hammer - you usually also need a saw and a screwdriver ... and maybe even a concrete truck.