Random Storage Comments, Answered

In my last posting, a lot of comments covered wide and varied ground. First, its important to note that even with CDP underlying ZFS pools, ZFS itself provides for its own integrity of state.  If CDP didn't complete a transaction, a re-sync will generally resolve it, but the actual hosted ZFS filesystem need not fear and its transactions won't be finished until the write is checksumed. I agree that there are failure modes here, but that leads to a good quote in one of the comments:

"To that end, it seems to be that whenever a choice can be had between doing something simple to accomplish a goal and chaining a bunch of parts together to accomplish the same goal with more sophistication, its likely the simpler solution will be more sustainable over time."

I concur. Nexenta marries two pieces of functionality to get auto-cdp, and they rely on the two components in whole to maintain overall simplicity of implementation. The real value that they have provided is in making the front end dead simple. If the management isn't simplified, any level of underlying functionality will be lost in the long run.

I want to focus more on the simplicity of the performant NAS solutions. Mentioning pNFS, lustre, and the like, we know that the client becomes a bit less transparent, and definitely the backend store of data becomes somewhat opaque as data is no longer consistent per one server, but is spread out across the whole back end. Even though you need newer clients with specific functionality in both cases, it can again be more simple than the alternative, which generally involves an NFS v3 client using automounts, LDAP-based auto mount maps, and heavy handed data management on the backend to scale out in similar ways over multiple NAS heads. The tact of taking a single high end head with best of breed backend hardware, such as IB interconnects to SAS disk arrays and 10GB ethernet out the front might seem to work, but we have already seen pathological conditions where a single heavy client writing millions of small files can make that enhanced hardware meaningless for performance.


There is no fast answer to solving both scale out with regards to capacity and performance without a little give on each aspect of the design. What makes it all reasonable to consider is if the entire solution is made greatly more simple to manage than the alternatives at either end of the design spectrum. In the end, simplicity of manageability will trump other considerations. As long as simplicity is strictly maintained in the product, the underlying complexity will seem well worth the effort. We just need to trust that someone gets the fine details. In the end, we don't mind that we can't muck much with a highly efficient but high performance car. As long as we feel mastery over its operation and trust in the quality of the build and service by the manufacturer, we all are willing to make the investment.