On Thu, Mar 14, 2013 at 11:27 AM, Kirk Jantzer <kirk.jant...@gmail.com> wrote: > For my testing, because of available resources, I chose local storage. > However, shared storage (like GlusterFS or the like) is appealing. What are > you using? Why did you chose what you're using? How many instances are you > running? etc. > Thanks!!
So I think it depends on workload and use case. If your applications are highly fault tolerant and largely stateless, you simply may not care about HA. Besides you'll almost certainly get better performance with local storage. If you are running a test/dev environment, you can probably tolerate instance failure, so why use shared storage. If people are going to come scream at you and threaten to take money away if something fails, perhaps you want something a bit more robust. The best of both worlds (with plenty of compromises too) is distributed, replicated shared storage like Ceph RBD. (GlusterFS 3.4, with all of the work that IBM has done around KVM is promising, but yet to be released. Early versions were robust, but had problems at any scale providing decent IO) Sheepdog is also promising and I keep hearing there are patches incoming for Sheepdog support. Of course these are all KVM-only for the moment. There are also plenty of people who do both local and shared storage. With higher internal costs for deploying to shared storage with the assumption that folks would do it for things that need a higher level of resiliency or less tolerance for failure. For shared storage, I've seen everything from NFS running on Linux to things like Isilon, Netapp, and EMC - again the choice depending on the tolerance for failure. --David
