| > Q1. I still don't get the use case for db storage on ephemeral storage. We're still not sure either, so we're investigating to see if it makes sense. One possible option will be to have the master on ephemeral disk with a hot backup on EBS so there is no data loss. Another option is use ephemeral for the master and all but one slave, so we got hot backups without a slowdown. Still need to look into it more. The one that we are doing ephemeral right now is Cassandra with continuous snapshots to EBS. Everything in there can be recalculated, and with an RF of 3, if we lose one node we can run a repair. > Q2. If EBS is the problem why are you migrating to S3 backed EBS boot vols? The problem with this is still the time in between snapshots even though it will be shortened. They are just easier to use. The root volume is rarely accessed after it is booted, so the EBS slowdowns aren't really a problem in that case. > Some Comments: It will only be a matter of time before S3 disks and hardware start dying like EBS...en masse I don't think so. It is a totally different product built by a totally different team with a different philosophy. S3 was build for durability above all else. In response to the rest of your comments, you are absolutely right, there are other options. We will certainly be investigating them. |