|
|
|
|
|
|
by cmurf
3695 days ago
|
|
|
For scaleable storage, get rid of conventional RAID for data. I'd like to see n-way (definable) copies of metadata, and single copies of data. On top is a cluster file system like GlusterFS. When a device dies, the file system merely rebuilds metadata, and then informs GlusterFS of the missing data due to the failed drive(s). And then the file system deletes the reference to all missing/damaged files from that missing drive. No degraded state ever happens. This way Gluster knows not to even make requests from that brick. If the brick were raid56 and the cluster fs isn't aware, requests happen with degraded read/writes which suck performance wise. Plus Gluster (or even Ceph for that matter) might use some logic and say, well that data doesn't even need to be replicated on that brick, it's better for the network/use loading if the new copy is on this other brick over here. |
|
|