On Mon, Jan 5, 2015 at 12:11 PM, Robert LeBlanc <robert@xxxxxxxxxxxxx> wrote: > If Ceph snapshots work like VM snapshots (and I don't have any reason to > believe otherwise), the snapshot will never grow larger than the size of the > base image. If the same blocks are rewritten, then they are just rewritten > in the snapshot and don't take any extra space. The snapshot functions > differently than a log (like databases). No matter how many times a block is > rewritten the read access should be roughly the same. Only when you increase > the number of snapshots on a single image is there the potential for > increased read latency. I believe there is a Ceph blueprint to implement a > snapshot bitmap which would make such look ups very inexpensive as it would > all be done efficiently in memory. > > There is some concern that large snapshot files (and thin or sparse > provisioning) increase fragmentation, but if you are running a large VM > environment, there is no such thing as sequential access on your storage > systems anyways. Exactly. If you take a snapshot, there's some bookkeeping but basically nothing changes. Then on a write, the original data (which is in the snapshot) is preserved either by copying to a new file or by preserving the blocks or whatever (this depends on what local storage system is in use) and the snapshot metadata is updated to point to it as appropriate, but the newly-written data is written to the same place it would normally go in the fs (on xfs we just overwrite the file as appropriate, in btrfs it's all COWed anyway, etc). -Greg > > On Fri, Dec 26, 2014 at 6:33 AM, Lindsay Mathieson > <lindsay.mathieson@xxxxxxxxx> wrote: >> >> On Tue, 16 Dec 2014 11:50:37 AM Robert LeBlanc wrote: >> > COW into the snapshot (like VMware, Ceph, etc): >> > When a write is committed, the changes are committed to a diff file and >> > the >> > base file is left untouched. This only has a single write penalty, if >> > you >> > want to discard the child, it is fast as you just delete the diff file. >> > The >> > negative side effects is that reads may have to query each diff file >> > before >> > being satisfied, and if you want to delete the snapshot, but keep the >> > changes (merge the snapshot into the base), then you have to copy all >> > the >> > diff blocks into the base image. >> >> >> Sorry to revive an old thread ... >> >> Does this mean with ceph snapshots if you leave the hanging around then >> the >> snapshot file will get larger and larger as writes are made? and reads >> will >> slow down? >> >> So not a good idea to leave snapshots of a VM undeleted for long? >> -- >> Lindsay >> _______________________________________________ >> ceph-users mailing list >> ceph-users@xxxxxxxxxxxxxx >> http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com >> > > > _______________________________________________ > ceph-users mailing list > ceph-users@xxxxxxxxxxxxxx > http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com > _______________________________________________ ceph-users mailing list ceph-users@xxxxxxxxxxxxxx http://lists.ceph.com/listinfo.cgi/ceph-users-ceph.com
 |