[adding ceph-devel] On Thu, 1 Dec 2016, LIU, Fei wrote: > Hi Sage, > > We have some workload that have requirements of latency and IOPS but > don’t care that much of data persistency. We do care a lot of TCO. We are > looking for a possible solution to deploy Ceph with single replica. We have > tried and put one OSD down intensively , the whole cluster was ok as long as > you don’t nee d to operate any data related to the down OSD. We need to be > carefully setup crushmap within a pool. However, what data is going to be > losed in the down OSD is very hard to track, any suggestions and adviced > will be appreciated. It of course depends on your application, but from teh ceph perspective, this is doable. The key is to streamline the data-loss procedure. If an OSD fails, you'll lose a random set of PGs, which will have a random subset of the overall set of objects in the cluster. You can tell the system to give up and recreate a PG as empty (acknowleding that data was *permanently* lost and telling ceph to carry on without it) with ceph pg force_create_pg <pgid> You'll want to have some automation to pick out which PGs are lost and recreate them. Of course, your application has to tolerate that... sage
