On 8/21/22 04:31, Chris Smart wrote: <snip>
Now I'm trying to understand the cause of delay in obtaining locks. I'd like to confirm first that when I see mds ops waiting to get locks I assume those are locks on objects in the data pool, right? And that this is data that a client somewhere has a lock on and is potentially doing something with? I assume that the MDS needs to ask the client to finish up their work and then release the cap so that the MDS can perform a metadata operation (without the file being changed again in the meantime), and it just has to wait for this to happen. (Maybe it's also stuck in an OSD ops queue). Is that right?
See this presentation: https://archive.fosdem.org/2020/schedule/event/sds_ceph_async_directory_ops/
Jeff Layton explains the client / MDS interaction. And also the optimizations that (by now) are available in newer MDS / kernel clients.
Gr. Stefan _______________________________________________ ceph-users mailing list -- ceph-users@xxxxxxx To unsubscribe send an email to ceph-users-leave@xxxxxxx
