> There are two types of "object", RBD-image-object and 8MiB-block-object. > When create a RBD image, a RBD-image-object is created and 12800 8MiB-block-objects > are allocated. That whole RBD-image-object is mapped to a single PG, which is mapped > to 3 OSDs (replica 3). That means, all user data on that RBD image is stored in those > 3 OSDs. Is my understanding correct? RBD image is not a object, is a bunch of objects as block device abstraction. Nope, each object of image may be placed to pseudo random placement. For example if you have 10000 osds and 10GiB image with 4MiB objects your image may be placed to 2560 different PGs on 100-1000-2560 OSDs... > I doubt it, because, for example, a Ceph cluster with bunch of 2TB drives, and user > won't be able to create RBD image bigger than 2TB. I don't believe that's true. > So, what am I missing here? User can create even 10TiB image in this case, 10000 of 2TiB drives gives you 5.8PiB of user space. All images are thin provisioned by default, so image size may be very big, but user may not fill it up to 100% or discard unused blocks k Sent from my iPhone > On 8 Aug 2021, at 00:00, Tony Liu <tonyliu0592@xxxxxxxxxxx> wrote: _______________________________________________ ceph-users mailing list -- ceph-users@xxxxxxx To unsubscribe send an email to ceph-users-leave@xxxxxxx
 |