On Wed, Nov 12, 2014 at 09:54:55PM +0800, Anthony Alba wrote: > Does the filesystem blobstore provider work on shared folders, with > concurrent clients? > > Does the provider handle locking and contention in naming?
Most blobstores do not provide follow locking semantics: instead they use a last-writer wins model. The filesystem blobstore lacks atomic updates[1]; it writes new and overwritten objects in-place, exposing partial updates. Instead it could write objects elsewhere then do an atomic rename. Would you like to investigate this further and perhaps provide a fix? Note that this will likely have some caveats on Windows due to how it handles renames/deletes of open file handles. [1] test_s3.py:_test_atomic_write, https://github.com/ceph/s3-tests/blob/master/s3tests/functional/test_s3.py#L4623 -- Andrew Gaul http://gaul.org/