If you follow the mailing list, you will see that a critical bug was
found in the s3 gateway and the recovery process. It basically boils down to
gateway chunking of big files (which is needed in s3) and a bug in the
recovery process of chunked files (note, there is no bug in the snapshotting
process). This is evident in the s3 gateway as it is chunked by default.
There is already a fix for this in master, and I will release 0.9.1 soon.
Since full gateway recovery only happens on a full cluster restart, you
might have not encountered it. Luckily, since the bug is only on the
recovery side, a way to solve this is to issue a full cluster shutdown
(always using the API), replace with latest master (or wait for 0.9.1), and
then start again the cluster.