As part of SERVER-105921  we introduced acquisitions from the shard role API for the filemd5 command.
The command is a just a collection scan that attempts to compute an hash. Since the operation is long, an interruption might cause the scan to be temporarily interrupted and the hash to be computed on document scanned so far.
Since md5 computation can be slow, locks needs to be yielded and restored  in between the computation.
However, in case of stepdown:

• The yield might happen on the node as primary
• the restore might happen on the node as secondary

That causes the restore to change the read source which must be kLastApplied, but that can't be changed if a snapshot is still open (on wiredtiger)

Solution are either

• Abandon the snapshot before restoring (simple fix)

• Think hard if this could be generalized by the restore functionality so that we don't have to let the caller deal with this