In general, when doing multi-shard operations which depend on versions, if a single or only two shards have incorrect versions the command / query needs to be retried on all shards. This is not usually necessary, and can cause unnecessary load if the command is costly (like mapreduce).

Fix : figure out which shards we need to refresh, just retry commands on those servers. This wouldn't totally eliminate the problem (if detected a version change on one shard but not the other after a migrate, for example), but would make it much less likely.