This becomes very relevant for $out to a sharded collection as well as sharded $lookup. Currently the cluster aggregate command will attach a single version based on the state of the execution namespace, however this may be incorrect if one of the shards wishes grab a collection lock on one of the other involved namespaces.

The title describes one possible approach, though I'm not sure how the implementation would look since it may involve modifying the operation context. Another option would be to have the shards temporarily modify the version on the op context when accessing a foreign collection.

Edit: Looks like such a map exists already, so the problem is that the involved namespaces do not have entries and thus default to unsharded.