Persisting the vector clock before setting the filtering metadata discovered after refreshing ensures causal consistency in the following scenario:

• Node A is primary and refreshes, discovering shard version X
• Step-down
• Node B becomes primary and refreshes, discovering shard version X - 1 because reading from the (stale) nearest CSRS node

In order to not block the shard for too long of a time, the checkpoint must be done only with local write concern.