The fix for DBZ-3331 implemented in these lines creates a table scan of the oplog because there are no (and can be no) indices. We have an oplog of ~50GB and we had to kill several queries (one per connector) that already ran for 4 hours, not a single one succeeded till then. 

One of our engineers suggested this:

the oplog is ordered, you could just query the first item and see if the timestamp is less than oplogStart