For large write operations that generate many oplog entires, say, a large multi-delete, investigate strategies to prevent secondaries from falling too far behind.

Flow control doesn't sufficiently account for this because it throttles operations globally, and only after lag has reached 5 seconds.

The idea here is to force operations that have written large amounts of data to periodically wait for their writes to majority-replicate before writing any more. In this way, we don't penalize smaller write operations, and only make problematic operations back off.