How catalog writes are timestamped is often different for writes that occur while a node is primary vs. secondary. Most of our rollback tests do not make this distinction. We should make some deterministic tests that exercise this behavior and also ensure that the rollback fuzzer is exercising it. We should make the writes occur on both sides of the common point. Getting operations applied as a secondary that have to be rolled back will likely require a 5 node replica set with a double rollback.