Going directly to the storage engine is cheaper then going through the query engine. It avoids the overhead of allocating/deallocating the QE stages and walking the associated extra code.

ReplicationConsistencyMarkersImpl::_upsertOplogTruncateAfterPointDocument is called as part of writing replication changes. Changing it to use the storage engine api will reduce the instruction count. While the POC made the change in ReplicationConsistencyMarkersImpl, a better approach may be to make it with upsertById in storage engine impl. The ReplicationConsistencyMarkersImpl does cache the _id record id. It also bypasses the opobservers to saving time applying the change

Note: I have not measured the perf of this change alone

POC: https://github.com/10gen/mongo/commit/2f09e0eb2cdb1acdc8d23ab397c2f26647204fd5