When trying to delete a record that doesn't exist, WiredTigerRecordStore::wtDeleteRecord currently logs an error, prints the stack trace, and moves on (in testing, it crashes the server). This can trigger false positives in replication when the secondary tries to apply the delete oplog entry and couldn't find the document. 

It will be better to return an error to indicate to the caller that a bad request was made and potentially tassert there.