The query system does repeated work when reading from a collection and updating a document: it creates one cursor to scan (or fetch) the document, and in the UpdateStage, creates a second cursor to update the document. This means two cursors have to be created, and two seeks have to be done.

We should consider making changes in the db/exec/ engine to avoid this. Currently stages do not share resources in this fashion, and are oblivious to what other stages do, so implementing this in a non-hacky way that preserves the independence of the stages would require some thinking.

Before considering this, we should do SERVER-105226 and the linked WT ticket to verify that the improvement is actually substantial.

We should also investigate a similar optimization for deletes.