When inserting, the cursor interface collects keys/values via the set_key or set_value API calls.  The data goes into the WT_CURSOR.key or value buffers, which may require malloc (addressed in WT-12811).  But if the key_format or value_format is "u", or the cursor was opened in "raw" mode, we really don't need to do a copy - provided the caller does not overwrite their own buffer between the set_key/value and the use of the data via a insert/update/search.

Note - the above conditions, format="u" and buffer remains valid, should be satisfied by MongoDB.  So I think we can often/always avoid a copy to the intermediate buffer in MongoDB.  WT_ITEM already has the capability to point outside its own allocated memory, but would we be violating any assumptions.  So potentially this is a pretty small change.

We'll need to read/modify the doc carefully.  While we haven't made a promise (that I can see) that we're making a copy, in cases where the format has a non-"u" schema (e.g. "iSS"), the arguments to set_value are clearly "by value".  So to be safe maybe we need/want to put the new behavior on a cursor or connection flag?

(Component: cursors)