In WT there is the notion of a raw mode cursor. In this mode the user must provide a WT_ITEM with the key bytes to use. These cursors are not in use within MongoDB right now.

In profiling runs we identified that the WT_CURSOR->set_key method ends up performing the following:

• First compute the necessary buffer size for the key bytes
• Allocate a buffer of that size
• Perform a struct pack of the RecordId into it

However, we know the following information:

• How many bytes the RecordId can take in the normal case of a Long key
• The buffer used by set_key never leaves the functions they are called in and the cursor is immediately positioned

With this information we could practically remove allocation costs by using a stack allocated buffer. We can also perform the struct packing ourselves before passing the raw key bytes into WT. This would turn the key setting into a single pass algorithm.

I did a POC of the changes required within WT and benchmarked the new algorithm which is attached in this ticket. Using a hybrid approach where we keep the cursors as is today but calling the today unexposed __wt_cursor_set_raw_key method yielded a 100% improvement going from ~100ns down to ~50ns. This might not be much but it is a heavily exercised path.