replace operation causes background fetch

A recent POC showed that under DGM, replace operations on ejected items caused them to be backfilled...significantly lowering the performance achievable.

Given what I know, this would not be expected as replace only needs to know whether the item exists or not and the hashtable should have sufficient information to do that.

A brief analysis of the code showed that replace always performed a background fetch (whereas add did not, a similar operation).

Seems like this could be a pretty straightforward fix if possible.