We have had various situations in which a scan did not return a lot of rows but was very slow, and the reason was because there were a lot of "invisible" keys with older MVCC values. Currently, there is no way for the SQL layer to obtain information about this (and show it with EXPLAIN ANALYZE).

I propose that KV returns (in ScanResponse) information about the number of underlying keys (and total bytes) that had to be processed to satisfy a ScanRequest. If collecting this information comes with performance overhead, we can add a flag in ScanRequest to only request it when necessary. This may apply to Get as well.

CC @asubiotto

Jira issue: CRDB-2835