Describe the problem

Following on from this excellent work, we see a foreground SQL latency spike when a changefeed starts (initial scan).

The above PR reduces that spike by ~20%, which is great, but ideally it would not exist or be measured in milliseconds. The above indicates ~12 seconds of latency, which will hurt production users.

The question to start with is: “what’s different about changefeed startup vs the steady state?”

To Reproduce

Reproduction is in the above PR.

Expected behavior

A changefeed backfill should have an insignificant effect on foreground SQL latency.

cc @nvanbenschoten if any thoughts, given your previous investigation.

Jira issue: CRDB-18774

Epic CRDB-24463