The lag-mitigation TTL logic in set() could be made smarter by:
- Checking $walltime in a better way to distinguish high/low regeneration times. Right now, I see "0.005120992660522461" being reported as "high" for 4vN48nkBmhjHfZRTRKyF (could just be someone manually debugging on mwdebug though)
- Ussing something higher than LOW_TTL (30 seconds) for the "slow callback case without lockTSE" case
- Possibly scheduling async regenerations and making the asyncHandler flush replica snapshots (also nice for avoiding similar problems with preemptive refresh updates since they run at the end of the request)
- Possibly making purge tombstones last longer (assuming the current automatic lockTSE and interim key logic is sufficient) so that MAX_READ_LAG could be higher
See T282985: Show cached information on NewcomerTasksInfo special page.