Mistake 2 is the one I see teams learn the hard way: writing the new value into the cache after the database write feels like the tidy option, and then a concurrent request that read the old row repopulates the stale value after your delete. The advice to delete on write and keep a short TTL as a backstop is the honest version — it narrows the race instead of pretending it doesn't exist. The checklist item I would add to the pre-launch list is a test that logs in as two users, which you do cover under mistake 6; that two-user test has caught more cache bugs for me than any amount of TTL tuning. One question: for the stampede fix, do you prefer the single-rebuild lock with the others serving the old copy, or background refresh before expiry, when the data changes every few seconds rather than every few minutes?
Prism Infoways