Two users edit the same record at the same time. How do you stop one silently overwriting the other?

Use optimistic locking for most cases: give the row a version column or timestamp, include it in the update predicate, and if zero rows are updated the record changed underneath you, so you return a conflict. Use pessimistic locking, a select for update inside a short transaction, when conflicts are frequent or the operation cannot be retried safely. The one thing you must not do is read, modify and write with no check at all.

The failure is read modify write, where both requests load version one, both compute from stale data and the second write wins silently. My default fix is optimistic: every row has a version, the update says where id equals this and version equals what I read, and it bumps the version. If the update reports zero rows changed, someone else got there first and I return a 409 rather than pretending it worked. That costs nothing when conflicts are rare, which they usually are. I switch to pessimistic locking when contention is real or a retry is unacceptable, for example decrementing inventory, where I take a select for update on the row, do the check and the write in one short transaction, and keep no network calls inside that transaction. The part I care about most is what the user sees. A raw conflict error is useless, so for a document editor we returned the current version alongside the error so the client could show a diff instead of just losing the work.