Delayed reminders without a dedicated queue only stay correct when the wake source, the idempotency key, and the catch-up path agree. Prefer storing the due timestamp with a stable reminder id, wake from a periodic sweeper or datastore TTL that cannot double-fire the same id, and bound how far catch-up may replay after a process gap so a restart cannot spam every overdue user at once. Keep the send path separate from the schedule write so a transient provider blip does not lose the due row.
Ops prove-out after wiring the sweeper: insert a synthetic reminder due in the near future and confirm exactly one send, then kill the process before due and restart to confirm the same id still fires once. Under a burst of overdue rows, watch send rate and duplicate-id count for one deploy window — if duplicates climb while the sweeper interval stays flat, the claim lock is missing or too short. Keep a canary that still uses the prior schedule store beside the new path so you can tell a sweeper bug from a real provider outage.