Design the reminder engine for a to-do and task app: users attach reminders to tasks ("nudge me at 9am", "an hour before it's due", "every weekday morning"), and the system fires each one on time and delivers it to their chosen channels. It is the piece that makes a task actually resurface at the right moment.

Operating context. Reminders vastly outnumber fires — most sit dormant for days or weeks. Fire times cluster hard on round numbers (the top of the hour, 9:00 local). A fired reminder can be snoozed to a later time, cancelled when its task is completed, or, if recurring, must schedule its next occurrence. Delivery goes to whichever channels the user picked. A late reminder is an annoyance; a lost or duplicated reminder erodes trust.

Out of scope. The task/to-do data model and app UI, the actual push/email/SMS transport internals (assume channel adapters exist), user authentication, and per-message content rendering. Assume a delivery service accepts a (user, channel, payload) and best-effort delivers it.

What to produce. A high-level architecture covering: how pending reminders are stored and efficiently queried by fire time without scanning everything; the dispatch path and how you tame the top-of-the-hour thundering herd; how you get effectively-once delivery from an at-least-once pipeline; how recurrence, snooze, and cancel interact; and the failure story when a dispatcher dies mid-fire. Sketch the components and the life of one reminder from set to fired; we will probe specifics in checkpoints.