Design the backend that ingests authoritative score events from live matches and fans the current state out to millions of concurrent viewers in near-real time (think the live-scores view in a popular sports app during a marquee final).

An upstream match feed is the single source of truth: it emits ordered, sequenced score updates per match (goals, wickets, point changes, clock ticks, status flips). Your system must accept those updates, apply them idempotently, and push the resulting state to everyone currently watching that match — while also handing any viewer who joins mid-match an immediate, correct snapshot so they never see a blank or stale board.

Operating context. A few hundred matches run concurrently on a busy evening; a single blockbuster match can hold 5–10 million simultaneous viewers, and viewership is extremely skewed (one match may hold 80% of all connections). Authoritative updates arrive at 1–50 events/sec per match, occasionally bursting to a few hundred during a chaotic passage of play. Viewers connect from everywhere; new viewers join continuously, so late-joiner snapshots dominate the request mix. The update-to-glass latency budget is p99 < 2 s from feed ingest to a watching client, and a late joiner must get a snapshot in p99 < 400 ms.

Out of scope. The upstream match-feed provider and its reliability, video/streaming, user accounts and auth, chat/reactions, betting or odds, and historical match archives. Assume the feed hands you clean, monotonically sequenced events per match.

What to produce. A high-level architecture covering: the ingest path for authoritative events and how updates are applied idempotently and in order; the fan-out path to millions of concurrent subscribers and how it scales with the skew; the edge-caching strategy for the last-known state so most late joiners never hit the origin; the snapshot flow for a mid-match joiner and how snapshot + live stream are reconciled without a gap or duplicate; and the data model for current match state versus the append-only event log. Sketch the major components and the request/event flow between them; we will probe specifics at checkpoints.