Design a Live Elimination Puzzle Arena With Spectators

You are designing PuzzleDrop, a mobile game. Players queue up and are placed in an arena of 50 to 150 people. Each round, every surviving player receives the same short puzzle (for example, a 4-option logic question) with a 12-second window. When the window closes:

• anyone who answered wrong or did not answer is knocked out;
• of those who answered correctly, the slowest 10% are also knocked out (at least one player, unless only one remains).

Rounds repeat until one player remains, or until a round knocks everyone out (then the fastest correct player of the previous round wins). Anyone can open a live arena as a spectator and see who is still alive, round number, and a top-20 leaderboard.

Functional requirements

1. Matchmaking that groups players of similar rating into arenas within ~15 seconds.
2. Synchronized puzzle release and a server-decided result for each round.
3. Live per-player state (alive/out, score, streak) for players in the arena.
4. Spectating any in-progress arena, plus a "featured arenas" list.
5. Final placements and rating updates written durably after each game.

Non-functional requirements

• Round results delivered to players within 500 ms of the window closing (p99).
• Fair timing: a player 200 ms from the server is not penalized against one 20 ms away; clients cannot fake speed.
• 2M concurrent players at peak (about 20k live arenas) and 10M concurrent spectators, heavily skewed toward a few streamer-hosted arenas.
• A crashed server must not end or corrupt games on other servers; a game whose server crashes mid-round should resume or be voided cleanly.
• Spectator spikes must never slow down gameplay.

Deliver

• High-level architecture and where state lives.
• How arenas are assigned to servers and how a round is settled.
• Your timing and anti-cheat approach.
• Failure handling for a game-server crash mid-round.
• How spectator traffic is served and isolated.
• Rough capacity numbers.