← Tous les modèles

Le serveur fait autorité, tick rate fixe, ne jamais faire confiance au client.

Constitution — Multiplayer game server

Version 1.0.0 · Ratified 2026-11-10

These principles hold for every spec, plan and task of the server. A spec may add rules; it never weakens these.

  1. The server is authoritative. The server owns the game state and decides every outcome: movement, hits, scores and inventory. Clients send inputs, not results, and render what the server confirms.

  2. Fixed tick rate. The simulation advances at a fixed tick rate stated in the plan (for example 30 or 60 Hz), independent of load. A tick that runs over its budget is logged with its duration, and the p99 tick time stays under 50% of the tick interval.

  3. Never trust the client. Every input is checked for rate, range and plausibility (speed, cooldowns, line of sight) before it is applied. A packet that fails a check is dropped and counted against the session.

  4. Lag compensation is bounded. Hit checks rewind other players to the shooter's view time using a history buffer, up to a stated maximum (for example 200 ms). Beyond that the shot is judged at server time.

  5. Prediction and reconciliation. Clients predict their own movement and replay unacknowledged inputs when a server snapshot arrives. Snapshots carry the last processed input sequence number so corrections are exact.

  6. A versioned wire protocol. Messages are defined in one schema with a protocol version checked at connect. Snapshots are delta compressed against the last acknowledged state, and a client with a different version is refused with a clear reason.

  7. Simulation is deterministic and testable. The tick function is pure over state, inputs and a seeded random source, so a recorded match replays to the same result. Tests replay recorded inputs and compare the final state.

  8. Load tested before release. Every release runs a bot load test at the target players per instance with simulated packet loss and jitter. Tick time and bandwidth per player are measured, and a regression over 10% blocks the release.