Real-time multiplayer netcode-inspired game in Go — client-side prediction, server reconciliation, and entity interpolation that make milliseconds of lag disappear. The game is two circles in an arena; built to solve network latency problems in online games.
The netcode is the project. Every technique that makes online games feel instant — despite lag, jitter, and diverging state — is built from scratch and visible live in the on-screen HUD:
- ⚡ Client-side prediction — your move applies instantly; no round-trip wait.
- 🎯 Server reconciliation — the authoritative server corrects mispredictions invisibly (no rubber-banding, even at 500 ms of lag).
- 🌊 Entity interpolation — opponents glide smoothly between 20 Hz snapshots.
- 🎬 Lag compensation — hits are judged against what the attacker actually saw.
- 📦 Binary delta protocol — MessagePack snapshots, ~½ the bytes of full state.
- 🔒 Fixed-timestep authoritative server — deterministic 20 Hz simulation; a single goroutine owns all state, so there are no mutexes.
- 🤖 A behavior-cloning AI opponent — a tiny MLP trained on human play, connecting as an ordinary WebSocket client. Practice against a live bot lobby.
Load-tested with a custom harness to its O(n²) failure point (~150 concurrent
clients) and profiled with pprof — zero goroutine leaks.
Go · gorilla/websocket · MessagePack · vanilla JS + Canvas ·
Python / numpy (bot training) · Docker · Render
Use arrow keys or WASD to move your circle. Charge into your opponent to drain them.
- Live: gonet.onrender.com
- Local:
Then
go run ./cmd/server # http://localhost:8080/play(open two tabs to play yourself) or/play?mode=practice(a server-populated bot lobby).
The full design — the six-stage pipeline a keypress travels, the single-owner concurrency model, the wire protocol, lag compensation, the AI bot, benchmarks, and how to run/verify everything — lives in ARCHITECTURE.md.
cmd/server wires the hubs, HTTP routes, embedded assets
cmd/bot headless WebSocket bot (MLP or heuristic)
cmd/loadtest concurrent-client load harness
internal/hub tick loop, physics, collisions, lag comp, delta encoding
internal/bot the bot's decision logic
client/ the game UI (prediction · reconciliation · interpolation)
site/ the landing page
scripts/ train_bot.py — behavior-cloning trainer
Rooms (2-player sharding → O(1) per session → thousands of concurrent games) and self-play RL bots (to beat the behavior-cloning ceiling). Details in ARCHITECTURE.md.