Skip to content

BACKLOG.md

Latest commit

 

History

History
27 lines (19 loc) · 4.43 KB

BACKLOG.md

File metadata and controls

27 lines (19 loc) · 4.43 KB
Raw

Backlog

Forward-looking work and known limitations. Items here are intentional gaps, not surprises — read this before "fixing" something that is already a deliberate trade-off.

Known limitations

  • Reconnect slot is claimed by code, not identity. While a match is paused for a reconnect (server_do), the next Join for that match re-attaches to the held-open slot. In a private two-player room that's the dropped player returning; in principle a third party who has the link could take the slot during the grace window. Acceptable for casual play — tighten with a per-player reconnect token if it ever matters.
  • Match codes are guessable in principle. Codes are 5 chars over a 36-char alphabet (~60M combinations) generated with thread_rng. Fine for short-lived two-capacity rooms; add rate limiting or longer codes if room-crashing is ever observed.

Nice to have

  • Reconnect resume polish. On resume, the reconnecting client briefly shows a reset board (0:0) during the ready-countdown until the first state snapshot arrives. Could buffer the last snapshot to avoid the blip.
  • Dependency freshness. A few crates trail latest (rand 0.8, getrandom 0.2, glam 0.27). None are security-relevant; bump opportunistically.
  • Spectators / >2 players. The Durable Object is hard-coded to two players. A spectator (read-only) connection would be a natural extension.
  • Region-aware DO placement. Each match runs in whatever location Cloudflare first instantiates the DO (near whoever created it), so a player far from that region sees light-speed latency. Pass a locationHint/jurisdiction so both players land near a shared region.

Code health

Behaviour-neutral cleanups deferred to keep changes small and safe to verify.

  • Collapse the CSS duplication. lobby_worker/style.css declares many layout selectors twice — a desktop value, then an unguarded "mobile-first" block that always wins — so the desktop declarations are effectively dead. Rework into one mobile-first base plus a single @media (min-width: …) block; verify on mobile and desktop widths since it touches the live layout.
  • Tune the wasm release profile. Add [profile.release] (e.g. opt-level = "z", lto = true) to shrink and speed up the deployed wasm; confirm the shared simulation stays bit-identical so client/server determinism holds.
  • Cover the time-based server paths. Give the test MockEnv an advanceable clock so the reconnect-grace expiry and idle-timeout triggers in the alarm loop are tested directly (today only their consequences are), and assert the broadcast bytes for the match-lifecycle messages.
  • Name LocalGame::step's return. Replace its six-element tuple (client_wasm/src/simulation.rs) with a small struct so call sites read by field, not position.
  • Bump the Workers compatibility_date (pinned at 2024-01-01) periodically, as a deliberate, smoke-tested change rather than a drive-by.
  • Dedupe paddle-input sends. The client polls sendInput() on a ~30 Hz timer (lobby_worker/script.js) and get_input_bytes() (client_wasm/src/lib.rs) always re-serialises the current paddle Y with an incrementing seq, so a stationary paddle still streams ~30 C2S::Input messages/sec — each a no-op, since the server treats input as an absolute, idempotent Y and ignores seq (server_do/src/lib.rs). Track the last-sent Y and return an empty buffer when it is unchanged; the JS send path already drops empty buffers (if (bytes.length > 0)), and the 30 Hz timer stays as the send-rate cap. Gameplay is unaffected — it just trims redundant Durable-Object message wakeups. Safe against the 120 s idle reaper because the 2 s Ping independently refreshes last_activity.
  • Decouple the server alarm rate from the sim rate (hosting-cost lever). The alarm reschedules every 16 ms (~60 Hz) during Playing (server_do/src/lib.rs), and each alarm invocation is a billed Durable-Object request — the dominant per-match cost. The alarm already runs an accumulator + catch-up loop that can step the sim several times per wake, so scheduling it at the 20 Hz broadcast cadence (~50 ms) and letting that loop run ~3 FIXED_DT steps keeps physics bit-identical while cutting alarm requests ~3×. Trade-off: input reaches the authoritative sim up to ~50 ms later — largely hidden because the client owns its own paddle — so verify game feel. This is the highest-leverage knob on the Durable-Object bill (far more than the input-dedup item above).