diff --git a/big-feature-phase/notes/constraints.md b/big-feature-phase/notes/constraints.md new file mode 100644 index 0000000..fefcef6 --- /dev/null +++ b/big-feature-phase/notes/constraints.md @@ -0,0 +1,58 @@ +# Constraints — In-Browser Swactor Runtime + +## Threading Model + +- **wasm-threads is mandatory** — the runtime uses SharedArrayBuffer + WebAssembly atomics for multi-worker parallelism. There is no single-threaded degraded mode for MVP. +- Browsers must serve pages with COOP/COEP headers: + - `Cross-Origin-Opener-Policy: same-origin` + - `Cross-Origin-Embedder-Policy: require-corp` +- Build requires nightly Rust + `-Z build-std=std,panic_abort` + target features `+atomics,+bulk-memory,+mutable-globals`. + +## Architecture Rules + +- **Platform abstractions live in core swactor** (`src/`), gated by `#[cfg(target_arch = "wasm32")]`. They do not belong in the wasm crate. +- **Do not add new modules** to `src/` — modify existing files only (TASK.md style rule). +- **Do not restructure** existing module boundaries. The abstraction is a thin layer (type aliases, cfg-gated imports), not a trait-based HAL. +- The browser crate (`crates/wasm-browser/` or evolved `crates/wasm/`) is a **thin wasm-bindgen shell**. All scheduling, routing, and actor logic stays in core Rust. + +## Actor Model + +- **Rust-only actors** — actors are written in Rust and compiled to wasm. JavaScript does not define actor behavior. +- JS interacts through the wasm-bindgen API: create runtime, spawn actors (by registered type), send messages, receive results. +- Actor types are registered at compile time via Rust generics, not dynamically from JS. + +## Performance Priorities + +- Maximize throughput: auto-scheduling via `setTimeout(0)` tight loop, not `requestAnimationFrame` (which caps at display refresh rate). +- Web Worker count defaults to `navigator.hardwareConcurrency` for full core utilization. +- Zero-copy where possible: SharedArrayBuffer eliminates serialization between workers. +- Minimize JS↔Wasm boundary crossings — batch operations where feasible. + +## Feature Scope + +- All core features that compile for wasm32: spawn, send, receive, tick, actor lifecycle, watching, extensions. +- swactor-std features (naming, groups, monitoring) should work if they compile. +- Transport: WebSocket adapter for distributed clusters. STUN/TURN (WebRTC) deferred to later. +- Features that require OS primitives not available in wasm (filesystem, raw TCP) are excluded. + +## Testing + +- Tests must pass on both native (`cargo test`) and wasm targets. +- Wasm tests use `wasm-pack test --headless --chrome` or Node.js with `--experimental-wasm-threads`. +- No test-only code paths that diverge native vs wasm behavior — if it works differently, it's a bug. +- Prefer scenario tests over structural tests (per project testing rules). + +## Dependencies + +- `web-time` — drop-in replacement for `std::time::Instant` on wasm32 +- `wasm-bindgen` + `js-sys` + `web-sys` — browser API bindings (in the wasm crate only, not core) +- `gloo-timers` — optional, for ergonomic setTimeout/setInterval +- No new dependencies in core swactor beyond `web-time` (which is no-op on native) + +## What We Don't Do + +- No async/await runtime (tokio, async-std) — swactor is synchronous tick-based +- No Emscripten — target is `wasm32-unknown-unknown` only +- No WASI — browser environment, not server-side wasm +- No JS actor definitions — Rust only +- No polyfills for missing atomics — if SharedArrayBuffer isn't available, the runtime doesn't start diff --git a/big-feature-phase/notes/feature-stages/01-platform-abstraction.md b/big-feature-phase/notes/feature-stages/01-platform-abstraction.md new file mode 100644 index 0000000..9fa5954 --- /dev/null +++ b/big-feature-phase/notes/feature-stages/01-platform-abstraction.md @@ -0,0 +1,110 @@ +# Stage 1 — Platform Abstraction Layer + +**Priority**: P0 +**Depends on**: Nothing +**Enables**: All subsequent stages + +## Goal + +Make core swactor compile for `wasm32-unknown-unknown` with `+atomics,+bulk-memory,+mutable-globals` target features. No behavioral changes on native targets. No new modules — only modify existing files with `cfg` gates. + +## What Changes + +### 1. Instant → web_time::Instant + +**Files**: `src/runtime.rs`, `src/worker.rs` + +Add `web-time` to `[dependencies]` (it's a no-op on non-wasm targets). Replace: +```rust +use std::time::Instant; +``` +with: +```rust +use web_time::Instant; +``` + +`web-time` is a drop-in replacement. The `Instant` type has identical API on native (delegates to `std::time::Instant`) and on wasm32 (uses `performance.now()`). + +**Scope**: 2 `use` statements, 0 logic changes. + +### 2. Thread Parking → ParkHandle + +**Files**: `src/runtime.rs`, `src/delivery.rs`, `src/worker.rs` + +Currently uses `OnceLock` + `thread::park_timeout` + `Thread::unpark`. On wasm32, there's no `Thread` type accessible from Rust (workers are JS objects). But wasm-threads supports `Atomics.wait`/`Atomics.notify` through Rust's `std::sync::atomic` and futex primitives. + +Approach: Define a `ParkHandle` abstraction in `src/runtime.rs`: + +**Native**: +```rust +#[cfg(not(target_arch = "wasm32"))] +mod parking { + pub type ParkHandle = OnceLock; + pub fn register(handle: &ParkHandle) { handle.set(thread::current()).ok(); } + pub fn unpark(handle: &ParkHandle) { if let Some(t) = handle.get() { t.unpark(); } } + pub fn park_timeout_us(micros: u64) { thread::park_timeout(Duration::from_micros(micros)); } + pub fn yield_now() { thread::yield_now(); } +} +``` + +**Wasm32**: +```rust +#[cfg(target_arch = "wasm32")] +mod parking { + // Use an AtomicI32 as a futex-like signal. Atomics.wait blocks the + // wasm thread, Atomics.notify wakes it — same semantics as park/unpark. + pub struct ParkHandle(AtomicI32); + pub fn register(_: &ParkHandle) {} // no-op, handle is pre-initialized + pub fn unpark(handle: &ParkHandle) { + handle.0.store(1, Ordering::Release); + std::sync::atomic::fence(Ordering::SeqCst); + // Atomics.notify via core::arch::wasm32::memory_atomic_notify + core::arch::wasm32::memory_atomic_notify(&handle.0 as *const _ as *mut i32, 1); + } + pub fn park_timeout_us(micros: u64) { + // Atomics.wait via core::arch::wasm32::memory_atomic_wait32 + core::arch::wasm32::memory_atomic_wait32(ptr, 0, timeout_ns as i64); + } + pub fn yield_now() {} // no-op on wasm +} +``` + +**Scope**: New `parking` sub-module in `runtime.rs` (~30 lines), update `TickContext` to use `ParkHandle` instead of `OnceLock`, update `worker.rs` backoff loop. + +### 3. Thread Spawning — No Change in Core + +Thread spawning (`std::thread::Builder::new().spawn()`) only happens in `Runtime::run()` (line 351). This method will be overridden/wrapped by the browser crate — it won't be called on wasm32. We can gate it: + +```rust +#[cfg(not(target_arch = "wasm32"))] +pub fn run(self) -> Result { ... } +``` + +The wasm browser crate will provide its own `run()` that spawns Web Workers instead. + +### 4. Validate crossbeam Compilation + +Test that `crossbeam-queue` compiles for wasm32 with atomics. If it doesn't, provide a cfg-gated fallback in `src/channel.rs` using `VecDeque` wrapped in `Mutex`. (Likely not needed — crossbeam uses `core::sync::atomic` which works with wasm atomics.) + +### 5. Feature Flag + +Add a `wasm` feature to `Cargo.toml`: +```toml +[features] +wasm = ["web-time", "no_random"] + +[dependencies] +web-time = { version = "0.2", optional = true } +``` + +On wasm32, this feature enables `web-time` and `no_random` together. + +## Verification + +1. `cargo test` passes unchanged on native +2. `cargo build --target wasm32-unknown-unknown --features wasm -Z build-std=std,panic_abort` compiles (may need `+atomics` RUSTFLAGS) +3. No runtime behavior changes on native (confirm with existing test suite) + +## Estimated Scope + +~50-80 lines of new/changed code across 4 files. No new modules. diff --git a/big-feature-phase/notes/feature-stages/02-single-worker-browser.md b/big-feature-phase/notes/feature-stages/02-single-worker-browser.md new file mode 100644 index 0000000..f48e655 --- /dev/null +++ b/big-feature-phase/notes/feature-stages/02-single-worker-browser.md @@ -0,0 +1,129 @@ +# Stage 2 — Single-Worker Browser Runtime + +**Priority**: P0 +**Depends on**: Stage 1 (platform abstraction) +**Enables**: Stage 3 (multi-worker), Stage 4 (feature parity) + +## Goal + +A working browser runtime on a single dedicated Web Worker with a JS API that supports spawning arbitrary (pre-registered) actor types, sending messages, receiving results, and auto-scheduled ticking. + +## What's Built + +### 1. New Crate: `crates/wasm-browser/` + +Replaces the PoC `crates/wasm/`. Structure: + +``` +crates/wasm-browser/ +├── Cargo.toml +├── src/ +│ ├── lib.rs # wasm-bindgen entry point +│ ├── runtime.rs # BrowserRuntime wrapping swactor::Runtime +│ ├── worker_glue.rs # Web Worker spawn/communication glue +│ └── scheduling.rs # Auto-tick scheduling (setTimeout loop) +├── js/ +│ ├── worker.js # Web Worker bootstrap script +│ └── index.js # Main thread API wrapper (optional) +└── tests/ + └── browser.rs # wasm-pack test suite +``` + +### 2. BrowserRuntime (wasm-bindgen API) + +```rust +#[wasm_bindgen] +pub struct BrowserRuntime { ... } + +#[wasm_bindgen] +impl BrowserRuntime { + #[wasm_bindgen(constructor)] + pub fn new(config: JsValue) -> Self; + + /// Spawn an actor by type name. Returns an opaque handle. + pub fn spawn(&mut self, type_name: &str, init: JsValue) -> JsValue; + + /// Send a message to an actor. + pub fn send(&self, addr: JsValue, msg: JsValue) -> bool; + + /// Drive one tick manually. + pub fn tick(&self); + + /// Start auto-scheduling. Calls tick() in a tight setTimeout(0) loop. + pub fn start(&self); + + /// Stop auto-scheduling. + pub fn stop(&self); + + /// Poll for results from a JS-visible inbox. + pub fn try_recv(&self) -> JsValue; + + /// Runtime stats snapshot. + pub fn stats(&self) -> JsValue; +} +``` + +### 3. Actor Registration + +Since Rust generics can't be dynamically dispatched from JS, actor types are registered at compile time: + +```rust +// In the user's wasm crate that depends on wasm-browser: +register_actors! { + "counter" => Counter, + "relay" => Relay, +} +``` + +This macro generates a factory map that `BrowserRuntime::spawn` indexes by string name. Each entry knows how to deserialize `JsValue` init args into the actor's constructor. + +### 4. Auto-Scheduling + +The runtime self-drives via a `setTimeout(0)` loop: + +```javascript +function tickLoop() { + runtime.tick(); + if (runtime.is_running()) { + setTimeout(tickLoop, 0); + } +} +``` + +This runs as fast as the browser allows (~4ms between ticks in most browsers, faster in Web Workers). The Rust side just calls `tick()` — no async runtime needed. + +### 5. Message Serialization + +JS ↔ Wasm boundary requires serialization. Options: +- **serde-wasm-bindgen**: Serialize Rust types to/from JsValue via serde. Zero-copy for simple types. +- **Manual**: Convert JsValue to bytes, route as `ByteMessage`. + +For Stage 2, use `serde-wasm-bindgen` for typed messages. Actor `Incoming` types must implement `serde::Deserialize`. + +### 6. Single Worker Architecture + +``` +┌─────────────────────┐ postMessage ┌──────────────────────┐ +│ Main Thread │ ◄──────────────────────► │ Web Worker │ +│ │ │ │ +│ JS application │ "spawn", "send", │ BrowserRuntime │ +│ calls API methods │ "tick", "recv" │ swactor::Runtime │ +│ │ │ (1 worker, tick()) │ +└─────────────────────┘ └──────────────────────┘ +``` + +The Web Worker runs the swactor runtime. The main thread sends commands via `postMessage`. This keeps the UI thread free. + +Alternative: run everything on the main thread (simpler, but blocks UI during tick). Support both modes — the user picks. + +## Verification + +1. `wasm-pack build --target web` succeeds +2. `wasm-pack test --headless --chrome` passes +3. Manual test: HTML page spawns actors, sends messages, receives results +4. Auto-scheduling: actors process messages continuously without manual tick calls +5. Performance: measure ticks/sec, compare to native single-threaded + +## Estimated Scope + +~300-500 lines of Rust + ~50 lines of JS glue. diff --git a/big-feature-phase/notes/feature-stages/03-multi-worker-parallelism.md b/big-feature-phase/notes/feature-stages/03-multi-worker-parallelism.md new file mode 100644 index 0000000..2e1a2a1 --- /dev/null +++ b/big-feature-phase/notes/feature-stages/03-multi-worker-parallelism.md @@ -0,0 +1,118 @@ +# Stage 3 — Multi-Worker Parallelism + +**Priority**: P1 +**Depends on**: Stage 1 (platform abstraction), Stage 2 (single-worker browser) +**Enables**: Stage 5 (transport) + +## Goal + +Spawn N Web Workers sharing the same swactor runtime via SharedArrayBuffer. Actors distributed across workers for true multi-core parallelism. Worker count configurable, defaults to `navigator.hardwareConcurrency`. + +## Architecture + +``` +┌─────────────┐ +│ Main Thread │ postMessage API +│ (JS app) │◄─────────────────────┐ +└──────┬───────┘ │ + │ spawn workers │ + ▼ │ +┌──────────────┐ SharedArrayBuffer ┌──────────────┐ +│ Web Worker 0 │◄───────────────────►│ Web Worker 1 │ +│ swactor │ (InboxRegistry, │ swactor │ +│ Worker #0 │ AddressMap, │ Worker #1 │ +│ tick loop │ transfer queues, │ tick loop │ +└──────────────┘ atomics) └──────────────┘ + ... + ┌──────────────┐ + │ Web Worker N │ + │ swactor │ + │ Worker #N │ + └──────────────┘ +``` + +## Key Challenge: Web Worker ↔ SharedArrayBuffer + +Web Workers can share `SharedArrayBuffer` instances. The swactor `Runtime` struct contains `Arc`-wrapped shared state (InboxRegistry, AddressMap, etc.). On native, this memory is shared via the process address space. On wasm with SharedArrayBuffer, it's shared via the underlying wasm linear memory. + +### How It Works + +1. **Main thread** creates the `Runtime` (allocates shared structures in wasm linear memory) +2. **Main thread** spawns N Web Workers, each loading the same `.wasm` module with `shared: true` memory +3. Each Web Worker receives a pointer (offset) to the `Runtime` shared state +4. Each Worker runs `worker.run(&tc, &is_running)` — the same tick loop as native +5. Crossbeam queues, atomics, Mutex/RwLock all work because the underlying memory is shared + +### wasm-bindgen + Web Workers + +The `web-sys::Worker` API creates workers. Each worker loads the same wasm module: + +```javascript +// worker.js (loaded by each Web Worker) +import init, { worker_entry } from './pkg/wasm_browser.js'; + +self.onmessage = async (e) => { + const { module, memory, worker_id, runtime_ptr } = e.data; + await init({ module, memory }); // shared memory! + worker_entry(worker_id, runtime_ptr); +}; +``` + +The Rust side: +```rust +#[wasm_bindgen] +pub fn worker_entry(worker_id: usize, runtime_ptr: u32) { + // Reconstruct the shared Runtime reference from the raw pointer + // Run the tick loop for this worker +} +``` + +### Memory Sharing + +With `wasm-threads`, the wasm linear memory is backed by a `SharedArrayBuffer`. All workers see the same memory. `Arc` increments are atomic operations on shared memory. Crossbeam queues use atomic compare-and-swap on shared memory. This is identical to how it works with OS threads. + +**Critical**: The wasm module must be compiled with `--shared-memory` and the `atomics` feature. The `Memory` import must use `shared: true`. + +## Park / Unpark + +The `ParkHandle` from Stage 1 uses `memory_atomic_wait32` / `memory_atomic_notify` — the wasm equivalents of futex. These work across Web Workers sharing the same memory. + +- `park_timeout_us(micros)` → `Atomics.wait(ptr, expected, timeout)` — blocks the Worker thread +- `unpark(handle)` → `Atomics.notify(ptr, 1)` — wakes one waiting Worker + +This gives the same backoff behavior as native: hot spin → yield → sleep with exponential backoff. + +## Worker Spawning + +Replace `std::thread::Builder::new().spawn()` in `Runtime::run()`: + +```rust +#[cfg(target_arch = "wasm32")] +pub fn run(self) -> Result { + let num_workers = self.config.num_threads.max(1); + let rt = Arc::new(self); + + for i in 0..num_workers { + let worker = web_sys::Worker::new("./worker.js")?; + worker.post_message(&JsValue::from(/* module, memory, worker_id, ptr */)); + } + // ... +} +``` + +## Placement + +The existing `Placement` strategy (load-aware round-robin) works unchanged — it reads `WorkerStats` atomics to pick the least-loaded worker. On wasm, these atomics are in SharedArrayBuffer, readable from any worker. + +## Verification + +1. Spawn runtime with `num_threads: 4`, verify 4 Web Workers created +2. Spawn actors, verify they're distributed across workers (check stats per-worker actor count) +3. Cross-worker message delivery works (actor on Worker 0 sends to actor on Worker 1) +4. Backoff/parking works (idle workers sleep, wake on new messages) +5. Throughput scales with worker count (benchmark: N workers vs 1 worker) +6. Shutdown: all workers terminate cleanly when `is_running` set to false + +## Estimated Scope + +~200-300 lines Rust + ~30 lines JS worker bootstrap. Most complexity is in the Web Worker ↔ shared memory plumbing, not the Rust logic (which is the same as native). diff --git a/big-feature-phase/notes/feature-stages/04-feature-parity.md b/big-feature-phase/notes/feature-stages/04-feature-parity.md new file mode 100644 index 0000000..2e49512 --- /dev/null +++ b/big-feature-phase/notes/feature-stages/04-feature-parity.md @@ -0,0 +1,109 @@ +# Stage 4 — Feature Parity + +**Priority**: P1 +**Depends on**: Stage 2 (single-worker) or Stage 3 (multi-worker) +**Enables**: Stage 5 (transport), Stage 6 (DX) + +## Goal + +All swactor features that are feasible in a browser environment work and are tested: actor watching, swactor-std extensions (naming, groups, monitoring), stats introspection. + +## Features to Enable + +### 1. Actor Watching / Death Notifications + +**Status**: Already in core (`src/runtime.rs`, `src/worker.rs`) + +Components: +- `ExitReason` enum — normal, panic, stopped +- `ActorExited` message — delivered to watchers +- `on_actor_exit()` default method on `ActorInterface` +- `WatchRegistry` — tracks who watches whom +- Phase 5b in `tick_once` — delivers death notifications + +**Wasm concern**: `WatchRegistry` is behind `Arc>`. With wasm-threads, `Mutex` works. The catch_unwind panic-safety model works identically in wasm. + +**Work needed**: Compile and test. Write wasm-specific tests for: +- Actor dies → watchers notified +- Watcher on different Web Worker receives notification (cross-worker) +- Panic in wasm actor → poisoned, watchers notified + +### 2. swactor-std Extension + +**Status**: Complete in `crates/std/` + +Components: +- `StdExtension` — wraps NameRegistry + MonitorRegistry + GroupRegistry +- `CtxMonitoring` — watch/unwatch actors +- `CtxNaming` — register/resolve actor names +- `CtxGroups` — join/leave groups, broadcast +- `RuntimeNaming` — resolve names from runtime handle +- `RuntimeGroups` — list groups, broadcast from outside +- Supervisor — restart policies + +**Wasm concern**: All use `Arc`, `Mutex`, `HashMap` — standard types that work with wasm-threads. No OS-specific dependencies. + +**Work needed**: +- Add `crates/std/` to wasm build verification +- Test naming: register name → resolve from another actor on different worker +- Test groups: broadcast reaches all group members across workers +- Test supervisor: child dies → supervisor restarts (factory-based) + +### 3. Stats and Introspection + +**Status**: In core (`src/stats.rs`) + +Components: +- `WorkerStats` — per-worker atomic counters (actors, depth, ticks, messages) +- `RuntimeStats` — aggregated snapshot +- `StatsHook` trait — called each tick with stats + +**Wasm concern**: Atomic counters work with wasm-threads. `StatsHook` is called in the tick loop — works. + +**Work needed**: +- Expose `RuntimeStats` to JS via `serde-wasm-bindgen` (JSON-serializable snapshot) +- Optional: periodic stats push to main thread via `postMessage` +- Test: spawn actors across workers, verify stats reflect correct counts + +### 4. Runtime Extensions + +**Status**: In core (`src/extension.rs`) + +The `RuntimeExtension` trait (`on_actor_death`, `cleanup_dead`, `as_any`) is called during phase 7 of tick_once. It uses `Arc` — works with wasm-threads. + +**Work needed**: Verify `StdExtension` as a `RuntimeExtension` compiles and works in wasm. Test the full lifecycle: actor death → extension notified → cleanup runs. + +## JS API Additions + +Extend the `BrowserRuntime` wasm-bindgen API: + +```rust +impl BrowserRuntime { + // Actor watching + pub fn watch(&self, watcher: JsValue, target: JsValue) -> bool; + + // Naming (if StdExtension enabled) + pub fn register_name(&self, name: &str, addr: JsValue) -> bool; + pub fn resolve_name(&self, name: &str) -> JsValue; + + // Groups + pub fn join_group(&self, group: &str, addr: JsValue) -> bool; + pub fn broadcast_group(&self, group: &str, msg: JsValue) -> bool; + + // Stats + pub fn stats(&self) -> JsValue; // JSON snapshot of RuntimeStats +} +``` + +## Verification + +1. All existing native tests for watching/std pass on wasm target +2. Cross-worker watching: actor on Worker 0 watches actor on Worker 1, Worker 1 actor dies → notification arrives +3. Naming works across workers: register on Worker 0, resolve on Worker 1 +4. Group broadcast reaches actors on all workers +5. Stats counters are accurate across workers (compare sum to expected) +6. `cargo test` still passes on native (no regressions) + +## Estimated Scope + +~100-200 lines of new wasm-bindgen API surface + ~200 lines of wasm tests. Core logic should work as-is once it compiles. diff --git a/big-feature-phase/notes/feature-stages/05-transport-foundation.md b/big-feature-phase/notes/feature-stages/05-transport-foundation.md new file mode 100644 index 0000000..3c3b96e --- /dev/null +++ b/big-feature-phase/notes/feature-stages/05-transport-foundation.md @@ -0,0 +1,107 @@ +# Stage 5 — Transport Foundation + +**Priority**: P2 +**Depends on**: Stage 3 (multi-worker), Stage 4 (feature parity) +**Enables**: Browser nodes joining distributed swactor clusters + +## Goal + +Browser nodes connect to native swactor clusters via WebSocket. A browser can spawn actors that communicate with actors on server nodes. Foundation for future STUN/TURN (WebRTC DataChannel) for browser-to-browser direct connections. + +## Architecture + +``` +┌──────────────────┐ WebSocket ┌──────────────────┐ +│ Browser Node │ ◄────────────────────────► │ Server Node │ +│ (wasm runtime) │ │ (native runtime)│ +│ │ swactor wire protocol │ │ +│ Actor A ──────►─┤───── msg for Actor B ─────►├──► Actor B │ +│ │ │ │ +│ Actor C ◄───────┤◄──── msg for Actor C ─────┤───── Actor D │ +└──────────────────┘ └──────────────────┘ +``` + +## Existing Transport Infrastructure + +swactor already has a transport layer (feature-gated under `transport`): + +- `src/transport.rs` — `TransportRouter`, `CodecRegistry`, remote message routing +- `crates/distribution/` — SWIM protocol, gossip, cluster membership +- `crates/distribution/src/driver.rs` — `NodeDriver` bridges `DistributedNode` ↔ TCP +- Wire protocol: Ping/Ack/PingReq with piggyback bytes + +The browser transport needs to implement the same wire protocol over WebSocket instead of raw TCP. + +## What's Built + +### 1. WebSocket Transport Adapter + +A new module in `crates/wasm-browser/` (not in core): + +```rust +pub struct WebSocketTransport { + ws: web_sys::WebSocket, + // ... +} + +impl TransportAdapter for WebSocketTransport { + fn send(&self, dest: SocketAddr, data: &[u8]) -> Result<(), Error>; + fn recv(&self) -> Option<(SocketAddr, Vec)>; +} +``` + +Uses `web-sys::WebSocket` for the browser side. The server side uses a WebSocket server (e.g., `tokio-tungstenite`) that bridges to the existing TCP transport. + +### 2. WebSocket ↔ TCP Bridge (Server Side) + +A thin relay server that accepts WebSocket connections from browsers and translates to/from the TCP wire protocol: + +``` +Browser ──WebSocket──► Bridge Server ──TCP──► swactor-node +``` + +This bridge is a separate binary/service, not part of the runtime. It's a protocol translator. + +### 3. Browser Node Identity + +Browser nodes need: +- A unique node ID (derived from random or assigned by the cluster) +- An address for the cluster to route messages to (the WebSocket endpoint) +- Membership in the SWIM protocol (lightweight — browsers are "client" members that don't participate in failure detection) + +### 4. Cluster Registry Integration + +The existing ClusterRegistry (LWW-Register CRDT in `crates/distribution/src/registry.rs`) should work from browsers: +- `register_name` / `resolve_name` / `registry_events` — all work over the wire +- Piggyback payloads carry registry updates through the WebSocket connection + +## Key Design Decisions + +| Decision | Choice | Rationale | +|----------|--------|-----------| +| Transport protocol | WebSocket (binary frames) | Universal browser support, bidirectional, binary-capable | +| Membership role | Client member (no failure detection) | Browsers are ephemeral; full SWIM overhead not justified | +| Bridge architecture | Separate relay server | Keeps swactor-node unchanged; bridge handles WebSocket↔TCP | +| Wire format | Same as TCP transport | No translation needed beyond framing (WebSocket frames ↔ TCP stream) | + +## STUN/TURN Foundation (Future) + +This stage establishes the transport abstraction. Stage 5 itself is WebSocket only. Future work: + +- **WebRTC DataChannel** — direct browser-to-browser, requires STUN/TURN for NAT traversal +- The `TransportAdapter` trait from this stage will have a WebRTC implementation +- STUN/TURN server infrastructure is out of scope for this feature phase + +## Verification + +1. Browser node connects to server cluster via WebSocket +2. Actor on browser sends message to actor on server → received +3. Actor on server sends message to actor on browser → received +4. Browser appears in cluster membership (visible in dashboard) +5. ClusterRegistry: name registered on server → resolvable from browser +6. Browser disconnects → cluster detects and removes membership +7. Reconnection: browser reconnects → re-joins cluster, actor addresses still valid + +## Estimated Scope + +~500-800 lines for WebSocket transport adapter + bridge server. Builds heavily on existing distribution infrastructure. diff --git a/big-feature-phase/notes/feature-stages/06-developer-experience.md b/big-feature-phase/notes/feature-stages/06-developer-experience.md new file mode 100644 index 0000000..dab3283 --- /dev/null +++ b/big-feature-phase/notes/feature-stages/06-developer-experience.md @@ -0,0 +1,102 @@ +# Stage 6 — Developer Experience + +**Priority**: P3 +**Depends on**: Stage 2 (single-worker browser), Stage 4 (feature parity) +**Enables**: Adoption, ecosystem growth + +## Goal + +Make it easy for Rust developers to build browser applications with swactor. TypeScript type safety, build tooling, and debugging support. + +## Features + +### 1. TypeScript Type Generation + +Derive TypeScript interfaces from Rust actor message types. When an actor defines: + +```rust +#[derive(Serialize, Deserialize)] +pub struct ChatMessage { + pub from: String, + pub text: String, +} +``` + +Generate: +```typescript +export interface ChatMessage { + from: string; + text: string; +} +``` + +**Approach**: Use `ts-rs` crate or a custom proc macro that emits `.d.ts` files during `wasm-pack build`. This gives TypeScript consumers compile-time type checking for messages. + +### 2. Build Tooling + +A `swactor-build` CLI or build script that wraps: +```bash +RUSTFLAGS='-C target-feature=+atomics,+bulk-memory,+mutable-globals' \ + cargo +nightly build --target wasm32-unknown-unknown \ + -Z build-std=std,panic_abort \ + --release +wasm-bindgen --target web --out-dir pkg/ ... +``` + +Into: +```bash +swactor-build --target browser +``` + +Features: +- Detects nightly toolchain, installs if missing +- Sets correct RUSTFLAGS for wasm-threads +- Runs wasm-bindgen with correct target +- Copies worker.js bootstrap into output +- Generates example HTML with correct COOP/COEP headers + +### 3. Example Project Template + +A `cargo generate` template or example project: + +``` +my-swactor-app/ +├── Cargo.toml +├── src/ +│ └── lib.rs # Define actors, register them +├── web/ +│ ├── index.html # With COOP/COEP headers +│ ├── main.js # Import wasm, create runtime, interact +│ └── worker.js # Web Worker bootstrap (auto-generated) +└── tests/ + └── browser.rs # wasm-pack tests +``` + +### 4. Browser Dev Tools Integration + +Expose runtime internals for debugging: + +- **Actor Inspector**: List all actors, their types, mailbox depths, message counts +- **Message Tracer**: Log messages between actors (opt-in, performance impact) +- **Performance Monitor**: Ticks/sec, messages/sec, worker utilization + +Implementation: A `console`-based logger that uses `web-sys::console` to output structured data. Optionally integrates with browser DevTools via custom formatters or a small React/Preact inspector panel. + +### 5. Documentation + +- Getting started guide +- Architecture overview for browser runtime +- Migration guide from native → browser (what works, what doesn't) +- API reference (generated from Rust doc comments) +- Example: Chat application with multiple browser tabs + +## Verification + +1. TypeScript types match Rust types (compile TS project against generated `.d.ts`) +2. Build tool produces working wasm output from example project +3. Template project builds and runs out of the box +4. Dev tools show actor state in browser console + +## Estimated Scope + +Variable — this stage is a collection of independent DX improvements. Each can be implemented and shipped independently. Total: ~500-1000 lines across Rust, JS, and documentation. diff --git a/big-feature-phase/notes/history.md b/big-feature-phase/notes/history.md new file mode 100644 index 0000000..67e6a3f --- /dev/null +++ b/big-feature-phase/notes/history.md @@ -0,0 +1,4 @@ +# Cycle History (append-only) + +## Cycle 0 — Research (complete) +Investigated Lunatic, wasmCloud, Actix-wasm attempts. Analyzed core swactor platform deps: 4 blockers (thread spawn, park/unpark, yield, Instant). User confirmed: performance-first, SharedArrayBuffer+wasm-threads, Rust-only actors, all feasible features, future STUN/TURN. Produced constraints.md, research_synthesis.md, 6 feature-stage docs. Next: Stage 1 platform abstraction. diff --git a/big-feature-phase/notes/research_synthesis.md b/big-feature-phase/notes/research_synthesis.md new file mode 100644 index 0000000..aaa5da6 --- /dev/null +++ b/big-feature-phase/notes/research_synthesis.md @@ -0,0 +1,87 @@ +# Research Synthesis — In-Browser Swactor Runtime + +## Ecosystem Landscape + +No established Rust actor framework runs natively in browsers. The closest projects: + +- **Lunatic** — Erlang-inspired Wasm actor runtime using wasmtime (server-side, not browser). Uses preemptive scheduling and work-stealing. Not applicable to browser constraints. +- **wasmCloud** — CNCF distributed actor platform. Single-threaded actors, NATS-backed lattice. Cloud/edge focus, no browser target. +- **Actix** — Tokio-dependent, network stack doesn't compile for wasm32. Community attempts to port failed due to `net2`/tokio dependencies. + +**Implication**: swactor would be the first Rust actor runtime with true multi-threaded browser execution via wasm-threads. This is a differentiated position. + +## Existing Work in This Codebase + +| Component | Status | Notes | +|-----------|--------|-------| +| `crates/wasm/` | Basic PoC | Hardcoded Counter/Relay actors, manual tick, u32-only messages | +| `no_random` feature | Working | Deterministic address generation without `getrandom` | +| `tick()` method | Working | Single-threaded tick for manual driving | +| Actor watching | In core | `ExitReason`, `ActorExited`, `on_actor_exit`, `WatchRegistry` | +| swactor-std | Complete | StdExtension, naming, groups, monitoring, supervisor | + +## Platform Dependencies Analysis + +### Works as-is with wasm-threads +- `crossbeam-queue` (ArrayQueue, SegQueue) — uses `core::sync::atomic` +- `std::sync::{Mutex, RwLock}` — stdlib uses futex on wasm with atomics +- `std::sync::atomic::*` — maps to wasm atomic instructions +- `Arc` — works with atomics +- `std::sync::OnceLock` — works with atomics + +### Requires platform abstraction (4 items) +1. `std::thread::spawn` → Web Worker via `web-sys::Worker` +2. `thread::park_timeout` / `Thread::unpark` → `Atomics.wait` / `Atomics.notify` +3. `thread::yield_now` → no-op (or `Atomics.wait(0)` as hint) +4. `std::time::Instant` → `web_time::Instant` (drop-in crate) + +## Priority Ranking + +### P0 — Must Have (enables everything else) + +1. **Platform abstraction layer** — cfg-gated replacements for thread spawn, park/unpark, yield, Instant. Core swactor compiles for wasm32 with atomics. +2. **Single-worker browser runtime** — Prove the runtime works in a browser. One Web Worker, auto-scheduled tick loop, generic JS API for spawn/send/recv. +3. **Multi-worker parallelism** — N Web Workers sharing runtime state via SharedArrayBuffer. Full utilization of browser CPU cores. + +### P1 — Should Have (full actor system) + +4. **Actor watching in browser** — Death notifications, exit reasons. Already in core, just needs to compile and pass wasm tests. +5. **swactor-std in browser** — Naming, groups, monitoring extensions. Compile and test for wasm32. +6. **Stats and introspection** — Runtime stats accessible from JS. Worker info, actor counts, message throughput. + +### P2 — Important (distributed peer) + +7. **WebSocket transport** — Adapter implementing swactor's transport traits over WebSocket. Browser node joins a distributed cluster. +8. **Browser-to-browser transport foundation** — WebRTC DataChannel scaffolding for future STUN/TURN. + +### P3 — Nice to Have (developer experience) + +9. **TypeScript type generation** — Derive TS interfaces from Rust actor message types. +10. **Build tooling** — wasm-pack wrapper script, example project template, CI configuration. +11. **Browser dev tools** — Actor inspector, message flow visualization, performance profiling. + +### P4 — Future (out of scope for this feature phase) + +12. **STUN/TURN integration** — Full NAT traversal for peer-to-peer browser connections. +13. **Hot code reload** — Swap actor implementations without restarting the runtime. +14. **Wasm component model** — Migrate from wasm-bindgen to component model when stabilized. + +## Key Design Decisions + +| Decision | Choice | Rationale | +|----------|--------|-----------| +| Threading model | SharedArrayBuffer + wasm-threads | swactor's shared-memory architecture (Arc, crossbeam queues, atomics) maps directly. postMessage isolation would require a rewrite. | +| Scheduling | `setTimeout(0)` tight loop | `requestAnimationFrame` caps at 60Hz. setTimeout(0) gives ~4ms resolution, sufficient for actor ticks. For rendering-coupled actors, RAF can be opt-in. | +| Actor definition | Rust only | Keeps the type system intact. JS actors would require dynamic dispatch and lose compile-time guarantees. | +| Platform abstraction approach | cfg-gated type aliases + inline functions | Minimal invasion. No trait-based HAL, no new modules. Just swap `std::time::Instant` → `web_time::Instant` etc. | +| Browser crate location | New `crates/wasm-browser/` | Clean separation from the existing PoC. The old `crates/wasm/` can be deprecated or kept as a minimal example. | + +## Risk Assessment + +| Risk | Likelihood | Impact | Mitigation | +|------|-----------|--------|------------| +| crossbeam-queue doesn't compile for wasm32+atomics | Low | High | Test early in Stage 1. Fallback: thin wrapper over `VecDeque` behind cfg. | +| COOP/COEP headers break third-party integrations | Medium | Medium | Document clearly. This is the standard trade-off for SharedArrayBuffer apps. | +| Nightly Rust + build-std is unstable | Medium | Medium | Pin to a known-good nightly. Revisit when wasm-threads stabilizes. | +| Web Worker spawn latency | Low | Low | Spawn workers eagerly at runtime init, not per-actor. | +| Browser compatibility gaps | Low | Medium | Target Chrome/Firefox/Safari latest. All support SharedArrayBuffer since 2021. | diff --git a/big-feature-phase/notes/state.md b/big-feature-phase/notes/state.md new file mode 100644 index 0000000..e0f952a --- /dev/null +++ b/big-feature-phase/notes/state.md @@ -0,0 +1,46 @@ +## Current Task +Cycle 0 — Research +Step: Complete +Attempt: 1 of 3 + +## Key Files (read these first on resume) +- `big-feature-phase/TASK.md` — workflow rules +- `big-feature-phase/notes/constraints.md` — guardrails for in-browser runtime +- `big-feature-phase/notes/research_synthesis.md` — P0-P4 priority ranking +- `big-feature-phase/notes/feature-stages/01-platform-abstraction.md` — Stage 1 spec +- `src/runtime.rs` — thread spawn (line 351), park handles (line 110), Instant (line 205) +- `src/worker.rs` — backoff loop (lines 495-523), Instant timing (9 calls) +- `src/delivery.rs` — TickContext.worker_threads (line 246) +- `src/channel.rs` — crossbeam HybridChannel (validate wasm compilation) +- `crates/wasm/src/lib.rs` — existing PoC browser runtime + +## Last Action & Result +Completed Cycle 0 research phase: +- Investigated similar projects (Lunatic, wasmCloud, Actix wasm attempts) +- Analyzed all platform-specific code in core swactor (4 blockers found) +- User confirmed: Rust-only actors, performance-first, Web Workers, all features, full peer +- Key decision: SharedArrayBuffer + wasm-threads (not postMessage isolation) +- Wrote constraints.md, research_synthesis.md, 6 feature-stage docs + +## Next Action +Begin Stage 1 (Platform Abstraction Layer) — `notes/feature-stages/01-platform-abstraction.md`: +1. Add `web-time` dependency to core Cargo.toml +2. Replace `std::time::Instant` with `web_time::Instant` in runtime.rs and worker.rs +3. Implement `ParkHandle` abstraction with cfg gates +4. Gate `Runtime::run()` thread spawning for non-wasm +5. Validate crossbeam-queue compiles for wasm32 with atomics +6. Add `wasm` feature flag + +## Completed This Session +- [x] Read TASK.md, state.md, history.md (context restart checklist) +- [x] Investigated similar codebases (Lunatic, wasmCloud, Actix-wasm) +- [x] Analyzed platform dependencies in core swactor +- [x] Asked and resolved design questions with user +- [x] Wrote `notes/constraints.md` +- [x] Wrote `notes/research_synthesis.md` +- [x] Created `notes/feature-stages/` with 6 stage docs + +## Open Questions / Blockers +- Need to verify crossbeam-queue compiles for wasm32 with atomics (Stage 1 task) +- Need to find/verify correct `core::arch::wasm32` APIs for memory_atomic_wait/notify +- Nightly Rust toolchain + wasm32 target needs to be set up in CI