Run the desktop's model layer in the browser (#557)

packages/platform/src/web/worker.ts

@@ -0,0 +1,191 @@
+/// <reference lib="webworker" />
+
+/**
+ * The process that owns the database.
+ *
+ * On the desktop that is the Rust binary: it holds SQLite, every window talks
+ * to it, and it pushes model writes to all of them. In a browser this worker
+ * plays that part. It loads the model layer compiled to wasm, opens the one
+ * database, answers each tab's commands over its port, and fans every
+ * `model_writes` out to every port — so two tabs are coherent for the same
+ * reason two desktop windows are, not because of a side channel.
+ *
+ * It is a SharedWorker, and only that: the browser hands every tab on the
+ * origin the same one, which is what makes "one database owner" true without
+ * anyone coordinating. It still takes a Web Lock before opening the database,
+ * for the one overlap the browser doesn't rule out — a tab reloading itself,
+ * whose old worker may still be letting go while the new one comes up.
+ */
+
+import { DB_LOCK_NAME, type FromWorker, type ToWorker } from "./protocol";
+
+/**
+ * The wasm is imported lazily, inside `boot()`, rather than at the top of the
+ * module. That keeps this script's own evaluation instant, so a tab's connect
+ * gets its `hello` immediately regardless of how long the model layer takes to
+ * download and compile — and the tab can therefore tell "this worker is dead"
+ * from "this worker is busy" with a short timeout.
+ */
+type Engine = typeof import("@yaakapp-internal/web");
+let engine: Engine | null = null;
+
+const ports = new Set<MessagePort>();
+
+/** Resolves once `boot()` has, or rejects with why it couldn't. */
+let booted: Promise<void> | null = null;
+let bootError: string | null = null;
+
+function send(port: MessagePort, message: FromWorker, transfer: Transferable[] = []): void {
+  port.postMessage(message, transfer);
+}
+
+function broadcast(message: FromWorker): void {
+  for (const port of ports) send(port, message);
+}
+
+function errorMessage(err: unknown): string {
+  if (err instanceof Error) return err.message;
+  return String(err);
+}
+
+/**
+ * How long to wait for the database lock before concluding someone else has it
+ * for good. The wait exists for one case: a tab reloading itself. Its old
+ * worker still holds the lock while it is torn down, and the new worker only
+ * needs it to let go — which takes milliseconds, not seconds. Anything longer
+ * is a live worker in another tab, and the honest answer is to say so.
+ */
+const LOCK_TIMEOUT_MS = 3000;
+
+const ALREADY_OPEN =
+  "Yaak's database is held by another worker that isn't letting go. Close Yaak's other tabs and reload.";
+
+/**
+ * Take the lock, or explain why not.
+ *
+ * Held for the life of the worker: the callback's promise never settles, so
+ * the browser keeps the lock until this worker is gone. Requested with a
+ * timeout rather than `ifAvailable`, so a dying predecessor's brief hold is
+ * waited out but a live one is reported.
+ */
+function acquireDatabaseLock(): Promise<void> {
+  // The tab checked for Web Locks before it ever connected; a worker without
+  // them would be a browser lying about its own features.
+  return new Promise<void>((resolve, reject) => {
+    navigator.locks
+      .request(DB_LOCK_NAME, { signal: AbortSignal.timeout(LOCK_TIMEOUT_MS) }, () => {
+        resolve();
+        return new Promise<void>(() => {});
+      })
+      .catch((err: unknown) => {
+        const name = (err as { name?: string } | null)?.name;
+        reject(name === "AbortError" || name === "TimeoutError" ? new Error(ALREADY_OPEN) : err);
+      });
+  });
+}
+
+/**
+ * Open the database, once, for everyone.
+ *
+ * Lock first, then load the model layer, then open — in that order, so a
+ * worker that will never own the database also never downloads and compiles
+ * the wasm for it.
+ */
+function bootOnce(): Promise<void> {
+  if (booted != null) return booted;
+
+  booted = (async () => {
+    await acquireDatabaseLock();
+    const loaded = await import("@yaakapp-internal/web");
+    await loaded.boot();
+    engine = loaded;
+  })();
+
+  booted.catch((err) => {
+    bootError = errorMessage(err);
+  });
+
+  return booted;
+}
+
+async function handle(port: MessagePort, message: ToWorker): Promise<void> {
+  if (message.type === "goodbye") {
+    ports.delete(port);
+    return;
+  }
+
+  // Every command waits for boot rather than the tab having to. Tabs post
+  // the moment they load; the port queues; this drains once the DB is open.
+  try {
+    await booted;
+  } catch {
+    send(port, { type: "error", id: message.id, message: bootError ?? "Database failed to open" });
+    return;
+  }
+  const { rpc, blob_get, blob_put, blob_delete } = engine!;
+
+  try {
+    switch (message.type) {
+      case "rpc": {
+        const outcome = rpc(message.cmd, message.payload, message.label) as {
+          result: unknown;
+          events: unknown[];
+        };
+        // Result first, to the caller; then the writes, to everyone including
+        // the caller. The store applies its own echo the same as any other
+        // window's, so it must arrive — and the caller's `await` resolving
+        // before its echo lands is fine, because it resolves on the same tick
+        // and the store reads on the next.
+        send(port, { type: "result", id: message.id, result: outcome.result });
+        if (outcome.events.length > 0) {
+          broadcast({ type: "event", event: "model_writes", payload: outcome.events });
+        }
+        return;
+      }
+      case "blob_get": {
+        const bytes = blob_get(message.blobId);
+        if (bytes == null) {
+          send(port, { type: "result", id: message.id, result: null });
+        } else {
+          // Copy into a fresh buffer we can transfer: the wasm's memory
+          // cannot leave the worker.
+          const out = new Uint8Array(bytes.byteLength);
+          out.set(bytes);
+          send(port, { type: "result", id: message.id, result: out.buffer }, [out.buffer]);
+        }
+        return;
+      }
+      case "blob_put": {
+        blob_put(message.blobId, new Uint8Array(message.bytes));
+        send(port, { type: "result", id: message.id, result: null });
+        return;
+      }
+      case "blob_delete": {
+        blob_delete(message.blobId);
+        send(port, { type: "result", id: message.id, result: null });
+        return;
+      }
+    }
+  } catch (err) {
+    send(port, { type: "error", id: message.id, message: errorMessage(err) });
+  }
+}
+
+function attach(port: MessagePort): void {
+  ports.add(port);
+  port.onmessage = (e: MessageEvent<ToWorker>) => void handle(port, e.data);
+  port.start?.();
+
+  // Proof of life, before boot: the tab is timing this.
+  send(port, { type: "hello" });
+
+  bootOnce().then(
+    () => send(port, { type: "ready" }),
+    (err) => send(port, { type: "boot_error", message: errorMessage(err) }),
+  );
+}
+
+// Each tab arrives as a connect event with its own port.
+(self as unknown as SharedWorkerGlobalScope).onconnect = (e: MessageEvent) => {
+  attach(e.ports[0]!);
+};