cairnq 0.6.0 → 0.7.0

package/dist/wait.js

@@ -4,6 +4,7 @@ import { isTerminal } from "./models.js";
 export const DEFAULT_POLL_MS = 100;
 export const MAX_POLL_MS = 500;
 const GROWTH = 1.5;
+const sleep = (ms) => new Promise((resolve) => setTimeout(resolve, ms));
 /**
  * Grow the polling interval towards the ceiling.
  *
@@ -15,22 +16,46 @@ const GROWTH = 1.5;
 export function nextPollMs(current, maxMs) {
     return Math.min(maxMs, Math.max(current + 1, Math.floor(current * GROWTH)));
 }
-/** Poll get() until terminal or timeout. Returns the terminal Task (any status).
- * Throws TaskTimeout, leaving the task running. `pollMs` is the *first* interval;
- * it backs off towards `maxPollMs`. */
-export async function pollWait(store, taskId, { timeoutMs, pollMs = DEFAULT_POLL_MS, maxPollMs = MAX_POLL_MS, }) {
+/**
+ * Poll `read` until it yields a terminal task, or the timeout elapses.
+ *
+ * `wake` is what the loop sleeps on between reads: a store with a push channel
+ * (Postgres) cuts it short when the task goes terminal, but the re-read is the
+ * source of truth either way, so a plain sleep is always a correct answer.
+ */
+async function poll(read, wake, subject, key, { timeoutMs, pollMs = DEFAULT_POLL_MS, maxPollMs = MAX_POLL_MS }) {
     const deadline = nowMs() + timeoutMs;
     let interval = pollMs;
     for (;;) {
-        const task = await store.get(taskId);
+        const task = await read();
         if (task && isTerminal(task))
             return task;
         const remaining = deadline - nowMs();
         if (remaining <= 0)
-            throw new TaskTimeout(taskId, { timeoutMs, task });
-        // A store with a push channel (Postgres) cuts the sleep short when the task
-        // goes terminal; the re-get above stays the source of truth either way.
-        await store.taskDoneWake(taskId, Math.min(interval, remaining));
+            throw new TaskTimeout(task?.id ?? subject, { timeoutMs, task, key });
+        await wake(task, Math.min(interval, remaining));
         interval = nextPollMs(interval, maxPollMs);
     }
 }
+/** Poll get() until terminal or timeout. Returns the terminal Task (any status).
+ * Throws TaskTimeout, leaving the task running. `pollMs` is the *first* interval;
+ * it backs off towards `maxPollMs`. */
+export function pollWait(store, taskId, opts) {
+    return poll(() => store.get(taskId), (_task, ms) => store.taskDoneWake(taskId, ms), taskId, null, opts);
+}
+/**
+ * The same wait, following a key instead of an id.
+ *
+ * The key is re-resolved on every read, because that is what a key means: a
+ * pointer to the task that is *current* under it. A `replace` landing mid-wait
+ * moves the wait onto the new task rather than reporting the cancellation of the
+ * old one, and a key that points at nothing yet is simply not finished — it
+ * polls until something appears, the same way waiting on an id that does not
+ * exist yet does.
+ *
+ * There is nothing to subscribe to before the key resolves, so those naps are
+ * plain sleeps; once it resolves, the store's push channel applies as usual.
+ */
+export function pollWaitByKey(store, key, opts) {
+    return poll(() => store.getByKey(key), (task, ms) => (task ? store.taskDoneWake(task.id, ms) : sleep(ms)), key, key, opts);
+}

package/dist/worker.js

@@ -1,6 +1,6 @@
 import { DEFAULT_RETRY_BACKOFF_MAX_MS, DEFAULT_RETRY_BACKOFF_MS, failDelayMs } from "./backoff.js";
 import { TaskContext } from "./context.js";
-import { asEnvelope, errorEnvelope, exceptionEnvelope, LostLease, SerializationError, } from "./errors.js";
+import { asEnvelope, errorEnvelope, EventLoopBlocked, exceptionEnvelope, LostLease, SerializationError, } from "./errors.js";
 import { newId } from "./ids.js";
 import { SQLiteStore } from "./store/sqlite.js";
 import { PostgresStore } from "./store/postgres.js";
@@ -703,6 +703,26 @@ export class Worker {
         let wake = null;
         // lease/3 gives two beats of slack; the floor only matters for sub-150ms leases.
         const interval = this.opts.heartbeatIntervalMs ?? Math.max(50, Math.floor(leaseMs / 3));
+        // When this heartbeat last ran. A loop cannot report its own absence — a
+        // handler that blocks for its whole attempt never lets the timer fire at all
+        // — so the check lives outside the loop and cancel() runs it too.
+        let lastBeatAt = Date.now();
+        /** Report a heartbeat that has not run for more than two intervals: a whole
+         * beat missed, which at lease/3 means the next such block loses the lease
+         * outright. Fires while the lease still holds — after it expires the only
+         * evidence is a task that ran twice, in two workers' logs, with no error in
+         * either. */
+        const checkBeat = () => {
+            const now = Date.now();
+            const lateMs = now - lastBeatAt - interval;
+            if (lateMs > interval) {
+                this.report(new EventLoopBlocked(lateMs, interval, leaseMs), {
+                    phase: "execute",
+                    taskId: ctxs[0].taskId,
+                });
+            }
+            lastBeatAt = now;
+        };
         const done = (async () => {
             while (active) {
                 // Cancellable sleep: cancel() resolves this immediately and clears the
@@ -717,6 +737,7 @@ export class Worker {
                 wake = null;
                 if (!active)
                     break;
+                checkBeat();
                 const live = ctxs.filter((c) => !c.settled && !c.lostLease);
                 if (!live.length)
                     break;
@@ -746,6 +767,10 @@ export class Worker {
                 active = false;
                 if (wake)
                     wake();
+                // The attempt is over, so this is the last chance to notice that the
+                // heartbeat never got one — the whole-attempt block, which is also the
+                // one where the lease is already gone.
+                checkBeat();
             },
             done,
         };

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "cairnq",
-  "version": "0.6.0",
+  "version": "0.7.0",
   "description": "SQLite-first, cross-language, storage-centered durable task runtime",
   "license": "MIT",
   "author": "Jannchie <jannchie@gmail.com>",

package/src/client.ts

@@ -1,11 +1,12 @@
 import type { BackpressureOptions } from "./backpressure.js";
+import { type RetentionOptions, RetentionSweeper } from "./retention.js";
 import { TaskCanceled, TaskFailed } from "./errors.js";
 import { isFailed, isSucceeded, type Task, type TaskStatus } from "./models.js";
 import { SQLiteStore } from "./store/sqlite.js";
 import { PostgresStore } from "./store/postgres.js";
 import type { ListInput, PurgeInput, SubmitInput, TaskStore } from "./store/base.js";
 import { type TaskDef, taskName } from "./task.js";
-import { pollWait } from "./wait.js";
+import { pollWait, pollWaitByKey } from "./wait.js";
 
 export type SubmitOptions = Omit<SubmitInput, "name" | "payload">;
 export interface CallOptions extends SubmitOptions {
@@ -15,10 +16,18 @@ export interface CallOptions extends SubmitOptions {
 
 /** Options this handle configures on the store it wraps, rather than the
  * store's own constructor arguments. */
-export type ClientOptions = Partial<BackpressureOptions>;
+export type ClientOptions = Partial<BackpressureOptions> & {
+  /** Delete terminal tasks older than a cutoff, on a schedule, for as long as
+   * this handle is open. Off unless set — and off means rows accumulate forever,
+   * because nothing else in CairnQ removes them. */
+  retention?: RetentionOptions;
+};
 
 /** API-side handle. Thin wrapper over a TaskStore + SDK-orchestrated wait/call. */
 export class CairnQ {
+  /** null unless `retention` was configured. */
+  private readonly sweeper: RetentionSweeper | null;
+
   constructor(
     private readonly _store: TaskStore,
     opts: ClientOptions = {},
@@ -28,6 +37,13 @@ export class CairnQ {
     if (opts.maxQueueDepth != null) {
       _store.useBackpressure(opts as BackpressureOptions);
     }
+    // Retention is the opposite case: it belongs to the handle, because a worker
+    // sharing the store must not also be deleting rows behind the API's back.
+    // Started here rather than in connect(), which is optional — every other
+    // path connects lazily, and retention that silently depends on an optional
+    // call is retention that silently does not happen.
+    this.sweeper = opts.retention ? new RetentionSweeper(_store, opts.retention) : null;
+    this.sweeper?.start();
   }
 
   static sqlite(path: string, opts: { busyTimeoutMs?: number } & ClientOptions = {}): CairnQ {
@@ -50,8 +66,11 @@ export class CairnQ {
     return this._store.connect();
   }
 
-  close(): Promise<void> {
-    return this._store.close();
+  /** Stop retention (waiting for a sweep in flight, so no purge outlives the
+   * store) and close the store. */
+  async close(): Promise<void> {
+    await this.sweeper?.stop();
+    await this._store.close();
   }
 
   /** Enqueue a task. With `maxQueueDepth` configured this blocks while the
@@ -114,6 +133,9 @@ export class CairnQ {
     return this._store.stats();
   }
 
+  /** Wait for a task to finish. Resolves with the terminal Task (any status);
+   * throws TaskTimeout without stopping the task, so `wait(err.taskId)` picks the
+   * same wait back up — from another process, or after a longer deadline. */
   wait(
     taskId: string,
     opts: { timeoutMs?: number; pollMs?: number } = {},
@@ -124,9 +146,25 @@ export class CairnQ {
     });
   }
 
+  /** Wait for whatever task the `key` currently points at — the cross-process
+   * form of picking a wait back up, when the id was never in hand or the process
+   * that held it is gone. Re-resolves the key on each poll, so a `replace`
+   * landing mid-wait moves the wait onto the new task, and a key with no task
+   * yet is waited for rather than rejected. */
+  waitByKey(key: string, opts: { timeoutMs?: number; pollMs?: number } = {}): Promise<Task> {
+    return pollWaitByKey(this._store, key, {
+      timeoutMs: opts.timeoutMs ?? 30_000,
+      pollMs: opts.pollMs,
+    });
+  }
+
   /** submit + wait. Resolves with the result on success; rejects with
    * TaskFailed / TaskCanceled / TaskTimeout otherwise. Pass a TaskDef and the
-   * resolved value is typed as its Result. */
+   * resolved value is typed as its Result.
+   *
+   * `waitTimeoutMs` bounds the wait, not the task: on timeout the task runs on,
+   * and `wait(err.taskId)` — or `waitByKey`, from a process that only has the
+   * key — resumes the wait rather than starting the work over. */
   async call(name: string, payload?: unknown, opts?: CallOptions): Promise<unknown>;
   async call<P, R>(task: TaskDef<P, R>, payload?: P, opts?: CallOptions): Promise<R>;
   async call(task: string | TaskDef, payload?: unknown, opts: CallOptions = {}): Promise<unknown> {

package/src/errors.ts

@@ -77,8 +77,12 @@ export class QueueFull extends CairnQError {
  * observed. No worker running, no handler for the name, wrong queue, and two
  * processes on different database files all look identical from the API side —
  * queued, never claimed — so that case names the likely causes. */
-function timeoutDetail(task: Task | null): string {
-  if (!task) return "task not found — wrong database file, or already purged?";
+function timeoutDetail(task: Task | null, key: string | null): string {
+  if (!task) {
+    return key === null
+      ? "task not found — wrong database file, or already purged?"
+      : "no task under this key — never submitted, or already purged?";
+  }
   if (isQueued(task)) {
     const delayMs = task.run_at_ms - nowMs();
     if (task.attempt === 0 && delayMs <= 0) {
@@ -94,23 +98,33 @@ function timeoutDetail(task: Task | null): string {
   return `still running (attempt ${task.attempt}/${task.max_attempts})`;
 }
 
-/** wait/call did not reach a terminal status in time. The task keeps running.
- * `task` is the last snapshot wait() observed (null if get() found nothing), and
- * the message says what state it was stuck in — a queued-never-claimed task is
- * the classic first-run failure (no worker, no handler, wrong queue or file). */
+/** wait/call did not reach a terminal status in time. The task keeps running, so
+ * `taskId` is the handle for picking the wait back up — `wait(err.taskId)`
+ * re-attaches to the same task from anywhere that can reach the store. `task` is
+ * the last snapshot wait() observed (null if the lookup found nothing), and the
+ * message says what state it was stuck in — a queued-never-claimed task is the
+ * classic first-run failure (no worker, no handler, wrong queue or file).
+ *
+ * `key` is set when the wait watched a key rather than an id; `taskId` is then
+ * the task the key pointed at, or the key itself when it pointed at nothing —
+ * there was no id to report. */
 export class TaskTimeout extends CairnQError {
   readonly task: Task | null;
+  readonly key: string | null;
   constructor(
     public taskId: string,
-    opts: { timeoutMs?: number; task?: Task | null } = {},
+    opts: { timeoutMs?: number; task?: Task | null; key?: string | null } = {},
   ) {
+    const key = opts.key ?? null;
+    const subject = key === null ? `task ${taskId}` : `key ${key}`;
     super(
       opts.timeoutMs == null
-        ? `task ${taskId} did not finish in time`
-        : `task ${taskId} did not finish within ${opts.timeoutMs}ms: ${timeoutDetail(opts.task ?? null)}`,
+        ? `${subject} did not finish in time`
+        : `${subject} did not finish within ${opts.timeoutMs}ms: ${timeoutDetail(opts.task ?? null, key)}`,
     );
     this.name = "TaskTimeout";
     this.task = opts.task ?? null;
+    this.key = key;
   }
 }
 
@@ -146,6 +160,36 @@ export class TaskCanceled extends CairnQError {
   }
 }
 
+/**
+ * A heartbeat beat came back later than its own interval allowed.
+ *
+ * The heartbeat shares the event loop with the handlers whose leases it renews,
+ * so a handler that blocks the loop stops the renewal with it: the lease expires,
+ * the task is recovered and redelivered, and a second worker starts computing
+ * what the first is still computing — one task, billed twice, with no error
+ * anywhere. Nothing inside the blocked handler can observe that, which is why it
+ * is reported through `onError` alongside the other things the run loop survived.
+ *
+ * The cause is always synchronous work in a handler: a tight loop, a large
+ * JSON.parse, a `*Sync` filesystem or crypto call. Node has one loop and no way
+ * to preempt it — move the work to a worker thread, a child process, or an async
+ * API that yields.
+ */
+export class EventLoopBlocked extends CairnQError {
+  constructor(
+    readonly lateMs: number,
+    readonly intervalMs: number,
+    readonly leaseMs: number,
+  ) {
+    super(
+      `heartbeat beat was ${lateMs}ms late (interval ${intervalMs}ms, lease ${leaseMs}ms): ` +
+        `the event loop was blocked long enough to miss a beat. Synchronous work in a ` +
+        `handler starves lease renewal — move it off the loop.`,
+    );
+    this.name = "EventLoopBlocked";
+  }
+}
+
 /** A worker write affected 0 rows: the lease expired and was reclaimed. */
 export class LostLease extends CairnQError {
   constructor(public taskId: string) {

package/src/index.ts

@@ -2,6 +2,8 @@ export { CairnQ } from "./client.js";
 export type { CallOptions, ClientOptions, SubmitOptions } from "./client.js";
 export { QueueDepthGate } from "./backpressure.js";
 export type { BackpressureOptions, QueueDepthLimit } from "./backpressure.js";
+export { RetentionSweeper } from "./retention.js";
+export type { RetentionOptions } from "./retention.js";
 export { Worker } from "./worker.js";
 export type { BatchHandler, Handler, TypedHandler, WorkerOptions } from "./worker.js";
 export { TaskContext } from "./context.js";
@@ -32,6 +34,7 @@ export {
   TaskCanceled,
   TaskError,
   LostLease,
+  EventLoopBlocked,
   ProtocolVersionMismatch,
   SerializationError,
 } from "./errors.js";

package/src/retention.ts

@@ -0,0 +1,136 @@
+import type { PurgeInput, TaskStore } from "./store/base.js";
+
+/** Sweep every hour unless asked otherwise — often enough that a queue with a
+ * day of retention never carries more than an hour of extra rows, rare enough
+ * that the sweep is invisible next to the task traffic. */
+const DEFAULT_INTERVAL_MS = 3_600_000;
+/** Rows per purge statement. The same bound `purge` defaults to: big enough that
+ * a backlog drains in few statements, small enough that each is a short write. */
+const DEFAULT_LIMIT = 1_000;
+
+export interface RetentionOptions {
+  /**
+   * How long a terminal task is kept after it finished. Required: there is no
+   * safe default for how long someone else's results stay readable.
+   */
+  olderThanMs: number;
+  /** Time between sweeps. Default 3_600_000 (one hour). */
+  intervalMs?: number;
+  /** Rows deleted per statement while draining. Default 1_000. */
+  limit?: number;
+  /**
+   * Called for a sweep that threw. The next sweep runs on schedule regardless —
+   * a purge that failed because the database was busy is not a reason to stop
+   * retaining — so without this a store quietly stops being swept. Must not throw.
+   */
+  onError?: (err: unknown) => void;
+}
+
+/**
+ * Deletes terminal tasks on a schedule, for as long as the handle is open.
+ *
+ * `purge` exists because nothing else in CairnQ removes rows, and a queue whose
+ * payloads carry real data — an image, a document, a batch of embeddings — turns
+ * that into a disk leak measured in gigabytes per backfill. Every deployment
+ * that runs longer than a demo needs the sweep; leaving it to an external
+ * scheduler means the leak is the default and remembering is the opt-in.
+ *
+ * It sweeps in bounded batches with a yield between them, so draining a backlog
+ * that accumulated while nothing was sweeping stays a sequence of short writes
+ * rather than one long one — on SQLite that matters, since a long write holds
+ * the single write lock against every producer and worker on the file.
+ */
+export class RetentionSweeper {
+  /** Whether the scheduled loop is running. */
+  private active = false;
+  /** Set by stop(), so a drain in progress can cut itself short too. */
+  private stopping = false;
+  /** Resolves the current sleep early, so stop() need not wait out an interval. */
+  private wake: (() => void) | null = null;
+  /** The loop itself, awaited by stop() so no purge outlives the store. */
+  private loop: Promise<void> | null = null;
+  private readonly intervalMs: number;
+  private readonly purgeInput: PurgeInput;
+
+  constructor(
+    private readonly store: TaskStore,
+    private readonly opts: RetentionOptions,
+  ) {
+    if (!Number.isFinite(opts.olderThanMs) || opts.olderThanMs < 0) {
+      throw new Error(`retention.olderThanMs must be >= 0, got ${opts.olderThanMs}`);
+    }
+    this.intervalMs = opts.intervalMs ?? DEFAULT_INTERVAL_MS;
+    if (!Number.isFinite(this.intervalMs) || this.intervalMs < 1) {
+      throw new Error(`retention.intervalMs must be >= 1, got ${this.intervalMs}`);
+    }
+    this.purgeInput = { olderThanMs: opts.olderThanMs, limit: opts.limit ?? DEFAULT_LIMIT };
+  }
+
+  start(): void {
+    if (this.active) return;
+    this.active = true;
+    this.stopping = false;
+    this.loop = this.run();
+  }
+
+  /** Stop sweeping and wait for the sweep in flight, if any. */
+  async stop(): Promise<void> {
+    this.stopping = true;
+    this.active = false;
+    this.wake?.();
+    await this.loop;
+    this.loop = null;
+  }
+
+  private async run(): Promise<void> {
+    // Sleep first: a process that restarts often would otherwise purge on every
+    // boot, which is a write burst exactly when the store is busiest.
+    while (!this.stopping) {
+      await this.sleep(this.intervalMs);
+      if (this.stopping) return;
+      try {
+        await this.sweep();
+      } catch (err) {
+        try {
+          this.opts.onError?.(err);
+        } catch {
+          // A reporting hook must never take the sweep down with it — the same
+          // rule the worker's onError follows.
+        }
+      }
+    }
+  }
+
+  /**
+   * Delete everything past the cutoff now, in bounded batches, and return how
+   * many rows went. The scheduled loop calls this; call it directly to drain on
+   * demand — after a backfill, or from a maintenance command.
+   */
+  async sweep(): Promise<number> {
+    const limit = this.purgeInput.limit as number;
+    let deleted = 0;
+    for (;;) {
+      const ids = await this.store.purge(this.purgeInput);
+      deleted += ids.length;
+      if (ids.length < limit || this.stopping) return deleted;
+      // Hand the loop back between batches: a large drain must not starve the
+      // submits and claims sharing this process.
+      await this.sleep(0);
+    }
+  }
+
+  /** Sleep, interruptible by stop(). Unref'd: retention is housekeeping, and a
+   * pending sweep must never be the reason a process refuses to exit. */
+  private sleep(ms: number): Promise<void> {
+    return new Promise<void>((resolve) => {
+      const timer = setTimeout(resolve, ms);
+      timer.unref?.();
+      this.wake = () => {
+        clearTimeout(timer);
+        resolve();
+      };
+    }).finally(() => {
+      this.wake = null;
+    });
+  }
+}

package/src/store/base.ts

@@ -6,7 +6,7 @@ import {
   ProtocolVersionMismatch,
   SerializationError,
 } from "../errors.js";
-import { rowToTask, STATUSES, type Task, type TaskStatus } from "../models.js";
+import { rowToTask, STATUSES, TERMINAL, type Task, type TaskStatus } from "../models.js";
 import { type BackpressureOptions, QueueDepthGate } from "../backpressure.js";
 
 const rejectMangled = function (this: unknown, _key: string, v: unknown): unknown {
@@ -62,9 +62,25 @@ export function checkProtocolVersion(version: number): void {
 // CONFLICTS is the canonical declaration; the type derives from it so the
 // runtime guard in submit() and the type can't drift apart (same pattern as
 // STATUSES/TaskStatus in models.ts).
-const CONFLICTS = ["reuse", "reject", "replace"] as const;
+const CONFLICTS = ["reuse", "reuse-succeeded", "reject", "replace"] as const;
 export type Conflict = (typeof CONFLICTS)[number];
 
+/**
+ * Whether a keyed submit's strategy accepts the task the key already points at.
+ *
+ * Both reuse strategies deduplicate work that is still in play — that is what a
+ * key is for, and the answer cannot depend on the outcome of a task that has no
+ * outcome yet. They differ only on what a *finished* task means: `reuse` treats
+ * the key as free again, while `reuse-succeeded` reads a succeeded task as a
+ * cached result. Neither ever hands back a failed or canceled one, which would
+ * poison the key for every later submit (see PROTOCOL.md "Key conflict").
+ */
+function reusable(conflict: Conflict, status: TaskStatus): boolean {
+  if (conflict === "replace") return false;
+  if (!TERMINAL.includes(status)) return true;
+  return conflict === "reuse-succeeded" && status === "succeeded";
+}
+
 /** The queue a submit lands on when it names none. Owned here, where the
  * default is applied, so nothing above has to re-derive it. */
 export const DEFAULT_QUEUE = "default";
@@ -283,10 +299,15 @@ export abstract class TaskStore {
         // free after all, whatever the strategy.
         const current = (await fetch("get", { id: existing[0].task_id }))[0];
         if (current) {
-          if (conflict === "reuse") return rowToTask(current);
           if (conflict === "reject") throw new AlreadyExists(key);
-          // "replace": cancel the recorded task, then repoint the key below.
-          await fetch("cancel", { id: existing[0].task_id });
+          if (reusable(conflict, current.status as TaskStatus)) return rowToTask(current);
+          // The strategy declined the recorded task, so the key repoints to the
+          // fresh one inserted below. Cancel only what is still live: a terminal
+          // task has nothing to stop, and cancelling it would rewrite a settled
+          // row (and hand a `canceled` back to whoever is waiting on it).
+          if (!TERMINAL.includes(current.status as TaskStatus)) {
+            await fetch("cancel", { id: existing[0].task_id });
+          }
         }
       }
       const row = (await fetch("insert_task", ins))[0];

package/src/wait.ts

@@ -7,6 +7,14 @@ export const DEFAULT_POLL_MS = 100;
 export const MAX_POLL_MS = 500;
 const GROWTH = 1.5;
 
+const sleep = (ms: number) => new Promise<void>((resolve) => setTimeout(resolve, ms));
+
+export interface PollOptions {
+  timeoutMs: number;
+  pollMs?: number;
+  maxPollMs?: number;
+}
+
 /**
  * Grow the polling interval towards the ceiling.
  *
@@ -19,28 +27,64 @@ export function nextPollMs(current: number, maxMs: number): number {
   return Math.min(maxMs, Math.max(current + 1, Math.floor(current * GROWTH)));
 }
 
-/** Poll get() until terminal or timeout. Returns the terminal Task (any status).
- * Throws TaskTimeout, leaving the task running. `pollMs` is the *first* interval;
- * it backs off towards `maxPollMs`. */
-export async function pollWait(
-  store: TaskStore,
-  taskId: string,
-  {
-    timeoutMs,
-    pollMs = DEFAULT_POLL_MS,
-    maxPollMs = MAX_POLL_MS,
-  }: { timeoutMs: number; pollMs?: number; maxPollMs?: number },
+/**
+ * Poll `read` until it yields a terminal task, or the timeout elapses.
+ *
+ * `wake` is what the loop sleeps on between reads: a store with a push channel
+ * (Postgres) cuts it short when the task goes terminal, but the re-read is the
+ * source of truth either way, so a plain sleep is always a correct answer.
+ */
+async function poll(
+  read: () => Promise<Task | null>,
+  wake: (task: Task | null, ms: number) => Promise<void>,
+  subject: string,
+  key: string | null,
+  { timeoutMs, pollMs = DEFAULT_POLL_MS, maxPollMs = MAX_POLL_MS }: PollOptions,
 ): Promise<Task> {
   const deadline = nowMs() + timeoutMs;
   let interval = pollMs;
   for (;;) {
-    const task = await store.get(taskId);
+    const task = await read();
     if (task && isTerminal(task)) return task;
     const remaining = deadline - nowMs();
-    if (remaining <= 0) throw new TaskTimeout(taskId, { timeoutMs, task });
-    // A store with a push channel (Postgres) cuts the sleep short when the task
-    // goes terminal; the re-get above stays the source of truth either way.
-    await store.taskDoneWake(taskId, Math.min(interval, remaining));
+    if (remaining <= 0) throw new TaskTimeout(task?.id ?? subject, { timeoutMs, task, key });
+    await wake(task, Math.min(interval, remaining));
     interval = nextPollMs(interval, maxPollMs);
   }
 }
+
+/** Poll get() until terminal or timeout. Returns the terminal Task (any status).
+ * Throws TaskTimeout, leaving the task running. `pollMs` is the *first* interval;
+ * it backs off towards `maxPollMs`. */
+export function pollWait(store: TaskStore, taskId: string, opts: PollOptions): Promise<Task> {
+  return poll(
+    () => store.get(taskId),
+    (_task, ms) => store.taskDoneWake(taskId, ms),
+    taskId,
+    null,
+    opts,
+  );
+}
+
+/**
+ * The same wait, following a key instead of an id.
+ *
+ * The key is re-resolved on every read, because that is what a key means: a
+ * pointer to the task that is *current* under it. A `replace` landing mid-wait
+ * moves the wait onto the new task rather than reporting the cancellation of the
+ * old one, and a key that points at nothing yet is simply not finished — it
+ * polls until something appears, the same way waiting on an id that does not
+ * exist yet does.
+ *
+ * There is nothing to subscribe to before the key resolves, so those naps are
+ * plain sleeps; once it resolves, the store's push channel applies as usual.
+ */
+export function pollWaitByKey(store: TaskStore, key: string, opts: PollOptions): Promise<Task> {
+  return poll(
+    () => store.getByKey(key),
+    (task, ms) => (task ? store.taskDoneWake(task.id, ms) : sleep(ms)),
+    key,
+    key,
+    opts,
+  );
+}