cairnq 0.4.0 → 0.6.0

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "cairnq",
-  "version": "0.4.0",
+  "version": "0.6.0",
   "description": "SQLite-first, cross-language, storage-centered durable task runtime",
   "license": "MIT",
   "author": "Jannchie <jannchie@gmail.com>",
@@ -34,6 +34,7 @@
   "engines": { "node": ">=22" },
   "scripts": {
     "bench": "tsx bench/run.ts",
+    "bench:sweep": "tsx bench/sweep.ts",
     "build": "tsc -p tsconfig.json",
     "test": "vitest run",
     "typecheck": "tsc -p tsconfig.json --noEmit"

package/src/backoff.ts

@@ -0,0 +1,53 @@
+/**
+ * Retry backoff, in its own module because two callers need it.
+ *
+ * The worker computes it when a handler's failure ends an attempt; TaskContext
+ * computes it when a handler fails one task of a batch itself. Keeping it in
+ * worker.ts would make context.ts import the module that imports it.
+ */
+
+export const DEFAULT_RETRY_BACKOFF_MS = 1_000;
+export const DEFAULT_RETRY_BACKOFF_MAX_MS = 30_000;
+
+/**
+ * Exponential backoff with equal jitter: the window doubles per attempt up to
+ * `maxMs`, and the delay lands uniformly in its upper half, `[w/2, w)`.
+ *
+ * The jitter is what keeps a fleet from retrying in lockstep. Failures align
+ * when the downstream fails fast enough that a whole concurrency batch raises
+ * at once (connection refused, DNS gone), and capped exponential backoff then
+ * *preserves* that alignment — once every task sits at `maxMs`, they all retry
+ * on the same beat forever. Spreading over half the window breaks it; keeping
+ * the lower half as a floor means jitter never shortens the wait to less than
+ * half of what plain exponential backoff would have asked for.
+ *
+ * `rand` is injected so tests can pin an exact delay.
+ */
+export function retryDelayMs(
+  attempt: number,
+  baseMs: number,
+  maxMs: number,
+  rand: () => number = Math.random,
+): number {
+  if (baseMs <= 0) return 0;
+  const exponent = Math.max(0, attempt - 1);
+  const window = Math.min(maxMs, baseMs * 2 ** exponent);
+  const floor = Math.floor(window / 2);
+  return floor + Math.floor(rand() * (window - floor));
+}
+
+/**
+ * The delay a `fail` write should carry. Not just the backoff: a permanent
+ * failure is never re-run, so it always delays 0. Both settlement paths — the
+ * worker's and a handler's `ctx.fail` — go through this, so they cannot end up
+ * backing off differently.
+ */
+export function failDelayMs(
+  attempt: number,
+  retryable: boolean,
+  baseMs: number,
+  maxMs: number,
+  rand: () => number = Math.random,
+): number {
+  return retryable ? retryDelayMs(attempt, baseMs, maxMs, rand) : 0;
+}

package/src/backpressure.ts

@@ -0,0 +1,140 @@
+import { setTimeout as delay } from "node:timers/promises";
+import { QueueFull } from "./errors.js";
+import type { TaskStore } from "./store/base.js";
+
+/**
+ * Most tasks a producer may enqueue on one probe's word.
+ *
+ * The gate probes only when its headroom runs out, so this is what the check
+ * costs amortized: one bounded index read per MAX_GRANT submits. It also bounds
+ * how far the limit can be overshot — see the class docstring on why several
+ * producers make this a soft limit, and why that overshoot is (N-1) * MAX_GRANT
+ * rather than unbounded.
+ */
+const MAX_GRANT = 64;
+
+// Named for probing, not polling: wait.ts exports DEFAULT_POLL_MS / MAX_POLL_MS
+// for the get() loop behind wait(), an order of magnitude tighter and answering
+// a different question. Two constants of the same name in one SDK would be read
+// as one policy.
+const INITIAL_PROBE_INTERVAL_MS = 250;
+const MAX_PROBE_INTERVAL_MS = 5_000;
+const DEFAULT_MAX_WAIT_MS = 600_000;
+
+/** Per-queue depth limits. A number applies one limit to every queue; a record
+ * gates only the queues it names and leaves the rest unbounded. */
+export type QueueDepthLimit = number | Record<string, number>;
+
+export interface BackpressureOptions {
+  /** Queued tasks a queue may hold before `submit` blocks. */
+  maxQueueDepth: QueueDepthLimit;
+  /** How long a blocked submit waits before raising QueueFull. Default 600_000. */
+  maxQueueWaitMs?: number;
+  /** First backoff between depth probes; doubles to a 5s ceiling. Default 250. */
+  queuePollIntervalMs?: number;
+}
+
+/**
+ * Blocks `submit` while a queue is at its depth limit.
+ *
+ * Without one of these a producer that outruns its workers is only bounded by
+ * disk: the backlog grows, every task's queue wait grows with it, and the
+ * failure is a database that filled up rather than a producer that slowed down.
+ * A queue is the wrong place to buffer an overload — pushing back on the
+ * producer is the point.
+ *
+ * **A soft limit under several producers.** The check is a read followed by a
+ * write that other producers can interleave with, and each holds its own grant,
+ * so N producers can overshoot the limit by up to (N-1) * MAX_GRANT tasks. Made
+ * exact it would need the depth check inside insert_task's transaction, which
+ * puts an unbounded-scan predicate on the hot path of every submit and turns
+ * concurrent submits into lock contention — a steep price for a bound whose
+ * whole purpose is approximate. Size the limit for the pushback you want, not as
+ * a capacity assertion.
+ */
+export class QueueDepthGate {
+  /** Remaining grant per queue: submits allowed before the next probe. */
+  private readonly headroom = new Map<string, number>();
+  /** In-flight probe per queue, so concurrent submits share one read rather
+   * than each issuing their own against a queue that is already known full. */
+  private readonly probing = new Map<string, Promise<void>>();
+  private readonly limits: QueueDepthLimit;
+  private readonly maxWaitMs: number;
+  private readonly initialProbeMs: number;
+
+  constructor(
+    private readonly store: TaskStore,
+    opts: BackpressureOptions,
+  ) {
+    this.limits = opts.maxQueueDepth;
+    this.maxWaitMs = opts.maxQueueWaitMs ?? DEFAULT_MAX_WAIT_MS;
+    this.initialProbeMs = opts.queuePollIntervalMs ?? INITIAL_PROBE_INTERVAL_MS;
+    if (typeof this.limits === "number") this.validate("*", this.limits);
+    else for (const [q, v] of Object.entries(this.limits)) this.validate(q, v);
+  }
+
+  private validate(queue: string, limit: number): void {
+    // A limit of 0 would block every submit forever, which is never what a
+    // caller means; catching it here beats a first submit that hangs for
+    // maxQueueWaitMs and then raises.
+    if (!Number.isInteger(limit) || limit < 1) {
+      throw new Error(`maxQueueDepth for ${queue} must be an integer >= 1, got ${limit}`);
+    }
+  }
+
+  /** The limit for `queue`, or null when it is not gated. */
+  limitFor(queue: string): number | null {
+    if (typeof this.limits === "number") return this.limits;
+    return this.limits[queue] ?? null;
+  }
+
+  /**
+   * Consume one unit of headroom for `queue`, waiting for room if it is full.
+   * Returns immediately for an ungated queue. Raises QueueFull on timeout,
+   * having enqueued nothing.
+   */
+  async acquire(queue: string): Promise<void> {
+    const limit = this.limitFor(queue);
+    if (limit == null) return;
+
+    const startedAt = Date.now();
+    let waitMs = this.initialProbeMs;
+    for (;;) {
+      const left = this.headroom.get(queue) ?? 0;
+      if (left > 0) {
+        this.headroom.set(queue, left - 1);
+        return;
+      }
+      await this.probe(queue, limit);
+      if ((this.headroom.get(queue) ?? 0) > 0) continue;
+
+      const waited = Date.now() - startedAt;
+      if (waited >= this.maxWaitMs) throw new QueueFull(queue, limit, waited);
+      // Back off: a queue at its limit will not drain within one poll interval,
+      // and re-probing tightly adds read load to a database already behind.
+      await delay(Math.min(waitMs, this.maxWaitMs - waited));
+      waitMs = Math.min(waitMs * 2, MAX_PROBE_INTERVAL_MS);
+    }
+  }
+
+  /**
+   * Refresh `queue`'s grant from the store, at most one probe in flight.
+   *
+   * Callers re-read `headroom` afterwards rather than using a returned value:
+   * only the caller that started the probe writes the grant, so waiters that
+   * joined it cannot overwrite the units already handed out.
+   */
+  private probe(queue: string, limit: number): Promise<void> {
+    let p = this.probing.get(queue);
+    if (!p) {
+      p = this.store
+        .queueDepth(queue, limit)
+        .then((headroom) => {
+          this.headroom.set(queue, Math.min(headroom, MAX_GRANT));
+        })
+        .finally(() => this.probing.delete(queue));
+      this.probing.set(queue, p);
+    }
+    return p;
+  }
+}

package/src/client.ts

@@ -1,3 +1,4 @@
+import type { BackpressureOptions } from "./backpressure.js";
 import { TaskCanceled, TaskFailed } from "./errors.js";
 import { isFailed, isSucceeded, type Task, type TaskStatus } from "./models.js";
 import { SQLiteStore } from "./store/sqlite.js";
@@ -12,18 +13,33 @@ export interface CallOptions extends SubmitOptions {
   pollMs?: number;
 }
 
+/** Options this handle configures on the store it wraps, rather than the
+ * store's own constructor arguments. */
+export type ClientOptions = Partial<BackpressureOptions>;
+
 /** API-side handle. Thin wrapper over a TaskStore + SDK-orchestrated wait/call. */
 export class CairnQ {
-  constructor(private readonly _store: TaskStore) {}
+  constructor(
+    private readonly _store: TaskStore,
+    opts: ClientOptions = {},
+  ) {
+    // Installed on the store, not held here: every submit path goes through the
+    // store, including TaskContext.submit, which this handle never sees.
+    if (opts.maxQueueDepth != null) {
+      _store.useBackpressure(opts as BackpressureOptions);
+    }
+  }
 
-  static sqlite(path: string, opts?: { busyTimeoutMs?: number }): CairnQ {
-    return new CairnQ(new SQLiteStore(path, opts));
+  static sqlite(path: string, opts: { busyTimeoutMs?: number } & ClientOptions = {}): CairnQ {
+    const { busyTimeoutMs, ...client } = opts;
+    return new CairnQ(new SQLiteStore(path, { busyTimeoutMs }), client);
   }
 
   /** Multi-host backend. `dsn` is a libpq connection string; requires the
    * optional `pg` package. */
-  static postgres(dsn: string, opts?: { max?: number }): CairnQ {
-    return new CairnQ(new PostgresStore(dsn, opts));
+  static postgres(dsn: string, opts: { max?: number } & ClientOptions = {}): CairnQ {
+    const { max, ...client } = opts;
+    return new CairnQ(new PostgresStore(dsn, { max }), client);
   }
 
   get store(): TaskStore {
@@ -38,12 +54,24 @@ export class CairnQ {
     return this._store.close();
   }
 
+  /** Enqueue a task. With `maxQueueDepth` configured this blocks while the
+   * target queue is at its limit, and raises QueueFull if it stays there for
+   * `maxQueueWaitMs` — see QueueDepthGate for why that bound is approximate
+   * across several producers. */
   submit(name: string, payload?: unknown, opts?: SubmitOptions): Promise<Task>;
   submit<P, R>(task: TaskDef<P, R>, payload?: P, opts?: SubmitOptions): Promise<Task>;
   submit(task: string | TaskDef, payload?: unknown, opts: SubmitOptions = {}): Promise<Task> {
     return this._store.submit({ name: taskName(task), payload, ...opts });
   }
 
+  /** How many more tasks fit on `queue` under `maxDepth` — 0 once it is full.
+   * The non-blocking read behind `maxQueueDepth`, for a producer that would
+   * rather shed load or pick another queue than wait. Cheaper than `stats()`:
+   * bounded at `maxDepth` index entries instead of aggregating the table. */
+  queueDepth(queue: string, maxDepth: number): Promise<number> {
+    return this._store.queueDepth(queue, maxDepth);
+  }
+
   get(taskId: string): Promise<Task | null> {
     return this._store.get(taskId);
   }

package/src/context.ts

@@ -1,24 +1,48 @@
-import { LostLease } from "./errors.js";
+import { DEFAULT_RETRY_BACKOFF_MAX_MS, DEFAULT_RETRY_BACKOFF_MS, failDelayMs } from "./backoff.js";
+import { asEnvelope, type FailReason, LostLease } from "./errors.js";
 import { cancelRequested, type Task } from "./models.js";
 import type { SubmitOptions } from "./client.js";
 import type { TaskStore } from "./store/base.js";
 import { type TaskDef, taskName } from "./task.js";
 import { pollWait } from "./wait.js";
 
-/** Handed to a task handler. Worker-side capabilities mirror the Python SDK. */
+export interface TaskContextOptions {
+  retryBackoffMs?: number;
+  retryBackoffMaxMs?: number;
+}
+
+/**
+ * Handed to a task handler. Worker-side capabilities mirror the Python SDK.
+ *
+ * One of these per task, whether a handler is delivered one task or a batch: a
+ * batch handler receives a `TaskContext[]`, so a single-task handler's `ctx` is
+ * literally the batch-of-one element. Lease, cancellation and settlement are per
+ * task, which is why they live here rather than on anything batch-shaped.
+ */
 export class TaskContext {
   private readonly abort = new AbortController();
   private leaseLost = false;
   // Cancellation is monotonic: once the DB has told us a cancel was requested it
   // can't be taken back, so canceled() can answer from this without a re-read.
   private cancelSeen = false;
+  // Set once this task reached a terminal state through succeed()/fail(). The
+  // worker reads it to know which tasks a batch handler already decided, so it
+  // neither settles them twice nor keeps renewing their leases — the bookkeeping
+  // every ack/nack-style handler otherwise has to carry itself.
+  private isSettled = false;
+  private readonly backoffMs: number;
+  private readonly backoffMaxMs: number;
 
   constructor(
     private readonly store: TaskStore,
     private readonly task: Task,
     public readonly workerId: string,
     private readonly leaseMs: number,
-  ) {}
+    opts: TaskContextOptions = {},
+  ) {
+    this.backoffMs = opts.retryBackoffMs ?? DEFAULT_RETRY_BACKOFF_MS;
+    this.backoffMaxMs = opts.retryBackoffMaxMs ?? DEFAULT_RETRY_BACKOFF_MAX_MS;
+  }
 
   get taskId(): string {
     return this.task.id;
@@ -44,6 +68,19 @@ export class TaskContext {
   get payload(): any {
     return this.task.payload;
   }
+  /**
+   * True once this task reached a terminal state — whether the handler settled
+   * it with succeed()/fail() or the worker settled it on the handler's behalf.
+   * The heartbeat and the settlement paths both read it.
+   */
+  get settled(): boolean {
+    return this.isSettled;
+  }
+
+  /** @internal Called by the worker when it finalizes this task itself. */
+  markSettled(): void {
+    this.isSettled = true;
+  }
 
   /**
    * True once this worker has lost the task's lease — it expired and another
@@ -70,17 +107,36 @@ export class TaskContext {
   // Every owned write returns the current row, so cancellation and lease loss
   // ride along on writes the handler was making anyway.
   private observe(task: Task): Task {
-    if (cancelRequested(task)) this.cancelSeen = true;
+    this.observeCancel(cancelRequested(task));
     return task;
   }
 
+  /**
+   * @internal The same observation from just the flag, for a caller that read it
+   * without materializing a Task — the shared heartbeat, whose statement returns
+   * only the id and the cancel column precisely so it does not have to drag
+   * every payload back on every beat.
+   */
+  observeCancel(cancelRequested: boolean): void {
+    if (cancelRequested) this.cancelSeen = true;
+  }
+
   private async owned(write: () => Promise<Task>): Promise<Task> {
-    // Short-circuit once the lease is known lost: nothing this context writes
-    // may be recorded any more. Locally, not just via the store's ownership
-    // check — after an abandoned (timed-out) attempt the same worker may
-    // re-claim this task under the same workerId, and a zombie handler's write
-    // would then pass ownership against the NEW attempt.
+    // One gate for every write through this context, so "may I still write?" is
+    // answered in one place rather than at each call site.
+    //
+    // Lease lost: nothing this context writes may be recorded any more. Checked
+    // locally, not just via the store's ownership check — after an abandoned
+    // (timed-out) attempt the same worker may re-claim this task under the same
+    // workerId, and a zombie handler's write would then pass ownership against
+    // the NEW attempt.
     if (this.leaseLost) throw new LostLease(this.task.id);
+    // Settled: the task is terminal, so the statement would match no row and come
+    // back as a lost lease — telling the handler "another worker took this" when
+    // the truth is "you already finished it", and flipping lostLease on the way.
+    // Refuse here instead, without the round trip and without corrupting the
+    // lease state.
+    if (this.isSettled) throw new LostLease(this.task.id);
     try {
       return this.observe(await write());
     } catch (err) {
@@ -119,6 +175,57 @@ export class TaskContext {
     return this.cancelSeen || t.status === "canceled";
   }
 
+  // ------------------------------------------------------------- settlement
+  // Finalizing a task is normally the worker's job, decided by whether the
+  // handler returned or threw. These two let a handler decide one task itself,
+  // which is what a batch needs: four of 256 tasks failing for four different
+  // reasons is the ordinary case, not the edge one, and it cannot be expressed
+  // by a single return value or a single throw.
+  //
+  // Settling twice is a no-op rather than an error. Handlers built on ack/nack
+  // queues all end up carrying a `finalizedIds` set to guarantee exactly that;
+  // holding it here instead is the point.
+
+  /**
+   * Finalize this task as succeeded, now, without waiting for the handler to
+   * return. `complete` semantics: a cancel requested while it ran wins and the
+   * task finalizes as canceled instead, its result discarded. Returns null if
+   * this task was already settled.
+   */
+  async succeed(result: unknown = null): Promise<Task | null> {
+    if (this.isSettled) return null;
+    const task = await this.owned(() =>
+      this.store.complete({ taskId: this.task.id, workerId: this.workerId, result }),
+    );
+    this.markSettled();
+    return task;
+  }
+
+  /**
+   * Finalize this task as failed, now. `error` may be a string reason, an Error,
+   * a TaskError (which carries its own retryability), or a ready envelope.
+   * Retryable failures get the worker's backoff and are re-queued while attempts
+   * remain, exactly as a thrown error would be. Returns null if already settled.
+   */
+  async fail(
+    error: FailReason = "task failed",
+    opts: { retryable?: boolean } = {},
+  ): Promise<Task | null> {
+    if (this.isSettled) return null;
+    const [envelope, retryable] = asEnvelope(error, opts.retryable ?? true);
+    const task = await this.owned(() =>
+      this.store.fail({
+        taskId: this.task.id,
+        workerId: this.workerId,
+        error: envelope,
+        retryable,
+        delayMs: failDelayMs(this.task.attempt, retryable, this.backoffMs, this.backoffMaxMs),
+      }),
+    );
+    this.markSettled();
+    return task;
+  }
+
   /** Submit a child task; parent/root/correlation are wired automatically. */
   submit(name: string, payload?: unknown, opts?: SubmitOptions): Promise<Task>;
   submit<P, R>(task: TaskDef<P, R>, payload?: P, opts?: SubmitOptions): Promise<Task>;

package/src/errors.ts

@@ -20,6 +20,24 @@ export function errorEnvelope(e: {
   };
 }
 
+/**
+ * How an arbitrary thrown value becomes an envelope. Split out from `asEnvelope`
+ * below because the worker also reaches it directly, for a thrown plain object —
+ * which `asEnvelope` reads as a ready envelope, the right call for `ctx.fail` and
+ * the wrong one for something that was thrown. Both must agree on `code` and on
+ * deriving `type` from the error's name, or the same error reads differently
+ * depending on which way it was recorded.
+ */
+export function exceptionEnvelope(err: unknown, retryable = true): Record<string, unknown> {
+  const e = err as { name?: string; message?: string };
+  return errorEnvelope({
+    type: e?.name ?? "Error",
+    code: "handler_error",
+    message: String(e?.message ?? err),
+    retryable,
+  });
+}
+
 export class CairnQError extends Error {
   constructor(message?: string) {
     super(message);
@@ -37,6 +55,24 @@ export class AlreadyExists extends CairnQError {
   }
 }
 
+/** A gated submit waited out `maxWaitMs` without the queue draining below its
+ * depth limit. Nothing was enqueued. Distinct from a slow submit on purpose: a
+ * queue this far behind is a capacity problem, and a caller that silently
+ * retries forever converts it into an invisible one. */
+export class QueueFull extends CairnQError {
+  constructor(
+    public queue: string,
+    public maxDepth: number,
+    public waitedMs: number,
+  ) {
+    super(
+      `queue ${queue} still holds ${maxDepth} or more queued tasks after ` +
+        `${waitedMs}ms; refusing to enqueue more`,
+    );
+    this.name = "QueueFull";
+  }
+}
+
 /** One line of "why hasn't this finished" from the last snapshot wait()
  * 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 —
@@ -165,3 +201,33 @@ export class TaskError extends CairnQError {
     });
   }
 }
+
+/** What a handler may pass to `ctx.fail`. */
+export type FailReason = string | Error | TaskError | Record<string, unknown>;
+
+/**
+ * Normalize anything that can end a task into [envelope, retryable].
+ *
+ * Shared by both ways a failure is recorded — a handler passing a reason to
+ * `ctx.fail`, and the worker classifying an error that ended an attempt — so the
+ * two cannot disagree about what a given error means. It lives here, beside the
+ * envelope constructors it dispatches to, rather than in the module that happens
+ * to expose it to handlers.
+ *
+ * A handler failing one task of a batch has a reason, not an exception object:
+ * `item.fail("no source records", { retryable: false })` is the shape the real
+ * code wants. A TaskError carries its own retryability and wins over the option;
+ * everything else takes the caller's. A ready envelope passes through, which is
+ * how the worker hands in the ones it composes itself.
+ */
+export function asEnvelope(
+  error: FailReason,
+  retryable: boolean,
+): [Record<string, unknown>, boolean] {
+  if (error instanceof TaskError) return [error.envelope(), error.retryable];
+  if (error instanceof Error) return [exceptionEnvelope(error, retryable), retryable];
+  if (typeof error === "object" && error !== null) return [error, retryable];
+  // A bare reason is a TaskError in everything but the throwing, so let
+  // TaskError own its own type/code defaults rather than restating them.
+  return [new TaskError(String(error), { retryable }).envelope(), retryable];
+}

package/src/index.ts

@@ -1,8 +1,11 @@
 export { CairnQ } from "./client.js";
-export type { CallOptions, SubmitOptions } from "./client.js";
+export type { CallOptions, ClientOptions, SubmitOptions } from "./client.js";
+export { QueueDepthGate } from "./backpressure.js";
+export type { BackpressureOptions, QueueDepthLimit } from "./backpressure.js";
 export { Worker } from "./worker.js";
-export type { Handler, TypedHandler, WorkerOptions } from "./worker.js";
+export type { BatchHandler, Handler, TypedHandler, WorkerOptions } from "./worker.js";
 export { TaskContext } from "./context.js";
+export type { TaskContextOptions } from "./context.js";
 export { defineTask } from "./task.js";
 export type { TaskDef } from "./task.js";
 export { SQLiteStore } from "./store/sqlite.js";
@@ -23,6 +26,7 @@ export {
 export {
   CairnQError,
   AlreadyExists,
+  QueueFull,
   TaskTimeout,
   TaskFailed,
   TaskCanceled,
@@ -31,3 +35,4 @@ export {
   ProtocolVersionMismatch,
   SerializationError,
 } from "./errors.js";
+export type { FailReason } from "./errors.js";