@crawlee/core 4.0.0-beta.88 → 4.0.0-beta.89

package/autoscaling/autoscaled_pool.js

@@ -3,28 +3,21 @@ import { addTimeoutToPromise } from '@apify/timeout';
 import { betterClearInterval, betterSetInterval } from '@apify/utilities';
 import { CriticalError } from '../errors.js';
 import { serviceLocator } from '../service_locator.js';
-import { Snapshotter } from './snapshotter.js';
-import { SystemStatus } from './system_status.js';
 /**
  * Manages a pool of asynchronous resource-intensive tasks that are executed in parallel.
- * The pool only starts new tasks if there is enough free CPU and memory available
- * and the Javascript event loop is not blocked.
- *
- * The information about the CPU and memory usage is obtained by the {@link Snapshotter} class,
- * which makes regular snapshots of system resources that may be either local
- * or from the Apify cloud infrastructure in case the process is running on the Apify platform.
- * Meaningful data gathered from these snapshots is provided to `AutoscaledPool` by the {@link SystemStatus} class.
+ * The pool only starts new tasks while its {@link IConcurrencySystem|concurrency system} reports free capacity —
+ * that governor is what monitors CPU, memory and event loop load and autoscales the concurrency budget.
  *
  * Before running the pool, you need to implement the following three functions:
- * {@link AutoscaledPoolOptions.runTaskFunction},
- * {@link AutoscaledPoolOptions.isTaskReadyFunction} and
- * {@link AutoscaledPoolOptions.isFinishedFunction}.
+ * {@link AutoscaledPoolOptions.runTaskFunction|`runTaskFunction`},
+ * {@link AutoscaledPoolPredicateOptions.isTaskReadyFunction|`isTaskReadyFunction`} and
+ * {@link AutoscaledPoolPredicateOptions.isFinishedFunction|`isFinishedFunction`}.
  *
  * The auto-scaled pool is started by calling the {@link AutoscaledPool.run} function.
- * The pool periodically queries the {@link AutoscaledPoolOptions.isTaskReadyFunction} function
- * for more tasks, managing optimal concurrency, until the function resolves to `false`. The pool then queries
- * the {@link AutoscaledPoolOptions.isFinishedFunction}. If it resolves to `true`, the run finishes after all running tasks complete.
- * If it resolves to `false`, it assumes there will be more tasks available later and keeps periodically querying for tasks.
+ * The pool periodically queries `isTaskReadyFunction` for more tasks, managing optimal concurrency, until the function
+ * resolves to `false`. The pool then queries `isFinishedFunction`. If it resolves to `true`, the run finishes after all
+ * running tasks complete. If it resolves to `false`, it assumes there will be more tasks available later and keeps
+ * periodically querying for tasks.
  * If any of the tasks throws then the {@link AutoscaledPool.run} function rejects the promise with an error.
  *
  * The pool evaluates whether it should start a new task every time one of the tasks finishes
@@ -33,8 +26,12 @@ import { SystemStatus } from './system_status.js';
  * **Example usage:**
  *
  * ```javascript
+ * const concurrencySystem = new ConcurrencySystem({ maxConcurrency: 50 });
+ * await concurrencySystem.start();
+ *
  * const pool = new AutoscaledPool({
- *     maxConcurrency: 50,
+ *     concurrencySystem,
+ *     consumer: { id: 'my-pool' },
  *     runTaskFunction: async () => {
  *         // Run some resource-intensive asynchronous operation here.
  *     },
@@ -49,169 +46,108 @@ import { SystemStatus } from './system_status.js';
  *     }
  * });
  *
- * await pool.run();
+ * try {
+ *     await pool.run();
+ * } finally {
+ *     await concurrencySystem.stop();
+ * }
  * ```
  * @category Scaling
  */
 export class AutoscaledPool {
     log;
     // Configurable properties.
-    desiredConcurrencyRatio;
-    scaleUpStepRatio;
-    scaleDownStepRatio;
     maybeRunIntervalMillis;
-    loggingIntervalMillis;
-    autoscaleIntervalMillis;
     taskTimeoutMillis;
     runTaskFunction;
     isFinishedFunction;
     isTaskReadyFunction;
-    maxTasksPerMinute;
+    concurrencySystem;
+    consumer;
     // Internal properties.
-    _minConcurrency;
-    _maxConcurrency;
-    _desiredConcurrency;
-    _currentConcurrency = 0;
     isStopped = false;
-    lastLoggingTime;
     resolve = null;
     reject = null;
-    snapshotter;
-    /** Additional SystemStatus loadSignals - tracked here for initialization and cleanup */
-    loadSignals;
-    systemStatus;
-    autoscaleInterval;
     maybeRunInterval;
     queryingIsTaskReady;
     queryingIsFinished;
-    tasksDonePerSecondInterval;
-    _tasksPerMinute = Array.from({ length: 60 }, () => 0);
+    /**
+     * This pool's *own* in-flight task count, as opposed to {@link AutoscaledPool.currentConcurrency}, which is the
+     * (possibly shared) governor's total. `pause()` and `maybeFinish()` care only about this pool draining.
+     */
+    ownConcurrency = 0;
     constructor(options) {
         ow(options, ow.object.exactShape({
             runTaskFunction: ow.function,
             isFinishedFunction: ow.function,
             isTaskReadyFunction: ow.function,
-            maxConcurrency: ow.optional.number.integer.greaterThanOrEqual(1),
-            minConcurrency: ow.optional.number.integer.greaterThanOrEqual(1),
-            desiredConcurrency: ow.optional.number.integer.greaterThanOrEqual(1),
-            desiredConcurrencyRatio: ow.optional.number.greaterThan(0).lessThan(1),
-            scaleUpStepRatio: ow.optional.number.greaterThan(0).lessThan(1),
-            scaleDownStepRatio: ow.optional.number.greaterThan(0).lessThan(1),
             maybeRunIntervalSecs: ow.optional.number.greaterThan(0),
-            loggingIntervalSecs: ow.any(ow.number.greaterThan(0), ow.nullOrUndefined),
-            autoscaleIntervalSecs: ow.optional.number.greaterThan(0),
             taskTimeoutSecs: ow.optional.number.greaterThanOrEqual(0),
-            systemStatusOptions: ow.optional.object,
-            snapshotterOptions: ow.optional.object,
             log: ow.optional.object,
-            maxTasksPerMinute: ow.optional.number.integerOrInfinite.greaterThanOrEqual(1),
+            concurrencySystem: ow.object,
+            consumer: ow.object.partialShape({ id: ow.string.nonEmpty }),
         }));
-        const { runTaskFunction, isFinishedFunction, isTaskReadyFunction, maxConcurrency = 200, minConcurrency = 1, desiredConcurrency, desiredConcurrencyRatio = 0.9, scaleUpStepRatio = 0.05, scaleDownStepRatio = 0.05, maybeRunIntervalSecs = 0.5, loggingIntervalSecs = 60, taskTimeoutSecs = 0, autoscaleIntervalSecs = 10, systemStatusOptions, snapshotterOptions, log = serviceLocator.getLogger(), maxTasksPerMinute = Infinity, } = options;
+        const { runTaskFunction, isFinishedFunction, isTaskReadyFunction, maybeRunIntervalSecs = 0.5, taskTimeoutSecs = 0, log = serviceLocator.getLogger(), concurrencySystem, consumer, } = options;
         this.log = log.child({ prefix: 'AutoscaledPool' });
         // Configurable properties.
-        this.desiredConcurrencyRatio = desiredConcurrencyRatio;
-        this.scaleUpStepRatio = scaleUpStepRatio;
-        this.scaleDownStepRatio = scaleDownStepRatio;
         this.maybeRunIntervalMillis = maybeRunIntervalSecs * 1000;
-        this.loggingIntervalMillis = loggingIntervalSecs * 1000;
-        this.autoscaleIntervalMillis = autoscaleIntervalSecs * 1000;
         this.taskTimeoutMillis = taskTimeoutSecs * 1000;
         this.runTaskFunction = runTaskFunction;
         this.isFinishedFunction = isFinishedFunction;
         this.isTaskReadyFunction = isTaskReadyFunction;
-        this.maxTasksPerMinute = maxTasksPerMinute;
+        this.concurrencySystem = concurrencySystem;
+        this.consumer = consumer;
         // Internal properties.
-        this._minConcurrency = minConcurrency;
-        this._maxConcurrency = maxConcurrency;
-        this._desiredConcurrency = Math.min(desiredConcurrency ?? minConcurrency, maxConcurrency);
-        this._currentConcurrency = 0;
         this.isStopped = false;
         this.resolve = null;
         this.reject = null;
-        this.autoscale = this.autoscale.bind(this);
         this.maybeRunTask = this.maybeRunTask.bind(this);
-        this.incrementTasksDonePerSecond = this.incrementTasksDonePerSecond.bind(this);
-        // Create instances with correct options.
-        const ssoCopy = { ...systemStatusOptions };
-        ssoCopy.snapshotter ??= new Snapshotter({
-            ...snapshotterOptions,
-            log: this.log,
-        });
-        this.snapshotter = ssoCopy.snapshotter;
-        this.loadSignals = ssoCopy.loadSignals ?? [];
-        this.systemStatus = new SystemStatus(ssoCopy);
     }
     /**
-     * Gets the minimum number of tasks running in parallel.
-     */
-    get minConcurrency() {
-        return this._minConcurrency;
-    }
-    /**
-     * Sets the minimum number of tasks running in parallel.
+     * The governor backing this pool, as supplied to the constructor — exposed as the read-only
+     * {@link IConcurrencySystem} contract.
      *
-     * *WARNING:* If you set this value too high with respect to the available system memory and CPU, your code might run extremely slow or crash.
-     * If you're not sure, just keep the default value and the concurrency will scale up automatically.
-     */
-    set minConcurrency(value) {
-        ow(value, ow.optional.number.integer.greaterThanOrEqual(1));
-        this._minConcurrency = value;
-    }
-    /**
-     * Gets the maximum number of tasks running in parallel.
+     * This and the two getters below are telemetry only: concurrency is configured and tuned on the concrete
+     * {@link ConcurrencySystem} its owner holds, never through the pool.
      */
-    get maxConcurrency() {
-        return this._maxConcurrency;
+    get system() {
+        return this.concurrencySystem;
     }
-    /**
-     * Sets the maximum number of tasks running in parallel.
-     */
-    set maxConcurrency(value) {
-        ow(value, ow.optional.number.integer.greaterThanOrEqual(1));
-        this._maxConcurrency = value;
-    }
-    /**
-     * Gets the desired concurrency for the pool,
-     * which is an estimated number of parallel tasks that the system can currently support.
-     */
+    /** The estimated number of parallel tasks the governor can currently support. */
     get desiredConcurrency() {
-        return this._desiredConcurrency;
-    }
-    /**
-     * Sets the desired concurrency for the pool, i.e. the number of tasks that should be running
-     * in parallel if there's large enough supply of tasks.
-     */
-    set desiredConcurrency(value) {
-        ow(value, ow.optional.number.integer.greaterThanOrEqual(1));
-        this._desiredConcurrency = value;
+        return this.concurrencySystem.desiredConcurrency;
     }
     /**
-     * Gets the number of parallel tasks currently running in the pool.
+     * The number of parallel tasks currently booked against the governor. When it is shared, this counts every
+     * borrowing pool's tasks, not just this one's.
      */
     get currentConcurrency() {
-        return this._currentConcurrency;
+        return this.concurrencySystem.currentConcurrency;
     }
     /**
      * Runs the auto-scaled pool. Returns a promise that gets resolved or rejected once
      * all the tasks are finished or one of them fails.
+     *
+     * Throws if the {@link IConcurrencySystem|concurrency system} it borrows was never started — the pool assumes
+     * a running governor and cannot start one it does not own.
      */
     async run() {
+        // Checked here, on an awaited path — the capacity queries inside the task loop run from intervals and
+        // `setImmediate`, where a throw would become an unhandled rejection and hang `run()` forever.
+        if (!this.concurrencySystem.isRunning) {
+            throw new CriticalError('The ConcurrencySystem this AutoscaledPool borrows has not been started, so system load would not be ' +
+                'monitored and the concurrency would never be adjusted. Whoever creates a ConcurrencySystem owns ' +
+                'its lifecycle: call `await concurrencySystem.start()` before running the pools or crawlers that ' +
+                'use it, and `await concurrencySystem.stop()` once they are all done.');
+        }
         const poolPromise = new Promise((resolve, reject) => {
             this.resolve = resolve;
             this.reject = reject;
         });
-        await this.snapshotter.start();
-        await Promise.all(this.loadSignals.map((s) => s.start()));
-        // This interval checks the system status and updates the desired concurrency accordingly.
-        this.autoscaleInterval = betterSetInterval(this.autoscale, this.autoscaleIntervalMillis);
         // This is here because if we scale down to let's say 1, then after each promise is finished
         // this.maybeRunTask() doesn't trigger another one. So if that 1 instance gets stuck it results
         // in the crawler getting stuck and even after scaling up it never triggers another promise.
         this.maybeRunInterval = betterSetInterval(this.maybeRunTask, this.maybeRunIntervalMillis);
-        if (this.maxTasksPerMinute !== Infinity) {
-            // Start the interval that resets the counter of tasks per minute.
-            this.tasksDonePerSecondInterval = betterSetInterval(this.incrementTasksDonePerSecond, 1000);
-        }
         try {
             await poolPromise;
         }
@@ -249,6 +185,10 @@ export class AutoscaledPool {
      *
      * The promise returned from the {@link AutoscaledPool.run} function will not resolve
      * when `.pause()` is invoked (unlike abort, which resolves it).
+     *
+     * > *NOTE:* Pausing the pool does not suspend the (possibly shared) {@link ConcurrencySystem} — its
+     * autoscaling and resource monitoring keep running, since other pools borrowing it may still be active. To silence
+     * it during a long pause, its owner can `stop()` and `start()` it again.
      */
     async pause(timeoutSecs) {
         if (this.isStopped)
@@ -267,7 +207,7 @@ export class AutoscaledPool {
                 }, timeoutSecs);
             }
             interval = setInterval(() => {
-                if (this._currentConcurrency <= 0) {
+                if (this.ownConcurrency <= 0) {
                     // Clean up timeout and interval to prevent process hanging.
                     if (timeout)
                         clearTimeout(timeout);
@@ -316,17 +256,14 @@ export class AutoscaledPool {
             this.log.perf('Task will not run. Waiting for a ready task.');
             return done();
         }
-        // - we would exceed desired concurrency.
-        if (this._currentConcurrency >= this._desiredConcurrency) {
-            this.log.perf('Task will not run. Desired concurrency achieved.');
-            return done();
-        }
-        // - system is overloaded now and we are at or above minConcurrency
-        const currentStatus = this.systemStatus.getCurrentStatus();
-        const { isSystemIdle } = currentStatus;
-        if (!isSystemIdle && this._currentConcurrency >= this._minConcurrency) {
-            this.log.perf('Task will not be run. System is overloaded.', currentStatus);
-            return done();
+        // - the budget has room for us.
+        if (!this.concurrencySystem.hasCapacityForTask(this.consumer)) {
+            done();
+            // A shared governor's budget can stay saturated by another pool indefinitely, so we still have to be able
+            // to notice that *this* pool has run out of work — `maybeFinish()` is the only thing that ever resolves
+            // `run()`. It no-ops while this pool has tasks of its own in flight, which is every case in which an
+            // unshared governor reports no capacity.
+            return this.maybeFinish();
         }
         // - a task is ready.
         this.queryingIsTaskReady = true;
@@ -354,17 +291,14 @@ export class AutoscaledPool {
             // No tasks could mean that we're finished with all tasks.
             return this.maybeFinish();
         }
-        // - we have already reached the maximum tasks per minute
-        // we need to check this *after* checking if a task is ready to prevent hanging the pool
-        // for an extra minute if there are no more tasks
-        if (this.isOverMaxRequestLimit) {
-            this.log.perf('Task will not run. Maximum tasks per minute reached.');
+        // - the budget still has room. Re-checked atomically, because another pool sharing the governor may have taken
+        // the last free slot while we awaited `isTaskReadyFunction` above.
+        if (!this.concurrencySystem.tryRegisterTaskStart(this.consumer)) {
             return done();
         }
+        this.ownConcurrency++;
         try {
             // Everything's fine. Run task.
-            this._currentConcurrency++;
-            this._tasksPerMinute[0]++;
             // Try to run next task to build up concurrency,
             // but defer it so it doesn't create a cycle.
             setImmediate(this.maybeRunTask);
@@ -379,8 +313,8 @@ export class AutoscaledPool {
                 await this.runTaskFunction();
             }
             this.log.perf('Task finished.');
-            this._currentConcurrency--;
-            // Run task after the previous one finished.
+            // Run task after the previous one finished. Only on success: a failed task rejects the pool, and
+            // nudging the loop afterwards could start work on an already destroyed pool.
             setImmediate(this.maybeRunTask);
         }
         catch (e) {
@@ -397,85 +331,11 @@ export class AutoscaledPool {
                 this.reject(err);
             }
         }
-        return undefined;
-    }
-    /**
-     * Gets called every autoScaleIntervalSecs and evaluates the current system status.
-     * If the system IS NOT overloaded and the settings allow it, it scales up.
-     * If the system IS overloaded and the settings allow it, it scales down.
-     */
-    autoscale(intervalCallback) {
-        // Don't scale if paused.
-        if (this.isStopped)
-            return intervalCallback();
-        // Don't scale if we've hit the maximum requests per minute
-        if (this.isOverMaxRequestLimit)
-            return intervalCallback();
-        // Only scale up if:
-        // - system has not been overloaded lately.
-        const systemStatus = this.systemStatus.getHistoricalStatus();
-        const { isSystemIdle } = systemStatus;
-        // - we're not already at max concurrency.
-        const weAreNotAtMax = this._desiredConcurrency < this._maxConcurrency;
-        // - current concurrency reaches at least the given ratio of desired concurrency.
-        const minCurrentConcurrency = Math.floor(this._desiredConcurrency * this.desiredConcurrencyRatio);
-        const weAreReachingDesiredConcurrency = this._currentConcurrency >= minCurrentConcurrency;
-        if (isSystemIdle && weAreNotAtMax && weAreReachingDesiredConcurrency)
-            this.scaleUp(systemStatus);
-        // Always scale down if:
-        // - the system has been overloaded lately.
-        const isSystemOverloaded = !isSystemIdle;
-        // - we're over min concurrency.
-        const weAreNotAtMin = this._desiredConcurrency > this._minConcurrency;
-        if (isSystemOverloaded && weAreNotAtMin)
-            this.scaleDown(systemStatus);
-        // On periodic intervals, print comprehensive log information
-        if (this.loggingIntervalMillis > 0) {
-            const now = Date.now();
-            if (this.lastLoggingTime == null) {
-                this.lastLoggingTime = now;
-            }
-            else if (now > this.lastLoggingTime + this.loggingIntervalMillis) {
-                this.lastLoggingTime = now;
-                this.log.info('state', {
-                    currentConcurrency: this._currentConcurrency,
-                    desiredConcurrency: this._desiredConcurrency,
-                    systemStatus,
-                });
-            }
+        finally {
+            this.concurrencySystem.registerTaskEnd(this.consumer);
+            this.ownConcurrency--;
         }
-        // Start a new interval cycle.
-        return intervalCallback();
-    }
-    /**
-     * Scales the pool up by increasing
-     * the desired concurrency by the scaleUpStepRatio.
-     *
-     * @param systemStatus for logging
-     */
-    scaleUp(systemStatus) {
-        const step = Math.ceil(this._desiredConcurrency * this.scaleUpStepRatio);
-        this._desiredConcurrency = Math.min(this._maxConcurrency, this._desiredConcurrency + step);
-        this.log.debug('scaling up', {
-            oldConcurrency: this._desiredConcurrency - step,
-            newConcurrency: this._desiredConcurrency,
-            systemStatus,
-        });
-    }
-    /**
-     * Scales the pool down by decreasing
-     * the desired concurrency by the scaleDownStepRatio.
-     *
-     * @param systemStatus for logging
-     */
-    scaleDown(systemStatus) {
-        const step = Math.ceil(this._desiredConcurrency * this.scaleDownStepRatio);
-        this._desiredConcurrency = Math.max(this._minConcurrency, this._desiredConcurrency - step);
-        this.log.debug('scaling down', {
-            oldConcurrency: this._desiredConcurrency + step,
-            newConcurrency: this._desiredConcurrency,
-            systemStatus,
-        });
+        return undefined;
     }
     /**
      * If there are no running tasks and this.isFinishedFunction() returns true then closes
@@ -486,7 +346,7 @@ export class AutoscaledPool {
     async maybeFinish() {
         if (this.queryingIsFinished)
             return;
-        if (this._currentConcurrency > 0)
+        if (this.ownConcurrency > 0)
             return;
         this.queryingIsFinished = true;
         try {
@@ -512,23 +372,6 @@ export class AutoscaledPool {
     async destroy() {
         this.resolve = null;
         this.reject = null;
-        betterClearInterval(this.autoscaleInterval);
         betterClearInterval(this.maybeRunInterval);
-        if (this.tasksDonePerSecondInterval)
-            betterClearInterval(this.tasksDonePerSecondInterval);
-        if (this.snapshotter)
-            await this.snapshotter.stop();
-        await Promise.all(this.loadSignals.map((s) => s.stop()));
-    }
-    incrementTasksDonePerSecond(intervalCallback) {
-        this._tasksPerMinute.unshift(0);
-        this._tasksPerMinute.pop();
-        return intervalCallback();
-    }
-    get isOverMaxRequestLimit() {
-        if (this.maxTasksPerMinute === Infinity) {
-            return false;
-        }
-        return this._tasksPerMinute.reduce((acc, curr) => acc + curr, 0) >= this.maxTasksPerMinute;
     }
 }

package/autoscaling/client_load_signal.d.ts

@@ -1,25 +1,59 @@
-import type { StorageBackend } from '@crawlee/types';
-import type { LoadSnapshot } from './load_signal.js';
-import { SnapshotStore } from './load_signal.js';
+import type { LoadSignal, LoadSignalStartContext, LoadSnapshot } from './load_signal.js';
+/**
+ * A snapshot produced by the built-in client (rate-limit) signal.
+ * @internal
+ */
 export interface ClientSnapshot extends LoadSnapshot {
     rateLimitErrorCount: number;
 }
+/**
+ * Tuning for the built-in **client** (rate-limit) load signal, as accepted both by {@link ClientLoadSignal} and by
+ * the {@link LoadSignalsOptions.client|`client`} shorthand on {@link LoadSignalsOptions}.
+ */
 export interface ClientLoadSignalOptions {
-    client: StorageBackend;
-    clientSnapshotIntervalSecs?: number;
-    maxClientErrors?: number;
+    /**
+     * Defines the interval of checking the current state of the remote API client, in seconds.
+     * @default 1
+     */
+    snapshotIntervalSecs?: number;
+    /**
+     * Defines the maximum number of new rate limit errors within the given interval.
+     * @default 3
+     */
+    maxErrors?: number;
+    /**
+     * Maximum ratio of overloaded snapshots in a sample before the client counts as overloaded.
+     * @default 0.3
+     */
     overloadedRatio?: number;
-    snapshotHistoryMillis?: number;
 }
 /**
- * Periodically checks the storage client for rate-limit errors (HTTP 429)
- * and reports overload when the error delta exceeds a threshold.
+ * Periodically checks the storage backend for rate-limit errors (HTTP 429) and reports overload when the error delta
+ * exceeds a threshold.
+ *
+ * Built by default; construct one yourself only to wrap or adapt it — see {@link LoadSignal}.
+ *
+ * Switch it off entirely ({@link LoadSignalsOptions.client|`client: false`}) if the storage backend reports no
+ * rate-limit statistics, since it otherwise polls it every second to no purpose.
+ *
+ * @category Scaling
  */
-// @ts-ignore optional peer dependency or compatibility with es2022
-export declare function createClientLoadSignal(options: ClientLoadSignalOptions): Omit<import("./load_signal.js").LoadSignal, "getSample"> & {
-    store: SnapshotStore<ClientSnapshot>;
-    handle: (cb: () => unknown) => void;
-    getSample(sampleDurationMillis?: number): ClientSnapshot[];
-};
-/** @internal Return type for backward compat in Snapshotter facade */
-export type ClientLoadSignal = ReturnType<typeof createClientLoadSignal>;
+export declare class ClientLoadSignal implements LoadSignal {
+    readonly name = "clientInfo";
+    readonly overloadedRatio: number;
+    private readonly store;
+    private readonly intervalMillis;
+    private readonly maxErrors;
+    private interval?;
+    private client?;
+    constructor(options?: ClientLoadSignalOptions);
+    start(context: LoadSignalStartContext): Promise<void>;
+    stop(): Promise<void>;
+    getSample(sampleDurationMillis?: number): LoadSnapshot[];
+    /**
+     * Records one snapshot, overloaded when rate-limit errors grew by more than the configured limit since the
+     * previous one.
+     * @internal Also lets tests drive the measurement without waiting on a timer.
+     */
+    handle(intervalCallback: () => unknown): void;
+}

package/autoscaling/client_load_signal.js

@@ -1,36 +1,73 @@
+import { betterClearInterval, betterSetInterval } from '@apify/utilities';
+import { serviceLocator } from '../service_locator.js';
 import { SnapshotStore } from './load_signal.js';
 const CLIENT_RATE_LIMIT_ERROR_RETRY_COUNT = 2;
 /**
- * Periodically checks the storage client for rate-limit errors (HTTP 429)
- * and reports overload when the error delta exceeds a threshold.
+ * Periodically checks the storage backend for rate-limit errors (HTTP 429) and reports overload when the error delta
+ * exceeds a threshold.
+ *
+ * Built by default; construct one yourself only to wrap or adapt it — see {@link LoadSignal}.
+ *
+ * Switch it off entirely ({@link LoadSignalsOptions.client|`client: false`}) if the storage backend reports no
+ * rate-limit statistics, since it otherwise polls it every second to no purpose.
+ *
+ * @category Scaling
  */
-export function createClientLoadSignal(options) {
-    const maxClientErrors = options.maxClientErrors ?? 3;
-    const signal = SnapshotStore.fromInterval({
-        name: 'clientInfo',
-        overloadedRatio: options.overloadedRatio ?? 0.3,
-        intervalMillis: (options.clientSnapshotIntervalSecs ?? 1) * 1000,
-        snapshotHistoryMillis: options.snapshotHistoryMillis,
-        handler(store, intervalCallback) {
-            const now = new Date();
-            const allErrorCounts = options.client.stats?.rateLimitErrors ?? [];
-            const currentErrCount = allErrorCounts[CLIENT_RATE_LIMIT_ERROR_RETRY_COUNT] || 0;
-            const snapshot = {
-                createdAt: now,
-                isOverloaded: false,
-                rateLimitErrorCount: currentErrCount,
-            };
-            const all = store.getAll();
-            const previousSnapshot = all[all.length - 1];
-            if (previousSnapshot) {
-                const { rateLimitErrorCount } = previousSnapshot;
-                const delta = currentErrCount - rateLimitErrorCount;
-                if (delta > maxClientErrors)
-                    snapshot.isOverloaded = true;
-            }
-            store.push(snapshot, now);
-            intervalCallback();
-        },
-    });
-    return signal;
+export class ClientLoadSignal {
+    name = 'clientInfo';
+    overloadedRatio;
+    store = new SnapshotStore();
+    intervalMillis;
+    maxErrors;
+    interval;
+    client;
+    constructor(options = {}) {
+        this.overloadedRatio = options.overloadedRatio ?? 0.3;
+        this.intervalMillis = (options.snapshotIntervalSecs ?? 1) * 1000;
+        this.maxErrors = options.maxErrors ?? 3;
+        this.handle = this.handle.bind(this);
+    }
+    async start(context) {
+        this.store.useSampleWindow(context.maxSampleWindowMillis);
+        // A new session starts from a clean slate, or its first measurement diffs the error count against the previous
+        // session's — possibly against a different backend, since the client is resolved afresh just below.
+        this.store.clear();
+        // Resolved here rather than in the constructor, where asking for the backend would instantiate a default one
+        // as a side effect - long before the crawler that owns the run has had a chance to register its own.
+        this.client = serviceLocator.getStorageBackend();
+        this.interval = betterSetInterval(this.handle, this.intervalMillis);
+    }
+    async stop() {
+        if (this.interval)
+            betterClearInterval(this.interval);
+        this.interval = undefined;
+        this.client = undefined;
+    }
+    getSample(sampleDurationMillis) {
+        return this.store.getSample(sampleDurationMillis);
+    }
+    /**
+     * Records one snapshot, overloaded when rate-limit errors grew by more than the configured limit since the
+     * previous one.
+     * @internal Also lets tests drive the measurement without waiting on a timer.
+     */
+    handle(intervalCallback) {
+        const now = new Date();
+        const allErrorCounts = this.client?.stats?.rateLimitErrors ?? [];
+        const currentErrCount = allErrorCounts[CLIENT_RATE_LIMIT_ERROR_RETRY_COUNT] || 0;
+        const snapshot = {
+            createdAt: now,
+            isOverloaded: false,
+            rateLimitErrorCount: currentErrCount,
+        };
+        const all = this.store.getAll();
+        const previousSnapshot = all[all.length - 1];
+        if (previousSnapshot) {
+            const delta = currentErrCount - previousSnapshot.rateLimitErrorCount;
+            if (delta > this.maxErrors)
+                snapshot.isOverloaded = true;
+        }
+        this.store.push(snapshot, now);
+        intervalCallback();
+    }
 }