Skip to content

api.md

Latest commit

 

History

History
452 lines (371 loc) · 22.2 KB

api.md

File metadata and controls

452 lines (371 loc) · 22.2 KB
Raw

API

Table of contents

Pool

pool = new FixedThreadPool(numberOfThreads, specifier, opts)

numberOfThreads (mandatory) Number of workers for this pool.
specifier (mandatory) Specifier to a file with a worker implementation.
opts (optional) An object with the pool options properties described below.

pool = new DynamicThreadPool(min, max, specifier, opts)

min (mandatory) Same as FixedThreadPool numberOfThreads, this number of workers will be always active.
max (mandatory) Max number of workers that this pool can contain, the newly created workers will die after a threshold (default is 1 minute, you can override it in your worker implementation).
specifier (mandatory) Specifier to a file with a worker implementation.
opts (optional) An object with the pool options properties described below.

pool.execute(data, name, abortSignal, transferList)

data (optional) An object that you want to pass to your worker task function implementation.
name (optional) A string with the task function name that you want to execute on the worker. Default: 'default'
abortSignal (optional) An abort signal to abort the task function execution.
transferList (optional) An array of transferable objects that you want to transfer to your ThreadWorker implementation. The array is copied when this method is called; later changes do not affect submission. An empty array is treated as no transfer list.

This method is available on both pool implementations and returns a promise with the task function execution response.

Task failures

  • If sending a task throws, its promise rejects with the original error, without failing other tasks.
  • On a native worker error or an unexpected worker close notification, unfinished dispatched tasks reject with an Error and are not retried. Responses already received by the pool retain their result or task error.
  • Queued tasks are redistributed to ready workers respecting worker node affinity, including while other workers initialize. Affinity uses current worker node indices after removal and any replacement attempt. Tasks without an eligible destination reject.
  • Eligible handoffs complete before backPressureEnd is emitted, including when worker-removal event listeners disable task queues. Disabling queues from such a listener dispatches eligible detached tasks directly instead of discarding them.
  • Aborted tasks reject with the original abort reason if it is an Error, a new Error with that message if it is a string, or a task-specific abort Error otherwise. This applies before dispatch, while queued and while executing.
  • An already-aborted signal rejects before worker selection, queuing or transfer of input data.
  • Native worker failures without an Error payload are wrapped in an Error using the native error message and the original payload as cause. The pool's error event exposes that Error.
  • Failure to construct a replacement worker does not prevent task settlement; the construction error still propagates.
  • Without transfers, task errors retain cloneable worker-side input data in the taskError event, including worker mutations.
  • With a nonempty transfer list, taskError.data is the original input object. This does not transfer ownership back; its transferred objects remain owned by the worker. Raw worker error messages omit transferred input data.
  • A task error that cannot be reported causes an unexpected worker failure. Unfinished tasks then follow the failure rules above.

Intentional pool or worker termination alone does not guarantee rejection of unfinished tasks. Worker errors during termination still follow the failure rules. Bun provides a native worker close notification. The standard Web Worker API used by Deno and browsers has no exit notification: a worker that silently closes without an error cannot be detected. Deno and Bun preserve buffered task responses and task-function operation results sent before an observed unexpected exit, using their native synchronous port receiver. Exit consumers run before the private channel closes. Browsers have no standard synchronous port receiver: only responses already delivered to the pool retain their result. No browser queue-drain guarantee applies.

pool.mapExecute(data, name, abortSignals, transferList)

data An iterable of objects that you want to pass to your worker task function implementation.
name (optional) A string with the task function name that you want to execute on the worker. Default: 'default'
abortSignals (optional) An iterable of AbortSignal to abort the matching object task function execution.
transferList (optional) An array of transferable objects that you want to transfer to your ThreadWorker) worker implementation.

This method is available on both pool implementations and returns a promise with the task function execution responses array.

Task failures follow the same rules as pool.execute().

Signal iterables are consumed once and validated before dispatch. They must contain one AbortSignal per input item.

The transfer list uses one array snapshot for the batch. Later caller mutations do not change queued tasks or transfer ownership before dispatch.

pool.start()

This method is available on both pool implementations and will start the minimum number of workers.

pool.destroy()

This method is available on both pool implementations and returns a promise for pool destruction. Workers already removed after failure do not block destruction.

The pool waits up to 1000 milliseconds for each worker to confirm cleanup before terminating it. Cleanup failure or timeout does not prevent termination; unfinished cleanup can be interrupted. Reported cleanup failures emit the pool's error event when events are enabled. This timeout does not bound total pool destruction time.

pool.hasTaskFunction(name)

name (mandatory) The task function name.

This method is available on both pool implementations and returns a boolean.

pool.addTaskFunction(name, fn)

name (mandatory) The task function name.
fn (mandatory) The task function (data?: Data) => Response | PromiseLike<Response> or task function object { taskFunction: (data?: Data) => Response | PromiseLike<Response>, priority?: number, strategy?: WorkerChoiceStrategy, workerNodeKeys?: number[] }. Priority range is the same as Unix nice levels. workerNodeKeys is an array of worker node keys to restrict task execution to specific workers (worker node affinity).

Worker Node Affinity Notes

  • Worker node keys are validated at registration time against the pool's maximum size (maximumNumberOfWorkers ?? minimumNumberOfWorkers).
  • The number of worker node keys cannot exceed the pool's maximum size (maximumNumberOfWorkers ?? minimumNumberOfWorkers).
  • In dynamic pools, you can reference worker node keys up to the maximum pool size. Workers that don't exist yet are automatically created when a task targeting them is executed.
  • At execution time, submissions wait for eligible workers restoring runtime task functions. Initialization failures on ineligible workers do not reject waiting tasks. Otherwise selection retries until a worker becomes available or retries are exhausted.
  • Task stealing ignores worker node affinity, both on idle and under back pressure. For strict placement, disable both taskStealing and tasksStealingOnBackPressure. Initial worker selection and failure redistribution still respect affinity.

This method is available on both pool implementations and returns a boolean promise.

The function definition and its options, including the affinity array, are captured when this method is called. Later changes to the supplied definition do not alter pending registrations or their replay on replacement workers.

Functions registered through this method are installed on newly created workers, including replacements for fixed and minimum workers.

Pool-wide calls to addTaskFunction(), removeTaskFunction() and setDefaultTaskFunction() run in call order. A rejected call does not prevent later calls or undo changes already applied to workers.

Workers contacted by a change must respond or exit before the call settles, including when a later send fails. Send failures preserve the original error; otherwise, at least one accepted response is required and any other failed response rejects the change. Without a send failure, an unanswered target's exit does not reject a change with a successful peer, including synchronous exits from protected dispatch hooks. All targets exiting unanswered rejects the change. Deno and Bun preserve buffered operation results before an observed unexpected exit; browsers preserve only already-delivered results.

New and replacement workers wait for ongoing task function changes. They accept tasks only after restoring successfully registered task functions, the selected default and scheduling options. Workers being terminated remain unavailable. Replay remains bound to its original worker when protected listener-registration hooks synchronously retire or reindex worker nodes.

Successful calls resolve only after a restoration attempt completes for workers whose initialization failed. If restoration fails again, the change still succeeds, the worker remains unavailable, and the pool emits error when events are enabled.

Custom worker subclasses that defer task function operations must echo the request's taskFunctionOperationId in their acknowledgement, or retain the original taskFunctionProperties object when using the standard transport. Preserve result metadata when forwarding acknowledgements. An acknowledgement without a recoverable UUID rejects the operation. Pool-wide changes still wait for all contacted workers to respond or exit. Status-only overrides that discard this correlation must be updated. Deno and Bun preserve incoming asynchronous context; browsers preserve only synchronous context. Deferred browser handlers must retain the UUID or original properties object.

pool.removeTaskFunction(name)

name (mandatory) The task function name.

This method is available on both pool implementations and returns a boolean promise.

For workers whose initialization failed, confirmed absence of the requested function also counts as successful removal. Other removal errors still reject the call.

pool.listTaskFunctionsProperties()

This method is available on both pool implementations and returns an array of the task function properties.

pool.setDefaultTaskFunction(name)

name (mandatory) The task function name.

This method is available on both pool implementations and returns a boolean promise.

The last successfully acknowledged explicit default survives replacement and pool restart. With no workers, this method resolves to true without changing the default selection. A successful removal of that runtime function clears the selection; a failed removal does not.

Pool options

An object with these properties:

  • messageEventHandler (optional) - A function that will listen for message event on each worker.
    Default: () => {}

  • messageEventErrorHandler (optional) - A function that will listen for message event processing error on each worker.
    Default: () => {}

  • errorEventHandler? (optional) - A function that will listen for error event on each worker.
    Default: () => {}

  • workerChoiceStrategy (optional) - The default worker choice strategy to use in this pool:

    • WorkerChoiceStrategies.ROUND_ROBIN: Submit tasks to worker in a round robin fashion
    • WorkerChoiceStrategies.LEAST_USED: Submit tasks to the worker with the minimum number of executing and queued tasks
    • WorkerChoiceStrategies.LEAST_BUSY: Submit tasks to the worker with the minimum tasks execution time

    WorkerChoiceStrategies.WEIGHTED_ROUND_ROBIN, WorkerChoiceStrategies.INTERLEAVED_WEIGHTED_ROUND_ROBIN and WorkerChoiceStrategies.FAIR_SHARE strategies are targeted to heavy and long tasks.
    Default: WorkerChoiceStrategies.LEAST_USED

  • workerChoiceStrategyOptions (optional) - The worker choice strategy options object to use in this pool.
    Properties:

    • measurement (optional) - The measurement to use in worker choice strategies: runTime, waitTime
    • runTime (optional) - Use the tasks simple moving median runtime instead of the tasks simple moving average runtime in worker choice strategies.
    • waitTime (optional) - Use the tasks simple moving median wait time instead of the tasks simple moving average wait time in worker choice strategies.
    • weights (optional) - The worker weights to use in weighted round robin worker choice strategies: Record<number, number>.

    Default: { runTime: { median: false }, waitTime: { median: false } }

    Importing the public API in Deno does not require environment or system permissions. Computing default weights requires --allow-sys=cpus; supplying explicit weights avoids reading CPU information.

  • startWorkers (optional) - Start the minimum number of workers at pool initialization.
    Default: true

  • restartWorkerOnError (optional) - Restart worker on uncaught error in this pool.
    Default: true

  • enableEvents (optional) - Pool events emission enablement.
    Default: true

  • enableTasksQueue (optional) - Tasks queue per worker enablement in this pool.
    Default: false

  • tasksQueueOptions (optional) - The worker tasks queue options object to use in this pool.
    Properties:

    • size (optional) - The maximum number of tasks that can be queued on a worker before flagging it as back pressured. It must be a positive integer.
    • concurrency (optional) - The maximum number of tasks that can be executed concurrently on a worker. It must be a positive integer.
    • taskStealing (optional) - Task stealing enablement on idle. Delayed stealing retries follow the original worker when worker node indices shift and stop if that worker leaves the pool.
    • tasksStealingOnBackPressure (optional) - Tasks stealing enablement under back pressure.
    • tasksStealingRatio (optional) - The ratio of worker nodes that can steal tasks from another worker node. It must be a number between 0 and 1.
    • tasksFinishedTimeout (optional) - Queued tasks finished timeout in milliseconds at worker termination.
    • agingFactor (optional) - Aging factor for priority queue anti-starvation mechanism. It must be a positive number.
    • loadExponent (optional) - Load exponent for priority queue anti-starvation mechanism. It must be a positive number.

    Default: { size: (pool maximum size)^2, concurrency: 1, taskStealing: true, tasksStealingOnBackPressure: true, tasksStealingRatio: 0.6, tasksFinishedTimeout: 2000, agingFactor: 0.001, loadExponent: 0.667 }

  • workerOptions (optional) - An object with the worker options to pass to worker. See web worker for more details.
    Default:

    • Deno: { type: 'module' }
    • Bun: { type: 'module', smol: true }
    • Browser: { type: 'module' }

Worker

class YourWorker extends ThreadWorker

taskFunctions (mandatory) The task function or task functions object Record<string, (data?: Data) => Response | PromiseLike<Response> | { taskFunction: (data?: Data) => Response | PromiseLike<Response>, priority?: number, strategy?: WorkerChoiceStrategy, workerNodeKeys?: number[] }> that you want to execute on the worker. Priority range is the same as Unix nice levels. workerNodeKeys is an array of worker node keys to restrict task execution to specific workers (worker node affinity). See Worker Node Affinity Notes above for validation behavior. Task functions can return native promises, foreign promises, or custom promise-like results. Their settlement is awaited with or without an abort signal, even when the function is not declared async. Callable then getters are read once and invoked with the original result as receiver. Custom thenables remain supported when their proxies reject prototype or own-property inspection. In browsers, local promise candidates pass through Promise.resolve() for safe brand validation. A native promise with an exotic modified constructor can therefore invoke that constructor getter or require assimilation; Deno and Bun use native promise branding without this browser-specific step.
opts (optional) An object with these properties:

  • killBehavior (optional) - Dictates if your worker will be deleted in case a task is active on it.
    KillBehaviors.SOFT: If currentTime - lastActiveTime is greater than maxInactiveTime but the worker is stealing tasks or a task is executing or queued, then the worker won't be deleted.
    KillBehaviors.HARD: If currentTime - lastActiveTime is greater than maxInactiveTime but the worker is stealing tasks or a task is executing or queued, then the worker will be deleted.
    This option only apply to the newly created workers.
    Default: KillBehaviors.SOFT

  • maxInactiveTime (optional) - Maximum waiting time in milliseconds for tasks on newly created workers. After this time newly created workers will die. It must be a positive integer greater or equal than 5.
    The last active time of your worker will be updated when it terminates a task.
    If killBehavior is set to KillBehaviors.HARD this value represents also the timeout for the tasks that you submit to the pool, when this timeout expires your tasks is interrupted before completion and removed. The worker is killed if is not part of the minimum size of the pool.
    If killBehavior is set to KillBehaviors.SOFT your tasks have no timeout and your workers will not be terminated until your task is completed.
    Default: 60000

  • killHandler (optional) - A function that will be called when a worker is killed. It can return a promise-like result, which settles before acknowledgment. See pool.destroy() for timeout and failure handling.
    Default: () => {}

YourWorker.hasTaskFunction(name)

name (mandatory) The task function name.

This method is available on both worker implementations and returns { status: boolean, error?: Error }.

YourWorker.addTaskFunction(name, fn)

name (mandatory) The task function name.
fn (mandatory) The task function (data?: Data) => Response | PromiseLike<Response> or task function object { taskFunction: (data?: Data) => Response | PromiseLike<Response>, priority?: number, strategy?: WorkerChoiceStrategy, workerNodeKeys?: number[] }. Priority range is the same as Unix nice levels. workerNodeKeys is an array of worker node keys to restrict task execution to specific workers (worker node affinity). See Worker Node Affinity Notes above for validation behavior.

This method is available on both worker implementations and returns { status: boolean, error?: Error }.

YourWorker.removeTaskFunction(name)

name (mandatory) The task function name.

This method is available on both worker implementations and returns { status: boolean, error?: Error }.

YourWorker.listTaskFunctionsProperties()

This method is available on both worker implementations and returns an array of the task function properties.

YourWorker.setDefaultTaskFunction(name)

name (mandatory) The task function name.

This method is available on both worker implementations and returns { status: boolean, error?: Error }.