Skip to main content
Version: june-2022

Workflows in TypeScript

@temporalio/workflow NPM API reference | GitHub

Background reading: Workflows in Temporal

Workflows are async functions that can orchestrate Activities and access special Workflow APIs, subject to deterministic limitations.

Each Workflow function has two parts:

  • The function name is known as the Workflow Type.
  • The function implementation code (body) is known as the Workflow Definition.

Each Workflow Definition is bundled with any third party dependencies, and registered by Workflow Type in a Worker. A Workflow function becomes a Workflow Execution (instance) only when started from a Workflow Client using its Workflow Type.

How to write a Workflow function

Workflow Definitions are "just functions", which can store state, and orchestrate Activity functions.

The snippet above uses proxyActivities to create functions that, when called, schedule a greet Activity in the system to say "Hello World".

A Workflow function can have multiple parameters, but we encourage you to use a single object parameter, as that helps with backward compatibility:

type ExampleArgs = {
name: string;
};

export async function example(args: ExampleArgs): Promise<{greeting: string}> {
const greeting = await greet(args.name);
return {greeting};
}

Workflow Limitations

Workflow code must be deterministic, and the TypeScript SDK replaces common sources of nondeterminism for you, like Date.now(), Math.random, and setTimeout (we recommend using our sleep API instead). However, there are other important limitations:

  • No Node built-ins like process or the path and fs modules
  • No filesystem access
  • No network access

These constraints don't apply inside Activities. If you need to ping an API, or access the filesystem (e.g. for building a CI/CD system), move that code into Activities.

How to Start and Cancel Workflows

See the TypeScript SDK Client docs for how to use WorkflowHandles to start, cancel, signal, query, describe and more.

Workflow Options

Workflows have options that determine what Task Queue they run on, what Search Attributes they are tagged with, Cron schedule, and more, but they are only set in the Temporal Client call (i.e. when you start or execute a Workflow) rather than inside the Workflow code itself.

Please see the Temporal Client docs or the API Reference for more info on Workflow Options.

Workflow APIs

The @temporalio/workflow package exports all the useful primitives that you can use in Workflows. See the API reference for the full list, but the main ones are:

APIsPurpose
proxyActivitiesMake idempotent side effects (like making a HTTP request) with Activities (see Activities doc)
proxyLocalActivitiesMake idempotent side effects (like making a HTTP request) with Activities (see Activities doc)
defineSignal/defineQuerySignal and Query Workflows while they are running
sleepDefer execution by sleeping for fixed time
conditionDefer execution until a condition is true, with optional timeout
startChild/executeChildSpawn new Child Workflows with customizable ParentClosePolicy
continueAsNewTruncate Event History for Entity Workflows
patched/deprecatePatchMigrate Workflows to new versions (see Patching doc)
uuid4Generate an RFC compliant V4 uuid without needing to call an Activity or Side Effect.
APIs for advanced usersincluding workflowInfo (to retrieve Workflow metadata), Workflow data Sinks, Cancellation Scopes, Failure types, and getExternalWorkflowHandle

You can import them individually or as a group:

// Option 1
import {sleep} from "@temporalio/workflow";

// Option 2
import * as wf from "@temporalio/workflow";

We fully expect that developers will bundle these into their own reusable Workflow libraries. If you do, please get in touch on Slack, we would love to work with you and promote your work.

The rest of this document explains the major Workflow APIs you should know:

  • Signals and Queries: defineSignal, defineQuery, and setHandler
  • Deferred Execution: sleep and condition
  • Child Workflows: startChild and executeChild
  • Entity (indefinitely long-running) Workflows: continueAsNew

Signals and Queries

Background reading: Signals and Queries in Temporal

How to define and receive Signals and Queries

Define Signals and Queries inside a Workflow

  • To add a Signal to a Workflow, call defineSignal with a name, and then attach a listener with setHandler.
  • To add a Query to a Workflow, call defineQuery with a name, and then attach a listener with setHandler.
  • Handlers for both Signals and Queries can take arguments, which can be used inside setHandler logic.
  • Only Signal Handlers can mutate state, and only Query Handlers can return values.

Define Signals and Queries Statically

If you know the name of your signals and queries upfront, we recommend declaring them outside of the Workflow Definition.

This helps provide type safety, since you can export the type signature of the signal or query to be called on the clientside.

Define Signals and Queries Dynamically

For more flexible usecases, you may want a dynamic Signal (such as a generated ID). You may handle it in two ways:

  • avoid making it dynamic by collapsing all signals in one handler and move the ID to the payload, or
  • actually make the Signal name dynamic by inlining the Signal definition per handler.
import * as wf from "@temporalio/workflow";

// "fat handler" solution
wf.setHandler(`genericSignal`, (payload) => {
switch (payload.taskId) {
case taskAId:
// do task A things
break;
case taskBId:
// do task B things
break;
default:
throw new Error("Unexpected task.");
}
});

// "inline definition" solution
wf.setHandler(wf.defineSignal(`task-${taskAId}`), (payload) => {
/* do task A things */
});
wf.setHandler(wf.defineSignal(`task-${taskBId}`), (payload) => {
/* do task B things */
});

// utility "inline definition" helper
const inlineSignal = (signalName, handler) =>
wf.setHandler(wf.defineSignal(signalName), handler);
inlineSignal(`task-${taskBId}`, (payload) => {
/* do task B things */
});
API Design FAQs

Why not new Signal and new Query?

The semantic of defineSignal/defineQuery is intentional, in that they return Signal/Query Definitions, not unique instances of Signals and Queries themselves. This is their entire source code:

/**
* Define a signal method for a Workflow.
*/
export function defineSignal<Args extends any[] = []>(
name: string
): SignalDefinition<Args> {
return {
type: "signal",
name,
};
}

/**
* Define a query method for a Workflow.
*/
export function defineQuery<Ret, Args extends any[] = []>(
name: string
): QueryDefinition<Ret, Args> {
return {
type: "query",
name,
};
}

Signals/Queries are only instantiated in setHandler and are specific to a particular Workflow Execution.

These distinctions may seem minor, but they model how Temporal works under the hood, because Signals and Queries are messages identified by "just strings" and don't have meaning independent of the Workflow having a listener to handle them. This will be clearer if you refer to the Client-side APIs below.

Why setHandler and not OTHER_API?

We named it setHandler instead of subscribe because Signals/Queries can only have one "handler" at a time, whereas subscribe could imply an Observable with multiple consumers, and is a higher level construct.

wf.setHandler(MySignal, handlerFn1);
wf.setHandler(MySignal, handlerFn2); // replaces handlerFn1

If you are familiar with RxJS, you are free to wrap your Signal and Query into Observables if you wish, or you could dynamically reassign the listener based on your business logic or Workflow state.

Invoke Signals and Queries from a Client

Sending Signals and making Queries requires having a Workflow handle from a Temporal Client.

  • You send a Signal with handle.signal(signal, ...args). A Signal has no return value by definition.
  • You make a Query with handle.query(query, ...args). A Query needs a return value, but can also take args.
  • You can refer to either by string name, which is useful for dynamic reference, but you will lose type inference.
// // inside Client code! not Workflow code!
import {increment, count} from "./workflow";

// init client code omitted - see Client docs
const handle = client.getHandle(workflowId);

// these three are equivalent
await handle.signal(increment, 1);
await handle.signal<[number]>("increment", 1);
await client.getHandle(workflowId).signal(increment, 1);

// these three are equivalent
let state = await handle.query(count);
let state = await handle.query<number>("count");
let state = await client.getHandle(workflowId).query(count);

By design of these Workflow handles, two different Workflows can use the same Signal or Query and there is still no ambiguity, because you always have to specify which Workflow you are signalling (workflowHandle1.signal(MySignal) vs workflowHandle2.signal(MySignal)).

Signals and Queries design patterns

Because Signals and Queries are intentionally flexible, you can wrap them up into reusable functions:

// implementation of queryable + signallable State in Workflow file
import * as wf from "@temporalio/workflow";

function useState<T = any>(name: string, initialValue: T) {
const signal = wf.defineSignal<[T]>(name);
const query = wf.defineQuery<T>(name);
let state: T = initialValue;
return {
signal,
query,
get value() {
// need to use closure because function doesn't rerun unlike React Hooks
return state;
},
set value(newVal: T) {
state = newVal;
},
};
}

// usage in Workflow file
const store = useState("my-store", 10);
function MyWorkflow() {
wf.setHandler(store.signal, (newValue: T) => {
// console.log('updating', newValue) // optional but useful for debugging
store.value = newValue;
});
wf.setHandler(store.query, () => store.value);
while (true) {
console.log("sleeping for ", store.value);
wf.sleep(store.value++ * 100); // you can mutate the value as well
}
}

// usage in Client file
await handle.signal(store.signal, 30);
const storeState = handle.query<number>(store.query); // 30

You can even conditionally set handlers, or set handlers inside handlers:

import * as wf from "@temporalio/workflow";
function MyWorkflow(signallable: boolean, signalNames: string[]) {
// conditional setting of handlers
if (signallable) {
wf.setHandler(MySignal, handler);
}

// set same handler for an array of signals by name
signalNames.forEach((name) => wf.setHandler(name, handler));

// signal handler that sets signal handlers
// // would be nice to send a function but we can't because it is not serializable
wf.setHandler(MySignal, (handlerName) => {
wf.setHandler(handlerName, handlers[handlerName]);
});
}

Additional Signals and Queries Notes

Type-safety for Signals and Queries

The Signals and Queries API has been designed with type safety in mind:

  • wf.defineQuery<Ret, Args>(name): QueryDefinition<Ret, Args>
  • wf.defineSignal<Args>(name): SignalDefinition<Args>
  • handle.query<Ret, Args>(def, ...args): Promise<Ret>
  • handle.signal<Args>(def, ...args): Promise<Ret>

You can either:

  • Define the argument type (and, for Queries, the return type) up front and import it for type inference with the WorkflowHandle
  • Define the expected type at the call site when you invoke the Signal/Query.
const increment =
wf.defineSignal<[number /* more args can be added here */]>("increment");
const count = wf.defineQuery<number /*, Arg[] can be added here */>("count");

// type safety inferred from definitions
await handle.signal(increment, 1);
await handle.signal(increment); // Expected 2 arguments, but got 1.
await handle.signal(increment, "1"); // Argument of type 'string' is not assignable to parameter of type 'number'

// common problems when you lack type safety
await handle.signal("increment"); // No TS error but insufficient arguments
await handle.signal("increment", "1"); // No TS error but sending in wrong type

// add type safety at callsite
await handle.signal<[number]>("increment"); // Expected 2 arguments, but got 1.
let state = await handle.query<number, [string]>("print", "Count: ");

Notes on Signals

  • Signal handlers are only guaranteed to be called in order per Signal Type, not across all of them. If you need strict ordering across multiple Signals, either:
    • Combine them into one Signal Type and use a switch statement.
    • Register handlers statically (call setHandler outside of the Workflow function).
  • WorkflowHandle.signal resolves as soon as Temporal Server has persisted the Signal, before the Workflow's Signal handler is called.
  • WorkflowHandle.signal Promise resolves with no value; Signal handlers cannot return data to the caller.
  • No Synchronous Updates. Users often want Signals to return a value, for example, a validation error. However Temporal has no way to surface any error to the external invocation. Signals are always asynchronous, in other words, a Signal always succeeds. Long term, the solution to this is "Synchronous Update" and it is under active development.

For now the best workaround is to use a Query to return Workflow state after signaling. Temporal guarantees read-after-write consistency of Signals-followed-by-Queries.

Notes on Queries

🚨 WARNING: NEVER mutate Workflow state inside a query! Generating Commands in Query handlers can lead to unexpected behaviors on subsequent executions.

How NOT to write a Query

This mutates Workflow state - do not do this:

export function badExample() {
let someState = 123;
wf.setHandler(query, () => {
return someState++; // bad! don't do this!
});
}

Reusing Signals and Queries in Libraries

Because Signal and Query Definitions are separate from Workflow Definitions, we can now compose them together:

// basic reusable Workflow component
export async function unblocked() {
let isBlocked = true;
wf.setHandler(unblockSignal, () => (isBlocked = false));
wf.setHandler(isBlockedQuery, () => isBlocked);
await wf.condition(() => !isBlocked);
}

// usage: signals can be sent to each Workflow separately
export async function myWorkflow1() {
await unblocked();
}
export async function myWorkflow2() {
await unblocked();
}

Another example of componentization can be found in our code samples.

signalWithStart

If you're not sure if a Workflow is running, you can signalWithStart a Workflow to send it a Signal and optionally start the Workflow if it is not running. Arguments for both are sent as needed.

// Signal With Start in Client file
const client = new WorkflowClient();
await client.signalWithStart(MyWorkflow, {
workflowId,
args: [arg1, arg2],
signal: MySignal,
signalArgs: [arg3, arg4],
});

See the Workflow Client docs for more notes on how starting Workflows and Workflow Options look like.

Deferred Execution

sleep and condition help you write durable asynchronous code in Temporal by offering an easy to use Promise-like API, but deferring, persisting, and resuming execution behind the scenes.

  • In other words, they do not "lock" the process, allowing one Worker to concurrently process hundreds of Workflows that sleep and await arbitrary conditions.
  • They are also "cancellation aware", allowing for graceful cleanup if the Workflow they are linked to is canceled. For more information, see Cancellation Scopes.

The Workflow's V8 isolate environment completely replaces the JavaScript setTimeout global, including inside libraries that you use, to provide a complete JavaScript runtime. We recommend using our sleep(timeout) API instead, because it is a cancellation-aware Promise wrapper for setTimeout.

Why Durable Timers Are a Hard Problem

JavaScript has a setTimeout, which seems relatively straightforward. However, they are held in memory - if your system goes down, those timers are gone.

A lot of careful code is required to make these timeouts fully reliable (aka recoverable in case of outage) and cancellation aware.

Beyond that, further engineering is needed to scale this - imagine 100,000 independently running timers in your system, firing every minute. That is the kind of scale Temporal handles.

sleep

sleep sets a durable timer for a fixed time period (an "Updatable Timer" pattern is documented below). It uses the ms package to take either a string or number of milliseconds, and returns a promise that you can await and catch when the Workflow Execution is cancelled.

import {sleep} from "@temporalio/workflow";

await sleep("30 days"); // string API
await sleep(30 * 24 * 60 * 60 * 1000); // numerical API

// `sleep` is cancellation-aware
// when workflow gets canceled during sleep, promise is rejected
await sleep("30 days").catch(() => {
// clean up code if workflow is canceled during sleep
});

// NOT VALID
await sleep("1 month"); // ms package doesnt support "months" https://github.com/vercel/ms/issues/57
// use date-fns and sleepUntil instead, see below

With this primitive, you can build other abstractions. For example, a sleepUntil function that converts absolute time to relative time with date-fns:

import * as wf from "@temporalio/workflow";
import differenceInMilliseconds from "date-fns/differenceInMilliseconds";

async function sleepUntil(futureDate, fromDate = new Date()) {
const timeUntilDate = differenceInMilliseconds(
new Date(futureDate),
fromDate
);
return wf.sleep(timeUntilDate);
}

sleepUntil("30 Sep " + (new Date().getFullYear() + 1)); // wake up when September ends
sleepUntil("5 Nov 2022 00:12:34 GMT"); // wake up at specific time and timezone

You can check the valid ISO string formats on MDN's Date docs. The upcoming ECMAScript Temporal API will offer more time utilities natively in JavaScript, alongside unfortunate name collision for Temporal developers.

Preventing Confusion: Workflow sleep vs Activity sleep

There is an unrelated sleep utility function available in Activity Context that is not durable, but is cancellation aware. See the Activities docs for details.

condition

The condition(fn, timeout?) API returns a promise that resolves:

  • true when the given predicate function (must be synchronous) returns true or
  • (optional) false if a timeout (given as a string or number of milliseconds) happens first.

This API is comparable to Workflow.await in other SDKs and often used to wait for Signals, since Signals are the main way to asynchronously update internal Workflow state (looped Activities are another).

The timeout also uses the ms package to take either a string or number of milliseconds.

// type signature
export function condition(
fn: () => boolean,
timeout: number | string
): Promise<boolean>;
export function condition(fn: () => boolean): Promise<void>;

// Usage
import * as wf from "@temporalio/workflow";

let x = 0;
// do stuff with x, eg increment every time you receive a signal
await wf.condition(() => x > 3);
// you only reach here when x > 3

// await either x > 3 or 30 minute timeout, whichever comes first
if (await wf.condition(() => x > 3, "30 mins")) {
// reach here if predicate true
} else {
// reach here if timed out
}

// track user progress with condition
export async function trackStepChanges(): Promise<void> {
let step = 0;
wf.setHandler(updateStep, (s) => void (step = s));
wf.setHandler(getStep, () => step);
await wf.condition(() => step === 1);
await wf.condition(() => step === 2);
}
Example usage in our Next.js One-Click Buy code sample

condition only returns true when the function evaluates to true; if the condition resolves as false, then a timeout has occurred. This leads to some nice patterns, like placing await condition inside an if:

condition Antipatterns
  • No time based condition functions are allowed in your function as this is very error prone. Use the optional timeout arg or a sleep timer.
  • condition only accepts synchronous functions that return a boolean. Do not put async functions, like Activities, inside the condition function.

Async design patterns

The real value of sleep and condition is in knowing how to use them to model asynchronous business logic. Here are some examples we use the most; we welcome more if you can think of them!

Racing Timers

Use Promise.race with Timers to dynamically adjust delays.

export async function processOrderWorkflow({
orderProcessingMS,
sendDelayedEmailTimeoutMS,
}: ProcessOrderOptions): Promise<void> {
let processing = true;
const processOrderPromise = processOrder(orderProcessingMS).then(() => {
processing = false;
});

await Promise.race([processOrderPromise, sleep(sendDelayedEmailTimeoutMS)]);

if (processing) {
await sendNotificationEmail();
await processOrderPromise;
}
}
Racing Signals

Use Promise.race with Signals and Triggers to have a promise resolve at the earlier of either system time or human intervention.

import {Trigger, sleep, defineSignal} from "@temporalio/workflow";

const userInteraction = new Trigger<boolean>();
const completeUserInteraction = defineSignal("completeUserInteraction");

export async function myWorkflow(userId: string) {
setHandler(completeUserInteraction, () => userInteraction.resolve(true)); // programmatic resolve
const userInteracted = await Promise.race([
userInteraction,
sleep("30 days"),
]);
if (!userInteracted) {
await sendReminderEmail(userId);
}
}

You can invert this to create a Reminder pattern where the promise resolves IF no Signal is received.

Antipattern: Racing sleep.then

Be careful when racing a chained sleep. This may cause bugs because the chained .then will still continue to execute.

await Promise.race([
sleep("5s").then(() => (status = "timed_out")),
somethingElse.then(() => (status = "processed")),
]);

if (status === "processed") await complete(); // takes more than 5 seconds
// status = timed_out
Updatable Timer

Here is how you can build an updatable timer with condition:

import * as wf from "@temporalio/workflow";

// usage
export async function countdownWorkflow(): Promise<void> {
const target = Date.now() + 24 * 60 * 60 * 1000; // 1 day!!!
const timer = new UpdatableTimer(target);
console.log("timer set for: " + new Date(target).toString());
wf.setHandler(setDeadlineSignal, (deadline) => {
// send in new deadlines via Signal
timer.deadline = deadline;
console.log("timer now set for: " + new Date(deadline).toString());
});
wf.setHandler(timeLeftQuery, () => timer.deadline - Date.now());
await timer; // if you send in a signal with a new time, this timer will resolve earlier!
console.log("countdown done!");
}

This is available in the third party temporal-time-utils package where you can also see the implementation:

// implementation
export class UpdatableTimer implements PromiseLike<void> {
deadlineUpdated = false;
#deadline: number;

constructor(deadline: number) {
this.#deadline = deadline;
}

private async run(): Promise<void> {
/* eslint-disable no-constant-condition */
while (true) {
this.deadlineUpdated = false;
if (
!(await wf.condition(
() => this.deadlineUpdated,
this.#deadline - Date.now()
))
) {
break;
}
}
}

then<TResult1 = void, TResult2 = never>(
onfulfilled?: (value: void) => TResult1 | PromiseLike<TResult1>,
onrejected?: (reason: any) => TResult2 | PromiseLike<TResult2>
): PromiseLike<TResult1 | TResult2> {
return this.run().then(onfulfilled, onrejected);
}

set deadline(value: number) {
this.#deadline = value;
this.deadlineUpdated = true;
}

get deadline(): number {
return this.#deadline;
}
}

Triggers

Triggers are an experimental Promise-like concept in the TypeScript SDK.

Triggers, like Promises, can be awaited and expose a then method. Unlike Promises, they export resolve or reject methods, so you can programmatically control them.

Trigger Code Example
import {Trigger, sleep, defineSignal} from "@temporalio/workflow";

const userInteraction = new Trigger<boolean>();
const completeUserInteraction = defineSignal("completeUserInteraction");

export async function myWorkflow(userId: string) {
setHandler(completeUserInteraction, () => userInteraction.resolve(true)); // programmatic resolve
const userInteracted = await Promise.race([
userInteraction,
sleep("30 days"),
]);
if (!userInteracted) {
await sendReminderEmail(userId);
}
}

Trigger is CancellationScope-aware. It is linked to the current scope on construction and throws when that scope is cancelled.

In most cases, you should now be able to use condition instead of Triggers, and we may deprecate Triggers in future.

Child Workflows

Besides Activities, a Workflow can also start other, "Child" Workflows. Child Workflows have a subset of APIs from Temporal Clients, including how to start/execute/handle Workflows.

startChild starts a child workflow without awaiting completion, and returns a ChildWorkflowHandle:

import {startChild} from "@temporalio/workflow";

export async function parentWorkflow(names: string[]) {
const childHandle = await startChild(childWorkflow, {
args: [name],
// workflowId, // add business-meaningful workflow id here
// // regular workflow options apply here, with two additions (defaults shown):
// cancellationType: ChildWorkflowCancellationType.WAIT_CANCELLATION_COMPLETED,
// parentClosePolicy: ParentClosePolicy.PARENT_CLOSE_POLICY_TERMINATE
});
// you can use childHandle to signal or get result here
await childHandle.signal("anySignal");
const result = childHandle.result();
// you can use childHandle to signal, query, cancel, terminate, or get result here
}

You should use cancellationScopes if you need to cancel Child Workflows.

executeChild starts a child workflow and awaits (blocks until) its completion:

To control any running Workflow from inside a Workflow, use getExternalWorkflowHandle(workflowId).

import {getExternalWorkflowHandle, workflowInfo} from "@temporalio/workflow";

export async function terminateWorkflow() {
const {workflowId} = workflowInfo(); // no await needed
const handle = getExternalWorkflowHandle(workflowId); // sync function, not async
await handle.cancel();
}

Special Notes:

  • Child Workflow options automatically inherit their values from the Parent Workflow options if they are not explicitly set. They have two advanced options unique to Child Workflows:
    • cancellationType: Controls at which point to throw the CanceledFailure exception when a child workflow is cancelled
    • parentClosePolicy: Explained below
  • Child Workflow executions are CancellationScope aware and will automatically be cancelled when their containing scope is cancelled.
When to use Child Workflows vs Activities

Child Workflows and Activities are both started from Workflows, so you may feel confused about when to use which. Here are some important differences:

  • Child Workflows have access to all Workflow APIs, but are subject to Workflow Limitations. Activities have the inverse pros and cons.
  • Child Workflows can continue on if their Parent is canceled, with a ParentClosePolicy of ABANDON, whereas Activities are always canceled when their Workflow is canceled (they may react to a cancellationSignal for cleanup if canceled). The decision is roughly analogous to spawning a child process in a terminal to do work vs doing work in the same process.
  • Temporal tracks all state changes within Child Workflows in Event History, whereas only the input, output, and retry attempts of Activities are tracked.

Activities usually model a single operation on the external world. Workflows are modeling composite operations that consist of multiple activities or other child workflows.

When in doubt, use Activities.

Parent Close Policy

A Parent Close Policy determines what happens to a Child Workflow Execution if its Parent changes to a Closed status (Completed, Failed, or Timed out).

There are three possible values:

  • Abandon: the Child Workflow Execution is not affected.
  • Terminate (default): the Child Workflow Execution is forcefully Terminated.
  • Request Cancel: a Cancellation request is sent to the Child Workflow Execution.

ParentClosePolicy proto definition

Each Child Workflow Execution may have its own Parent Close Policy. This policy applies only to Child Workflow Executions and has no effect otherwise.

Parent Close Policy entity relationship

You can set policies per child, which means you can opt out of propagating terminates / cancels on a per-child basis. This is useful for starting Child Workflows asynchronously (see relevant issue here or the corresponding SDK docs).

continueAsNew

We need to call continueAsNew before our Workflow hits the 50,000 Event limit. Events are generated when a Workflow does various things involving Temporal Server, including calling an Activity, receiving a Signal, or calling sleep, but not handling a Query.

More info

continueAsNew stops the current Workflow Execution and starts another one with new arguments and an empty Event History. Note that this is done immediately, so make sure that your Signal handlers have finished running before calling continueAsNew.

You can also call continueAsNew from a signal handler or continueAsNew to a different Workflow (or different Task Queue) using makeContinueAsNewFunc.

If you need to know whether a Workflow was started via continueAsNew, you can pass an optional last argument as true:

export async function loopingWorkflow(foo: any, isContinued?: boolean) {
// some logic based on foo, branching on isContinued

(await continueAsNew)<typeof loopingWorkflow>(foo, true);
}

Don't overuse

You should not try to call continueAsNew too often - if at all! It's primary purpose is to truncate event history, which if too large may slow down your workflows and eventually cause an error. Calling it too frequently to be preemptive can cause other performance issues as each new Workflow Execution has overhead.

Temporal's default limits are set to warn you at 10,000 events in a single Workflow Execution, and error at 50,000. This is sufficient for:

  • If executing one activity a day, it can support an infinite loop for over 2 decades (27 years)
  • If executing one activity an hour, it can support an infinite loop for over 1 year (417 days)
  • If executing one activity a minute, it can support an infinite loop for over 1 week (7 days)

without even resorting to continueAsNew.

Our recommendation is to size it to continue as new between once a day to once a week, to ensure old version branches can be removed in a timely manner.

Example

Here is a simple pattern that we recommend to represent a single entity. It keeps track of the number of iterations regardless of frequency, and calls continueAsNew while properly handling pending updates from Signals.

interface Input {
/* define your workflow input type here */
}
interface Update {
/* define your workflow update type here */
}

const MAX_ITERATIONS = 1;

export async function entityWorkflow(
input: Input,
isNew = true
): Promise<void> {
try {
const pendingUpdates = Array<Update>();
setHandler(updateSignal, (updateCommand) => {
pendingUpdates.push(updateCommand);
});

if (isNew) {
await setup(input);
}

for (let iteration = 1; iteration <= MAX_ITERATIONS; ++iteration) {
// Ensure that we don't block the Workflow Execution forever waiting
// for updates, which means that it will eventually Continue-As-New
// even if it does not receive updates.
await condition(() => pendingUpdates.length > 0, "1 day");

while (pendingUpdates.length) {
const update = pendingUpdates.shift();
await runAnActivityOrChildWorkflow(update);
}
}
} catch (err) {
if (isCancellation(err)) {
await CancellationScope.nonCancellable(async () => {
await cleanup();
});
}
throw err;
}
await continueAsNew<typeof entityWorkflow>(input, false);
}

Putting it all together

Individually, the core Workflow APIs (Signals/Queries, sleep/condition, startChild/executeChild, and continueAsNew) are interesting, but they become truly powerful when wielded together.

We can illustrate this by building an example Workflow that combines them.

Schedule Workflow Example

One common request from users is for more powerful alternatives to Cron Workflows. We can try implementing them with the Workflow API primitives we have learned here.

Some desirable requirements:

  • One Parent Workflow that schedules ChildWorkflows based on either a:
    • Cron string (with timezone support, eg "at 8am every day")
    • or "unaligned" sleep period (eg "every 3 hours")
  • Allows setting:
    • a random "jitter" period to spread out execution
    • a maximum number of invocations, or have the schedules end by a set date
    • a "paused" or "running" state (that can also be queried)
  • Allows querying:
    • the expected times of the next N invocations
    • the number of invocations so far
  • Allows manual trigger at any point

Take some time to think about how you would implement these features, and then look at our suggested solution below.

Example CronScheduleWorkflow

The desired clientside usage would look something like this:

// client.ts
const handle = await client.start(MyScheduleWorkflow, {
args: [
{
cronParser: {
expression: "0 8 * * *", // every day 8am
options: {
currentDate: "2016-03-27 00:00:01",
endDate: new Date("Wed, 26 Dec 2012 14:40:00 UTC"),
tz: "Europe/Athens",
},
},
maxInvocations: 500,
jitterMs: 1000,
userId, // defined elsewhere
},
],
taskQueue: "scheduler",
workflowId: "schedule-for-" + userId,
});

This Workflow would want a sleepUntil timer at its core to power the scheduling - ideal for implementing clock-aligned "run at a set time" semantics. Temporal doesn't export sleepUntil for you - you can write your own with some simple time math.

import * as wf from "@temporalio/workflow";
import parser from "cron-parser";
import differenceInMilliseconds from "date-fns/differenceInMilliseconds";

// example atomic unit of work you are scheduling, can be workflow or activity or whatever
async function spawnChild(
userId: string,
nextTime: string,
invocation: number
) {
return wf.executeChild(childWorkflow, {
args: [userId],
workflowId: `childWorkflow-${invocation}-${nextTime}`,
// // regular workflow options apply here, with two additions (defaults shown):
// cancellationType: ChildWorkflowCancellationType.WAIT_CANCELLATION_COMPLETED,
// parentClosePolicy: ParentClosePolicy.PARENT_CLOSE_POLICY_TERMINATE
});
}

export async function sleepUntil(futureDate: string, fromDate = new Date()) {
const timeUntilDate = differenceInMilliseconds(
new Date(futureDate),
fromDate
);
return wf.sleep(timeUntilDate);
}

// queries
export const numInvocationsQuery = wf.defineQuery("numInvocationsQuery");
export const futureScheduleQuery = wf.defineQuery("futureScheduleQuery");
export const manualTriggerSignal = wf.defineSignal("manualTriggerSignal");
export type ScheduleWorkflowState = "RUNNING" | "PAUSED" | "STOPPED";
export const stateSignal =
wf.defineSignal<[ScheduleWorkflowState]>("stateSignal");
export const stateQuery = wf.defineQuery<ScheduleWorkflowState>("stateQuery");

export async function CronScheduleWorkflow(
args: {
cronParser: {
expression: string;
options: parser.ParserOptions;
};
callbackFn: (nextTime?: string, invocations?: number) => Promise<void>;
maxInvocations?: number;
jitterMs?: number;
userId?: string;
},
invocations = 1
) {
// signal and query handlers
wf.setHandler(numInvocationsQuery, () => invocations);
wf.setHandler(manualTriggerSignal, () =>
spawnChild(userId, nextTime.toString(), invocations++)
);
let scheduleWorkflowState = "RUNNING" as ScheduleWorkflowState;
wf.setHandler(stateQuery, () => scheduleWorkflowState);
wf.setHandler(stateSignal, (state) => void (scheduleWorkflowState = state));

const interval = parser.parseExpression(
args.cronParser.expression,
args.cronParser.options
);
const nextTime = interval.next().toString();
wf.setHandler(futureScheduleQuery, (numEntriesInFutureSchedule?: number) => {
const interval = parser.parseExpression(
args.cronParser.expression,
args.cronParser.options
); // reset interval
return {
futureSchedule: genNextTimes(numEntriesInFutureSchedule, () =>
interval.next().toString()
),
timeLeft: differenceInMilliseconds(new Date(nextTime), new Date()),
};
});

// timer logic
try {
await sleepUntil(nextTime);
if (args.jitterMs) {
await wf.sleep(Math.floor(Math.random() * (args.jitterMs + 1)));
}
if (scheduleWorkflowState === "PAUSED") {
await wf.condition(() => scheduleWorkflowState === "RUNNING");
}
await spawnChild(userId, nextTime.toString(), invocations); // no need to increment invocations bc relying on continueAsNew for that
if (args.maxInvocations && args.maxInvocations > invocations) {
await wf.continueAsNew<typeof CronScheduleWorkflow>(
args,
invocations + 1
);
} else {
scheduleWorkflowState = "STOPPED";
}
} catch (err) {
if (wf.isCancellation(err)) scheduleWorkflowState = "STOPPED";
else throw err;
}
}

// shared
function genNextTimes<T extends string | Date>(
number = 5,
getNextTimes: () => T
): T[] {
const times = [];
for (let i = 0; i < number; i++) {
times.push(getNextTimes());
}
return times;
}

You can extend or add features as you please. For example, notice that we only implemented a very trivial cancellation cleanup step. By default, if a Parent Workflow is cancelled, all child workflows will be cancelled as well. What if you wanted them to carry on to completion? (Hint: check the ParentClosePolicy).

Workflow Utility Libraries

As you build up strong opinions of how you'd like to compose behavior, you may want to publish reusable Temporal utility function or Temporal Workflow libraries. Let us know and we'd be happy to feature them here!

  • temporal-time-utils: Contains reusable versions of sleepUntil, UpdatableTimer, and ScheduleWorkflow described on this page.

Just keep in mind the difference between utility functions (deterministic, uses Workflow APIs but have to be inlined into Workflows rather than used standalone) and Workflow functions (can be used standalone, and subject to all Workflow limitations, including that all args and results must be JSON-serializable.)