Orchestrator function code constraints

Build stateful apps with Durable Functions. It's an extension of Azure Functions. Use an orchestrator function to coordinate other Durable Functions in your function app. Orchestrator functions are stateful, reliable, and they're built to run for a long time.

Build stateful, fault-tolerant workflows with the Durable Task SDKs in .NET, Python, and Java. Use an orchestrator to coordinate activities and sub-orchestrations. Orchestrators are stateful, reliable, and they're built to run for a long time.

Orchestrator functions use event sourcing to ensure reliable execution and to maintain local variable state. The replay behavior of orchestrator code creates constraints on the type of code you can write in an orchestrator function. For example, orchestrator functions must be deterministic: an orchestrator function replays multiple times, and it must produce the same result each time.

Orchestrators use event sourcing to ensure reliable execution and to maintain local variable state. The replay behavior of orchestrator code creates constraints on the type of code you can write in an orchestrator. For example, orchestrators must be deterministic: an orchestrator replays multiple times, and it must produce the same result each time.

Here are some simple guidelines to help ensure your code is deterministic.

Call APIs from your target languages in orchestrator functions, but use only deterministic APIs. A deterministic API always returns the same value for the same input, no matter when or how often it's called.

The following sections provide guidance on APIs and patterns you should avoid because they're not deterministic. These restrictions apply only to orchestrator functions. Other function types don't have such restrictions.

Here are some simple guidelines to help ensure your code is deterministic.

Call APIs from your target languages in orchestrators, but use only deterministic APIs. A deterministic API always returns the same value for the same input, no matter when or how often it's called.

The following sections provide guidance on APIs and patterns you should avoid because they're not deterministic. These restrictions apply only to orchestrators. Activities don't have such restrictions.

Time-based APIs are nondeterministic and should never be used in orchestrator functions. Each orchestrator function replay produces a different value. Instead, use the Durable Functions equivalent API for getting the current date or time, which remains consistent across replays.

DateTime startTime = context.CurrentUtcDateTime;
// do some work
TimeSpan totalTime = context.CurrentUtcDateTime.Subtract(startTime);

// create a timer that expires 2 minutes from now
const expiration = moment.utc(context.df.currentUtcDateTime).add(2, "m");
const timeoutTask = context.df.createTimer(expiration.toDate());

# create a timer that expires 2 minutes from now
expiration = context.current_utc_datetime + timedelta(seconds=120)
timeout_task = context.create_timer(expiration)

$expiryTime = $Context.Input.ExpiryTime
while ($Context.CurrentUtcDateTime -lt $expiryTime) {
    # do work
}

Instant startTime = ctx.getCurrentInstant();
// do some work
Duration totalTime  = Duration.between(startTime, ctx.getCurrentInstant());

Time-based APIs are nondeterministic and should never be used in orchestrators. Each orchestrator replay produces a different value. Instead, use the Durable Task SDK equivalent API for getting the current date or time, which remains consistent across replays.

using Microsoft.DurableTask;

public class TimerExample : TaskOrchestrator<object?, TimeSpan>
{
    public override async Task<TimeSpan> RunAsync(TaskOrchestrationContext context, object? input)
    {
        // Use context.CurrentUtcDateTime instead of DateTime.Now or DateTime.UtcNow
        DateTime startTime = context.CurrentUtcDateTime;

        // do some work
        await context.CallActivityAsync("DoWork", null);

        TimeSpan totalTime = context.CurrentUtcDateTime.Subtract(startTime);
        return totalTime;
    }
}

from durabletask import task
from datetime import timedelta

def timer_example(ctx: task.OrchestrationContext, _):
    # Use ctx.current_utc_datetime instead of datetime.now() or datetime.utcnow()
    start_time = ctx.current_utc_datetime

    # Create a timer that expires 2 minutes from now
    expiration = ctx.current_utc_datetime + timedelta(minutes=2)
    yield ctx.create_timer(expiration)

    return "Timer completed"

import com.microsoft.durabletask.TaskOrchestration;
import com.microsoft.durabletask.TaskOrchestrationContext;
import java.time.Duration;
import java.time.Instant;

public class TimerExample implements TaskOrchestration {
    @Override
    public void run(TaskOrchestrationContext ctx) {
        // Use ctx.getCurrentInstant() instead of Instant.now() or LocalDateTime.now()
        Instant startTime = ctx.getCurrentInstant();

        // do some work
        ctx.callActivity("DoWork", null, Void.class).await();

        Duration totalTime = Duration.between(startTime, ctx.getCurrentInstant());
        ctx.complete(totalTime);
    }
}

APIs that return a random GUID or UUID are nondeterministic because the generated value is different for each replay. Depending on your language, a built-in API for generating deterministic GUIDs or UUIDs might be available. Otherwise, use an activity function to return a randomly generated GUID or UUID.

Guid randomGuid = context.NewGuid();

const randomGuid = context.df.newGuid();

randomGuid = context.new_guid()

APIs that return a random GUID or UUID are nondeterministic because the generated value is different for each replay. Depending on your language, a built-in API for generating deterministic GUIDs or UUIDs might be available. Otherwise, use an activity to return a randomly generated GUID or UUID.

using Microsoft.DurableTask;

public class GuidExample : TaskOrchestrator<object?, Guid>
{
    public override async Task<Guid> RunAsync(TaskOrchestrationContext context, object? input)
    {
        // Use context.NewGuid() instead of Guid.NewGuid()
        Guid randomGuid = context.NewGuid();
        return randomGuid;
    }
}

from durabletask import task

def guid_example(ctx: task.OrchestrationContext, _):
    # Use ctx.new_uuid() instead of uuid.uuid4()
    random_guid = ctx.new_uuid()
    return str(random_guid)

import com.microsoft.durabletask.TaskOrchestration;
import com.microsoft.durabletask.TaskOrchestrationContext;
import java.util.UUID;

public class GuidExample implements TaskOrchestration {
    @Override
    public void run(TaskOrchestrationContext ctx) {
        // Use ctx.newUUID() instead of UUID.randomUUID()
        UUID randomGuid = ctx.newUUID();
        ctx.complete(randomGuid.toString());
    }
}

Use an activity function to return random numbers to an orchestrator function. The return values of activity functions are always safe for replay because they're saved into the orchestration history.

Alternatively, you can use a random number generator with a fixed seed value directly in an orchestrator function. This approach is safe as long as the same sequence of numbers is generated for each orchestration replay.

Use an activity to return random numbers to an orchestrator. The return values of activities are always safe for replay because they're saved into the orchestration history.

Alternatively, you can use a random number generator with a fixed seed value directly in an orchestrator. This approach is safe as long as the same sequence of numbers is generated for each orchestration replay.

Don't use bindings in an orchestrator function, including the orchestration client and entity client bindings. Use input and output bindings only in a client or activity function. Orchestrator functions can replay multiple times, causing nondeterministic and duplicate I/O with external systems.

Orchestrators shouldn't perform direct I/O operations with external systems. Move I/O operations to activities. Orchestrators can replay multiple times, causing nondeterministic and duplicate I/O with external systems.

Static variables can change over time, making them unsafe for orchestrator functions. Avoid using static variables in orchestrator functions because their values can change over time, resulting in nondeterministic runtime behavior. Instead, use constants, or limit the use of static variables to activity functions.

Static variables can change over time, making them unsafe for orchestrators. Avoid using static variables in orchestrators because their values can change over time, resulting in nondeterministic runtime behavior. Instead, use constants, or limit the use of static variables to activities.

Environment variables in orchestrator functions can change over time, resulting in nondeterministic runtime behavior. If an orchestrator function needs configuration defined in an environment variable, you must pass the configuration value into the orchestrator function as an input or as the return value of an activity function.

Environment variables in orchestrators can change over time, resulting in nondeterministic runtime behavior. If an orchestrator needs configuration defined in an environment variable, you must pass the configuration value into the orchestrator as an input or as the return value of an activity.

Use activity functions to make outbound network calls. If you need to make an HTTP call from your orchestrator function, you can also use the durable HTTP APIs.

Use activities to make outbound network calls. Orchestrators should never make direct HTTP calls or other network requests because these operations are nondeterministic.

Blocking APIs like sleep can cause performance and scale problems for orchestrator functions and can result in unnecessary execution time charges in the Azure Functions Consumption plan. Use alternatives when they're available. For example, use Durable timers to create delays that are safe for replay and don't count toward orchestrator execution time.

Blocking APIs like "sleep" can cause performance and scale problems for orchestrators and should be avoided. Use durable timers to create delays that are safe for replay.

// Don't use Task.Delay() or Thread.Sleep()
// Use context.CreateTimer() instead
await context.CreateTimer(context.CurrentUtcDateTime.AddMinutes(5), CancellationToken.None);

from durabletask import task
from datetime import timedelta

def delay_example(ctx: task.OrchestrationContext, _):
    # Don't use time.sleep() or asyncio.sleep()
    # Use ctx.create_timer() instead
    fire_at = ctx.current_utc_datetime + timedelta(minutes=5)
    yield ctx.create_timer(fire_at)

// Don't use Thread.sleep()
// Use ctx.createTimer() instead
ctx.createTimer(Duration.ofMinutes(5)).await();

Orchestrator code must never start any async operation, except operations defined by the orchestration trigger's context object. For example, never use Task.Run, Task.Delay, and HttpClient.SendAsync in .NET or setTimeout and setInterval in JavaScript. An orchestrator function should only schedule async work using Durable SDK APIs, like scheduling activity functions. Any other type of async invocations should be done inside activity functions.

Orchestrator code must never start any async operation, except operations defined by the orchestration context object. For example, never use Task.Run, Task.Delay, and HttpClient.SendAsync in .NET. An orchestrator should only schedule async work using Durable Task SDK APIs, like scheduling activities. Any other type of async invocations should be done inside activities.

Async JavaScript functions

Declare JavaScript orchestrator functions as synchronous generator functions. Don't declare JavaScript orchestrator functions as async because the Node.js runtime doesn't guarantee deterministic behavior for async functions.

Don't declare Python orchestrator functions as coroutines. Don't use the async keyword because coroutine semantics don't align with the Durable Functions replay model. Declare Python orchestrator functions as generators, and use yield instead of await with the context API.

You must not declare Python orchestrators as coroutines. In other words, never declare Python orchestrators with the async keyword because coroutine semantics don't align with the Durable Task replay model. You must always declare Python orchestrators as generators, meaning that you should use yield instead of await when calling context APIs.

from durabletask import task

# CORRECT - use yield (generator function)
def my_orchestrator(ctx: task.OrchestrationContext, input: str):
    result = yield ctx.call_activity(my_activity, input=input)
    return result

# WRONG - don't use async/await
async def bad_orchestrator(ctx: task.OrchestrationContext, input: str):
    result = await ctx.call_activity(my_activity, input=input)  # This won't work!
    return result

The Durable Task Framework runs orchestrator code on a single thread and can't interact with any other threads. Running async continuations on a worker pool thread in an orchestration's execution can result in nondeterministic execution or deadlocks. For this reason, your orchestrator functions should almost never use threading APIs. For example, never use ConfigureAwait(continueOnCapturedContext: false) in an orchestrator function to ensure task continuations run on the orchestrator function's original SynchronizationContext.

The Durable Task Framework runs orchestrator code on a single thread and can't interact with any other threads. Running async continuations on a worker pool thread in an orchestration's execution can result in nondeterministic execution or deadlocks. For this reason, your orchestrators should almost never use threading APIs. For example, never use ConfigureAwait(continueOnCapturedContext: false) in an orchestrator to ensure task continuations run on the orchestrator's original SynchronizationContext.

A durable orchestration can run for days, months, years, or even as an eternal orchestration. Code changes that affect running orchestrations can break replay behavior, so plan carefully before you update your app. For more information, see Versioning.

A durable orchestration can run for days, months, years, or even indefinitely. Code changes that affect running orchestrations can break replay behavior, so plan carefully before you update your app. Common versioning strategies include side by side deployment and using version specific task hub names.

Tasks that can safely wait in orchestrator functions are sometimes called durable tasks. The Durable Task Framework creates and manages these tasks. Examples include the tasks returned by CallActivityAsync, WaitForExternalEvent, and CreateTimer in .NET orchestrator functions.

A list of TaskCompletionSource objects in .NET manages these durable tasks internally. During replay, orchestrator code creates these tasks. The dispatcher completes them as it enumerates the corresponding history events.

The runtime executes the tasks synchronously on a single thread until it replays the history. If a durable task doesn't finish by the end of history replay, the runtime takes the appropriate actions. For example, the runtime can enqueue a message to call an activity function.

This runtime behavior explains why your orchestrator function can't use await or yield in a nondurable task. The dispatcher thread can't wait for the task to finish, and callbacks from that task can corrupt the orchestrator function's tracking state. The runtime includes checks to help detect these violations.

To learn more about how the Durable Task Framework executes orchestrator functions, see the Durable Task source code on GitHub. In particular, see TaskOrchestrationExecutor.cs and TaskOrchestrationContext.cs.

Tasks that can safely wait in orchestrators are sometimes called durable tasks. The Durable Task Framework creates and manages these tasks. Examples include the tasks returned by CallActivityAsync, WaitForExternalEvent, and CreateTimer in .NET orchestrators.

A list of TaskCompletionSource objects in .NET manages these durable tasks internally. During replay, orchestrator code creates these tasks. The dispatcher completes them as it enumerates the corresponding history events.

The runtime executes the tasks synchronously on a single thread until it replays the history. If a durable task doesn't finish by the end of history replay, the runtime takes the appropriate actions. For example, the runtime can enqueue a message to call an activity.

This runtime behavior explains why your orchestrator can't use await or yield in a nondurable task. The dispatcher thread can't wait for the task to finish, and callbacks from that task can corrupt the orchestrator's tracking state. The runtime includes checks to help detect these violations.

To learn more about how the Durable Task Framework executes orchestrators, see the Durable Task source code on GitHub. In particular, see TaskOrchestrationExecutor.cs and TaskOrchestrationContext.cs.