TimeProviderExtensions 1.0.0-rc.3

TimeProvider Extensions

Testing extensions for the System.TimeProvider API. It includes:

• An advanced test/fake version of the TimeProvider type, named ManualTimeProvider, that allows you to control the progress of time during testing deterministically (see the difference to Microsoft's FakeTimeProvider below).
• A backported version of PeriodicTimer that supports TimeProvider in .NET 6.

Quick start

This describes how to get started:

1. Get the latest release from https://www.nuget.org/packages/TimeProviderExtensions.

2. Take a dependency on TimeProvider in your production code. Inject the production version of TimeProvider available via the TimeProvider.System property during production.

3. During testing, inject the ManualTimeProvider from this library. This allows you to write tests that run fast and predictably.

   • Advance time by calling Advance(TimeSpan) or SetUtcNow(DateTimeOffset) or
   • Jump ahead in time using Jump(TimeSpan) or Jump(DateTimeOffset).

4. See the ManualTimeProvider API page for the full API documentation for ManualTimeProvider.

5. Read the rest of this README for further details and examples.

API Overview

These pages have all the details of the API included in this package:

• ManualTimeProvider
• AutoAdvanceBehavior
• ManualTimer

.NET 7 and earlier:

• PeriodicTimerWrapper
• TimeProviderPeriodicTimerExtensions

Known limitations and issues:

• If running on .NET versions earlier than .NET 8.0, there is a constraint when invoking CancellationTokenSource.CancelAfter(TimeSpan) on the CancellationTokenSource object returned by CreateCancellationTokenSource(TimeSpan delay). This action will not terminate the initial timer indicated by the delay argument initially passed the CreateCancellationTokenSource method. However, this restriction does not apply to .NET 8.0 and later versions.
• To enable controlling PeriodicTimer via TimeProvider in versions of .NET earlier than .NET 8.0, the TimeProvider.CreatePeriodicTimer returns a PeriodicTimerWrapper object instead of a PeriodicTimer object. The PeriodicTimerWrapper type is just a lightweight wrapper around the original System.Threading.PeriodicTimer and will behave identically to it.
• If ManualTimeProvider is created via AutoFixture, be aware that will set writable properties with random values. This behavior can be overridden by providing a customization to AutoFixture, e.g.:

  fixture.Customize<ManualTimeProvider>(x => x.OmitAutoProperties()));
  

  or by using the [NoAutoProperties] attribute, if using AutoFixture.Xunit2.

Installation and Usage

Get the latest release from https://www.nuget.org/packages/TimeProviderExtensions

Set up in production

To use in production, pass in TimeProvider.System to the types that depend on TimeProvider. This can be done directly or via an IoC Container, e.g., .NETs built-in IServiceCollection like so:

services.AddSingleton(TimeProvider.System);

If you do not want to register the TimeProvider with your IoC container, you can instead create an additional constructor in the types that use it, which allows you to pass in a TimeProvider, and in the existing constructor(s) you have, just new up TimeProvider.System directly. For example:

public class MyService
{
  private readonly TimeProvider timeProvider;

  public MyService() : this(TimeProvider.System)
  {
  }

  public MyService(TimeProvider timeProvider)
  {
    this.timeProvider = timeProvider;
  }
}

This allows you to explicitly pass in a ManualTimeProvider during testing.

Example - control time during tests

If a system under test (SUT) uses things like Task.Delay, DateTimeOffset.UtcNow, Task.WaitAsync, or PeriodicTimer, it becomes hard to create tests that run fast and predictably.

The idea is to replace the use of e.g. Task.Delay with an abstraction, the TimeProvider, that in production is represented by the TimeProvider.System, which just uses the real Task.Delay. During testing it is now possible to pass in ManualTimeProvider, which allows the test to control the progress of time, making it possible to skip ahead, e.g. 10 minutes, and also pause time, leading to fast and predictable tests.

As an example, let us test the "Stuff Service" below that performs specific tasks every 10 seconds with an additional 1-second delay. We have two versions, one that uses the standard types in .NET, and one that uses the TimeProvider.

// Version of stuff service that uses the built-in DateTimeOffset, PeriodicTimer, and Task.Delay
public class StuffService
{
  private static readonly TimeSpan doStuffDelay = TimeSpan.FromSeconds(10);
  private readonly List<DateTimeOffset> container;

  public StuffService(List<DateTimeOffset> container)
  {
    this.container = container;
  }

  public async Task DoStuff(CancellationToken cancelllationToken)
  {
    using var periodicTimer = new PeriodicTimer(doStuffDelay);

    while (await periodicTimer.WaitForNextTickAsync(cancellationToken))
    {
      await Task.Delay(TimeSpan.FromSeconds(1));
      container.Add(DateTimeOffset.UtcNow);
    }
  }
}

// Version of stuff service that uses the built-in TimeProvider
public class StuffServiceUsingTimeProvider
{
  private static readonly TimeSpan doStuffDelay = TimeSpan.FromSeconds(10);
  private readonly TimeProvider timeProvider;
  private readonly List<DateTimeOffset> container;

  public StuffServiceUsingTimeProvider(TimeProvider timeProvider, List<DateTimeOffset> container)
  {
    this.timeProvider = timeProvider;
    this.container = container;
  }

  public async Task DoStuff(CancellationToken cancelllationToken)
  {
    using var periodicTimer = timeProvider.CreatePeriodicTimer(doStuffDelay);

    while (await periodicTimer.WaitForNextTickAsync(cancellationToken))
    {
      await timeProvider.Delay(TimeSpan.FromSeconds(1));
      container.Add(timeProvider.GetUtcNow());
    }
  }
}

The test, using xUnit and FluentAssertions, could look like this:

[Fact]
public void DoStuff_does_stuff_every_11_seconds()
{
  // Arrange
  var timeProvider = new ManualTimeProvider();
  var container = new List<DateTimeOffset>();
  var sut = new StuffServiceUsingTimeProvider(timeProvider, container);

  // Act
  _ = sut.DoStuff(CancellationToken.None);
  timeProvider.Advance(TimeSpan.FromSeconds(11));

  // Assert
  container
    .Should()
    .ContainSingle()
    .Which
    .Should()
    .Be(timeProvider.GetUtcNow());
}

This test will run in nanoseconds and is deterministic.

Compare that to the similar test below for StuffService that needs to wait for 11 seconds before it can safely assert that the expectation has been met.

[Fact]
public async Task DoStuff_does_stuff_every_11_seconds()
{
  // Arrange
  var container = new List<DateTimeOffset>();
  var sut = new StuffService(container);

  // Act
  _ = sut.DoStuff(CancellationToken.None);
  await Task.Delay(TimeSpan.FromSeconds(11));

  // Assert
  container
    .Should()
    .ContainSingle()
    .Which
    .Should()
    .BeCloseTo(DateTimeOffset.UtcNow, precision: TimeSpan.FromMilliseconds(50));
}

Difference between ManualTimeProvider and FakeTimeProvider

The .NET team has published a similar test-specific time provider, the Microsoft.Extensions.Time.Testing.FakeTimeProvider.

The public API of both FakeTimeProvider and ManualTimeProvider are compatible, but there are some differences in when time is set before timer callbacks. Let's illustrate this with an example:

For example, if we create an ITimer with a due time and period set to 1 second, the DateTimeOffset returned from GetUtcNow() during the timer callback may be different depending on the amount passed to Advance() (or SetUtcNow()).

If we call Advance(TimeSpan.FromSeconds(1)) three times, effectively moving time forward by three seconds, the timer callback will be invoked once at times 00:01, 00:02, and 00:03, as illustrated in the drawing below. Both FakeTimeProvider and ManualTimeProvider behave like this:

If we instead call Advance(TimeSpan.FromSeconds(3)) once, the two implementations behave differently. ManualTimeProvider will invoke the timer callback at the same time (00:01, 00:02, and 00:03) as if we had called Advance(TimeSpan.FromSeconds(1)) three times, as illustrated in the drawing below:

However, FakeTimeProvider will invoke the timer callback at time 00:03 three times, as illustrated in the drawings below:

Technically, both implementations are correct since the ITimer abstractions only promise to invoke the callback timer on or after the due time/period has elapsed, never before.

However, I strongly prefer the ManualTimeProvider approach since it behaves consistently independent of how time is moved forward. It seems much more in the spirit of how a deterministic time provider should behave and avoids users being surprised when writing tests. I imagine users may get stuck for a while trying to debug why the time reported by GetUtcNow() is not set as expected due to the subtle difference in the behavior of FakeTimeProvider.

That said, it can be useful to test that your code behaves correctly if a timer isn't allocated processor time immediately when it's callback should fire, and for that, ManualTimeProvider includes a different method, Jump.

Jumping to a point in time

A real ITimer's callback may not be allocated processor time and be able to fire at the moment it has been scheduled, e.g. if the processor is busy doing other things. The callback will eventually fire (unless the timer is disposed of).

To support testing this scenario, ManualtTimeProvider includes a method that will jump time to a specific point, and then invoke all scheduled timer callbacks between the start and end of the jump. This behavior is similar to how FakeTimeProviders Advance method works, as described in the previous section.