Hedging resilience strategy

About

Option(s):
- HedgingStrategyOptions<T>
Extension(s):
- AddHedging
Exception(s): -

The hedging reactive strategy enables the re-execution of the callback if the previous execution takes too long. This approach gives you the option to either run the original callback again or specify a new callback for subsequent hedged attempts. Implementing a hedging strategy can boost the overall responsiveness of the system. However, it's essential to note that this improvement comes at the cost of increased resource utilization. If low latency is not a critical requirement, you may find the retry strategy more appropriate.

This strategy also supports multiple concurrency modes to flexibly tailor the behavior for your own needs.

Note

Please do not start any background work when executing actions using the hedging strategy. This strategy can spawn multiple parallel tasks, and as a result multiple background tasks can be started.

Usage

// Hedging with default options.
// See https://www.pollydocs.org/strategies/hedging#defaults for defaults.
var optionsDefaults = new HedgingStrategyOptions<HttpResponseMessage>();

// A customized hedging strategy that retries up to 3 times if the execution
// takes longer than 1 second or if it fails due to an exception or returns an HTTP 500 Internal Server Error.
var optionsComplex = new HedgingStrategyOptions<HttpResponseMessage>
{
    ShouldHandle = new PredicateBuilder<HttpResponseMessage>()
        .Handle<SomeExceptionType>()
        .HandleResult(response => response.StatusCode == HttpStatusCode.InternalServerError),
    MaxHedgedAttempts = 3,
    Delay = TimeSpan.FromSeconds(1),
    ActionGenerator = static args =>
    {
        Console.WriteLine("Preparing to execute hedged action.");

        // Return a delegate function to invoke the original action with the action context.
        // Optionally, you can also create a completely new action to be executed.
        return () => args.Callback(args.ActionContext);
    }
};

// Subscribe to hedging events.
var optionsOnHedging = new HedgingStrategyOptions<HttpResponseMessage>
{
    OnHedging = static args =>
    {
        Console.WriteLine($"OnHedging: Attempt number {args.AttemptNumber}");
        return default;
    }
};

// Add a hedging strategy with a HedgingStrategyOptions<TResult> instance to the pipeline
new ResiliencePipelineBuilder<HttpResponseMessage>().AddHedging(optionsDefaults);

Defaults

Property	Default Value	Description
`ShouldHandle`	Any exceptions other than `OperationCanceledException`.	Defines a predicate to determine what results and/or exceptions are handled by the hedging strategy.
`MaxHedgedAttempts`	1	The maximum number of hedged actions to use, in addition to the original action.
`Delay`	2 seconds	The maximum waiting time before spawning a new hedged action.
`ActionGenerator`	It returns the original callback that was passed to this strategy.	This delegate allows you to dynamically calculate the hedged action by utilizing information that is only available at runtime (like the attempt number).
`DelayGenerator`	`null`	This optional delegate allows you to dynamically calculate the delay by utilizing information that is only available at runtime (like the attempt number).
`OnHedging`	`null`	If provided then it will be invoked before the strategy performs the hedged action.

You can use the following special values for Delay or in DelayGenerator:

0 seconds - the hedging strategy immediately creates a total of MaxHedgedAttempts and completes when the fastest acceptable result is available.
-1 millisecond - this value indicates that the strategy does not create a new hedged task before the previous one completes. This enables scenarios where having multiple concurrent hedged tasks can cause side effects.

Note

If both Delay and DelayGenerator are specified then Delay will be ignored.

Telemetry

The hedging strategy reports the following telemetry events:

Event Name	Event Severity	When?
`ExecutionAttempt`	`Information` / `Warning`	Just after the original/hedged action completes with success/failure
`OnHedging`	`Warning`	Just before the strategy calls the `OnHedging` delegate

Here are some sample events:

The reported Execution Attempt telemetry events' severity depends on the action's outcome:

If it succeeded then the severity is Information
It it failed then the severity is Warning

Resilience event occurred. EventName: 'OnHedging', Source: 'MyPipeline/MyPipelineInstance/Hedging', Operation Key: 'MyHedgingOperation', Result: ''

Execution attempt. Source: 'MyPipeline/MyPipelineInstance/Hedging', Operation Key: 'MyHedgingOperation', Result: '1', Handled: 'False', Attempt: '0', Execution Time: 1505.3839ms

Execution attempt. Source: 'MyPipeline/MyPipelineInstance/Hedging', Operation Key: 'MyHedgingOperation', Result: 'Exception of type 'CustomException' was thrown.', Handled: 'True', Attempt: '1', Execution Time: 1525.2899ms
    CustomException: Exception of type 'CustomException' was thrown.
        at Program.<>c.<<Main>b__0_2>d.MoveNext()
        ...
        at Polly.ResiliencePipeline.<>c__8`1.<<ExecuteAsync>b__8_0>d.MoveNext() in /_/src/Polly.Core/ResiliencePipeline.AsyncT.cs:line 95

Note

Please note that the OnHedging telemetry event will be reported only if the hedging strategy performs any hedged actions.

On the other hand the Execution attempt event will be always reported regardless whether the strategy has to perform hedging.

Also remember that Attempt: '0' relates to the original execution attempt.

For further information please check out the telemetry page.

Concurrency modes

In the sections below, explore the different concurrency modes available in the hedging strategy. The behavior is primarily controlled by the Delay property value.

Latency mode

When the Delay property is set to a value greater than zero, the hedging strategy operates in latency mode. In this mode, additional executions are triggered when the initial ones take too long to complete. By default, the Delay is set to 2 seconds.

The primary execution is initiated.
If the initial execution either fails or takes longer than the Delay to complete, a new execution is initiated.
If the first two executions fail or exceed the Delay (calculated from the last initiated execution), another execution is triggered.
Happy path: The final result is the result of fastest successful execution.
Unhappy path: If all executions fail, the final result will be the primary execution's failure.

Latency: happy path sequence diagram

The hedging strategy does not trigger because the response arrives faster than the threshold.

sequenceDiagram
    autonumber
    actor C as Caller
    participant P as Pipeline
    participant H as Hedging
    participant HUC as HedgedUserCallback

    C->>P: Calls ExecuteAsync
    P->>H: Calls ExecuteCore
    Note over H: Wait start

    activate H
    H->>+HUC: Invokes
    HUC-->>HUC: Processes request
    HUC->>-H: Returns result
    deactivate H

    Note over H: Wait end
    H->>P: Returns result
    P->>C: Returns result

Latency: unhappy path sequence diagram

The hedging strategy triggers because the response arrives slower than the threshold.

sequenceDiagram
    autonumber
    actor C as Caller
    participant P as Pipeline
    participant H as Hedging
    participant D as HedgedUserCallback

    C->>P: Calls ExecuteAsync
    P->>H: Calls ExecuteCore
    Note over H: Wait start

    activate H
    activate D
    H->>D: Invokes (R1)
    D-->>D: Processes R1<br/> slowly ...
    Note over H: Wait end

    H->>D: Invokes (R2)
    activate D
    D-->>D: Processes R2<br/> quickly
    D->>H: Returns result (R2)
    deactivate D
    deactivate D

    H->>D: Propagates cancellation (R1)
    activate D
    D-->>D: Cancellation of pending actions (R1)
    deactivate D
    D->>H: Cancellation finished (R1)

    deactivate H

    H->>P: Returns result (R2)
    P->>C: Returns result (R2)

Note

Notice that the hedging waits until all additional pending actions are cancelled before returning the accepted result. Therefore it's important for the hedged actions to respect the cancellation token passed to the execution. If the cancellation token is not correctly respected, the hedging is unnecessarily delayed.

Fallback mode

In fallback mode, the Delay value should be less than TimeSpan.Zero. This mode allows only a single execution to proceed at a given time.

An execution is initiated, and the strategy waits for its completion.
If the initial execution fails, new one is initiated.
Happy path: The final result will be the first successful execution.
Unhappy path: If all executions fail, the final result will be the primary execution's failure.

Fallback: happy path sequence diagram

The hedging strategy triggers because the first attempt fails. It succeeds because the retry attempts do not exceed the MaxHedgedAttempts.

sequenceDiagram
    autonumber
    actor C as Caller
    participant P as Pipeline
    participant H as Hedging
    participant HUC as HedgedUserCallback

    C->>P: Calls ExecuteAsync
    P->>H: Calls ExecuteCore
    activate H

    H->>+HUC: Invokes (R1)
    HUC-->>HUC: Processes R1
    HUC->>-H: Fails (R1)

    H->>+HUC: Invokes (R2)
    HUC-->>HUC: Processes R2
    HUC->>-H: Returns result (R2)

    deactivate H
    H->>P: Returns result (R2)
    P->>C: Returns result (R2)

Fallback: unhappy path sequence diagram

The hedging strategy triggers because the first attempt fails. It fails because all retry attempts failed as well.

sequenceDiagram
    autonumber
    actor C as Caller
    participant P as Pipeline
    participant H as Hedging
    participant HUC as HedgedUserCallback

    C->>P: Calls ExecuteAsync
    P->>H: Calls ExecuteCore
    activate H

    H->>+HUC: Invokes (R1)
    HUC-->>HUC: Processes R1
    HUC->>-H: Fails (R1)

    H->>+HUC: Invokes (R2)
    HUC-->>HUC: Processes R2
    HUC->>-H: Fails (R2)

    deactivate H
    H->>P: Propagates failure (R1)
    P->>C: Propagates failure (R1)

Parallel mode

The hedging strategy operates in parallel mode when the Delay property is set to TimeSpan.Zero. In this mode, all executions are initiated simultaneously, and the strategy waits for the fastest completion.

Important

Use this mode only when absolutely necessary, as it consumes the most resources, particularly when the hedging strategy uses remote resources such as remote HTTP services.

All executions are initiated simultaneously, adhering to the MaxHedgedAttempts limit.
Happy path: The final result will be the fastest successful execution.
Unhappy path: If all executions fail, the final result will be the primary execution's failure

Parallel: happy path sequence diagram

The hedging strategy triggers because the Delay is set to zero. It succeeds because one of the requests succeeds.

sequenceDiagram
    actor C as Caller
    participant P as Pipeline
    participant H as Hedging
    participant HUC as HedgedUserCallback

    C->>P: Calls ExecuteAsync
    P->>H: Calls ExecuteCore
    activate H

    par
    H->>HUC: Invokes (R1)
    activate HUC

    and
    H->>+HUC: Invokes (R2)

    end

    HUC-->>HUC: Processes R1<br/> slowly ...
    HUC-->>HUC: Processes R2<br/> quickly ...
    HUC->>-H: Returns result (R2)

    deactivate HUC

    H->>HUC: Propagates cancellation (R1)
    activate HUC
    HUC-->>HUC: Cancellation of pending actions (R1)
    deactivate HUC
    HUC->>H: Cancellation finished (R1)

    deactivate H
    H->>P: Returns result (R2)
    P->>C: Returns result (R2)

Note

Parallel: unhappy path sequence diagram

The hedging strategy triggers because the Delay is set to zero. It fails because all requests fail.

sequenceDiagram
    actor C as Caller
    participant P as Pipeline
    participant H as Hedging
    participant HUC as HedgedUserCallback

    C->>P: Calls ExecuteAsync
    P->>H: Calls ExecuteCore
    activate H

    par
    H->>HUC: Invokes (R1)
    activate HUC

    and
    H->>+HUC: Invokes (R2)

    end

    HUC-->>HUC: Processes R1
    HUC-->>HUC: Processes R2
    HUC->>-H: Fails (R2)
    HUC->>-H: Fails (R1)

    deactivate H
    H->>P: Propagates failure (R1)
    P->>C: Propagates failure (R1)

Dynamic mode

In dynamic mode, you have the flexibility to control how the hedging strategy behaves during each execution. This control is achieved through the DelayGenerator property.

Note

The Delay property is disregarded when DelayGenerator is set.

Example scenario:

First, initiate the first two executions in parallel mode.
Subsequently, switch to fallback mode for additional executions.

To configure hedging according to the above scenario, use the following code:

new ResiliencePipelineBuilder<HttpResponseMessage>()
    .AddHedging(new()
    {
        MaxHedgedAttempts = 3,
        DelayGenerator = args =>
        {
            var delay = args.AttemptNumber switch
            {
                0 or 1 => TimeSpan.Zero, // Parallel mode
                _ => TimeSpan.FromSeconds(-1) // switch to Fallback mode
            };

            return new ValueTask<TimeSpan>(delay);
        }
    });

With this configuration, the hedging strategy:

Initiates a maximum of 4 executions. This includes initial action and an additional 3 attempts.
Allows the first two executions to proceed in parallel, while the third and fourth executions follow the fallback mode.

Dynamic: happy path sequence diagram

The hedging strategy triggers and switches between modes due to our DelayGenerator. It succeeds because the last request succeeds.

sequenceDiagram
    actor C as Caller
    participant P as Pipeline
    participant H as Hedging
    participant DG as DelayGenerator
    participant HUC as HedgedUserCallback

    C->>P: Calls ExecuteAsync
    P->>H: Calls ExecuteCore
    activate H

    Note over H: Parallel mode
    par
    H->>DG: Gets delay
    DG->>H: 0 second
    H->>HUC: Invokes (R1)
    activate HUC
    and
    H ->> DG: Gets delay
    DG ->> H: 0 second
    H->>+HUC: Invokes (R2)
    end

    HUC-->>HUC: Processes R1
    HUC-->>HUC: Processes R2
    HUC->>-H: Fails (R2)
    HUC->>-H: Fails (R1)

    Note over H: Fallback mode

    H->>DG: Gets delay
    DG->>H: -1 second
    H->>+HUC: Invokes (R3)
    HUC-->>HUC: Processes R3
    HUC->>-H: Fails (R3)

    H->>DG: Gets delay
    DG->>H: -1 second
    H->>+HUC: Invokes (R4)
    HUC-->>HUC: Processes R4
    HUC->>-H: Returns result (R4)

    deactivate H
    H->>P: Returns result (R4)
    P->>C: Returns result (R4)

Dynamic: unhappy path sequence diagram

The hedging strategy triggers and switches between modes due our DelayGenerator. It fails because all requests fail.

sequenceDiagram
    actor C as Caller
    participant P as Pipeline
    participant H as Hedging
    participant DG as DelayGenerator
    participant HUC as HedgedUserCallback

    C->>P: Calls ExecuteAsync
    P->>H: Calls ExecuteCore
    activate H

    Note over H: Parallel mode
    par
    H->>DG: Gets delay
    DG->>H: 0 second
    H->>HUC: Invokes (R1)
    activate HUC
    and
    H->>DG: Gets delay
    DG->>H: 0 second
    H->>+HUC: Invokes (R2)
    end

    HUC-->>HUC: Processes R1
    HUC-->>HUC: Processes R2
    HUC->>-H: Fails (R2)
    HUC->>-H: Fails (R1)

    Note over H: Fallback mode

    H->>DG: Gets delay
    DG->>H: -1 second
    H->>+HUC: Invokes (R3)
    HUC-->>HUC: Processes R3
    HUC->>-H: Fails (R3)

    H -> DG: Gets delay
    DG->>H: -1 second
    H->>+HUC: Invokes (R4)
    HUC-->>HUC: Processes R4
    HUC->>-H: Fails (R4)

    deactivate H
    H->>P: Propagates failure (R3)
    P->>C: Propagates failure (R3)

Hedging and resilience context

The hedging strategy supports the concurrent execution and cancellation of multiple hedged actions. To safeguard against concurrency issues when these actions access the same ResilienceContext, the hedging strategy ensures that each hedged action has its own unique ResilienceContext. The hedging strategy distinguishes between two types of contexts:

Primary context: The original resilience context that the hedging strategy receives.
Action context: A deep copy of the primary context with a distinct cancellation token for a single hedged action.

Here's the flow:

The strategy gets the primary context and preserves it for deep-cloning.
After the strategy has an accepted result from a hedged action, the resilience context from the action is merged back into the primary context.
All ongoing hedged actions are cancelled and discarded. The hedging strategy awaits the propagation of cancellation.

Merging action context into the primary context

After merging, the primary context contains:

New properties created during the action context's execution.
Existing properties updated or soft-deleted (i.e. setting the value to null) during execution.

Hedging callbacks

Some insights about the hedging callback behavior in relation to hedging contexts:

ActionGenerator: This callback provides arguments with PrimaryContext and ActionContext properties. Use PrimaryContext to get values that remain consistent throughout the hedging strategy execution. Use ActionContext to save values specific to a single hedged action execution.
OnHedging: This callback, like the previous one, provides arguments with PrimaryContext and ActionContext properties and is invoked after ActionGenerator. Any changes to both PrimaryContext and ActionContext are preserved and visible here.

Note

The hedging strategy ensures that both ActionGenerator and OnHedging are never executed concurrently. Any modifications to the contexts are thread-safe in these callbacks.

Sequence diagram about contexts and callbacks

For the sake of conciseness the Caller is not depicted on the diagram.

sequenceDiagram
    autonumber
    participant P as Pipeline
    participant H as Hedging
    participant AG as ActionGenerator
    participant OH as OnHedging
    participant UC as UserCallback
    participant HUC as HedgedUserCallback

    P->>H: Calls ExecuteCore <br/>with Primary Context
    H-->>H: Deep clones <br/>Primary Context

    H->>+UC: Invokes <br/>with Action Context
    UC-->>UC: Processes <br/>+ Modifies Context
    UC->>-H: Fails

    H-->>H: Deep clones <br/>Primary Context
    H->>+AG: Invokes <br/>with both Contexts
    AG-->>AG: Executes callback <br/>+ Modifies Primary <br/> and / or Action Context
    AG->>-H: Returns factory

    H->>+OH: Invokes <br/>with both Contexts
    OH-->>OH: Executes callback <br/>+ Modifies Primary <br/> and / or Action Context
    OH->>-H: Finishes

    H-->>H: Invokes factory
    H->>+HUC: Invokes <br/>with Action Context
    HUC-->>HUC: Processes <br/>+ Modifies Context
    HUC->>-H: Fails

    H-->>H: Merges Action Context <br/>onto Primary Context
    H->>P: Propagates failure <br/>with Primary Context

Action generator

The hedging options include an ActionGenerator property, allowing you to customize the actions executed during hedging. By default, the ActionGenerator returns the original callback passed to the strategy. The original callback also includes any logic introduced by subsequent resilience strategies. For more advanced scenarios, the ActionGenerator can be used to return entirely new hedged actions, as demonstrated in the example below:

new ResiliencePipelineBuilder<HttpResponseMessage>()
    .AddHedging(new()
    {
        ActionGenerator = args =>
        {
            // You can access data from the original (primary) context here
            var customData = args.PrimaryContext.Properties.GetValue(customDataKey, "default-custom-data");

            Console.WriteLine($"Hedging, Attempt: {args.AttemptNumber}, Custom Data: {customData}");

            // Here, we can access the original callback and return it or return a completely new action
            var callback = args.Callback;

            // A function that returns a ValueTask<Outcome<HttpResponseMessage>> is required.
            return async () =>
            {
                try
                {
                    // A dedicated ActionContext is provided for each hedged action.
                    // It comes with a separate CancellationToken created specifically for this hedged attempt,
                    // which can be cancelled later if needed.
                    //
                    // Note that the "MyRemoteCallAsync" call won't have any additional resilience applied.
                    // You are responsible for wrapping it with any additional resilience pipeline.
                    var response = await MyRemoteCallAsync(args.ActionContext.CancellationToken);

                    return Outcome.FromResult(response);
                }
                catch (Exception e)
                {
                    // Note: All exceptions should be caught and converted to Outcome.
                    return Outcome.FromException<HttpResponseMessage>(e);
                }
            };
        }
    });

Action generator: sequence diagram

sequenceDiagram
    autonumber
    actor C as Caller
    participant P as Pipeline
    participant H as Hedging
    participant AG as ActionGenerator
    participant UC as UserCallback
    participant HUC as HedgedUserCallback

    C->>P: Calls ExecuteAsync
    P->>H: Calls ExecuteCore

    activate H
    H->>+UC: Invokes (R1)
    UC-->>UC: Processes R1
    UC->>-H: Fails (R1)
    H->>+AG: Invokes
    AG->>-H: Returns factory
    H-->>H: Invokes factory

    H->>+HUC: Invokes (R2)
    HUC-->>HUC: Processes R2
    HUC->>-H: Fails (R2)

    H-->>H: Invokes factory
    H->>+HUC: Invokes (R3)
    HUC-->>HUC: Processes R3
    HUC->>-H: Returns result (R3)
    deactivate H

    H->>P: Returns result (R3)
    P->>C: Returns result (R3)

Parameterized callbacks and action generator

When you have control over the callbacks that the resilience pipeline receives, you can parameterize them. This flexibility allows for reusing the callbacks within an action generator.

A common use case is with DelegatingHandler. Here, you can parameterize the HttpRequestMessage:

internal sealed class HedgingHandler : DelegatingHandler
{
    private readonly ResiliencePipeline<HttpResponseMessage> _pipeline;

    public HedgingHandler(ResiliencePipeline<HttpResponseMessage> pipeline)
    {
        _pipeline = pipeline;
    }

    protected override async Task<HttpResponseMessage> SendAsync(HttpRequestMessage request, CancellationToken cancellationToken)
    {
        var context = ResilienceContextPool.Shared.Get(cancellationToken);

        // Store the incoming request in the context
        context.Properties.Set(ResilienceKeys.RequestMessage, request);

        try
        {
            return await _pipeline.ExecuteAsync(async cxt =>
            {
                // Allow the pipeline to use request message that was stored in the context.
                // This allows replacing the request message with a new one in the resilience pipeline.
                request = cxt.Properties.GetValue(ResilienceKeys.RequestMessage, request);

                return await base.SendAsync(request, cxt.CancellationToken);
            },
            context);
        }
        finally
        {
            ResilienceContextPool.Shared.Return(context);
        }
    }
}

Where ResilienceKeys is defined as:

internal static class ResilienceKeys
{
    public static readonly ResiliencePropertyKey<HttpRequestMessage> RequestMessage = new("MyFeature.RequestMessage");
}

In your ActionGenerator, you can easily provide your own HttpRequestMessage to ActionContext, and the original callback will use it:

new ResiliencePipelineBuilder<HttpResponseMessage>()
    .AddHedging(new()
    {
        ActionGenerator = args =>
        {
            if (!args.PrimaryContext.Properties.TryGetValue(ResilienceKeys.RequestMessage, out var request))
            {
                throw new InvalidOperationException("The request message must be provided.");
            }

            // Prepare a new request message for the callback, potentially involving:
            //
            // - Cloning the request message
            // - Providing alternate endpoint URLs
            request = PrepareRequest(request);

            // Override the request message in the action context
            args.ActionContext.Properties.Set(ResilienceKeys.RequestMessage, request);

            // Then, execute the original callback
            return () => args.Callback(args.ActionContext);
        }
    });

Table of Contents

Hedging resilience strategy

About

Note

Usage

Defaults

Note

Telemetry

Note

Concurrency modes

Latency mode

Latency: happy path sequence diagram

Latency: unhappy path sequence diagram

Note

Fallback mode

Fallback: happy path sequence diagram

Fallback: unhappy path sequence diagram

Parallel mode

Important

Parallel: happy path sequence diagram

Note

Parallel: unhappy path sequence diagram

Dynamic mode

Note

Dynamic: happy path sequence diagram

Dynamic: unhappy path sequence diagram

Hedging and resilience context

Merging action context into the primary context

Hedging callbacks

Note

Sequence diagram about contexts and callbacks

Action generator

Action generator: sequence diagram

Parameterized callbacks and action generator