akka-net-best-practices（akka 网络最佳实践）

Akka.NET Best Practices

When to Use This Skill

Use this skill when:

• Designing actor communication patterns
• Deciding between EventStream and DistributedPubSub
• Implementing error handling in actors
• Understanding supervision strategies
• Choosing between Props patterns and DependencyResolver
• Designing work distribution across nodes
• Creating testable actor systems that can run with or without cluster infrastructure
• Abstracting over Cluster Sharding for local testing scenarios

Reference Files

• work-distribution-patterns.md: Database queues, Akka.Streams throttling, outbox pattern
• cluster-local-abstractions.md: GenericChildPerEntityParent, IPubSubMediator, execution mode wiring
• async-cancellation-patterns.md: Actor-scoped CancellationToken, linked CTS, timeout handling

1. EventStream vs DistributedPubSub

Critical: EventStream is LOCAL ONLY

Context.System.EventStream is local to a single ActorSystem process. It does NOT work across cluster nodes.

// BAD: This only works on a single server
// When you add a second server, subscribers on server 2 won't receive events from server 1
Context.System.EventStream.Subscribe(Self, typeof(PostCreated));
Context.System.EventStream.Publish(new PostCreated(postId, authorId));

When EventStream is appropriate:

• Logging and diagnostics within a single process
• Local event bus for truly single-process applications
• Development/testing scenarios

Use DistributedPubSub for Multi-Node

For events that must reach actors across multiple cluster nodes, use Akka.Cluster.Tools.PublishSubscribe:

using Akka.Cluster.Tools.PublishSubscribe;

public class TimelineUpdatePublisher : ReceiveActor
{
    private readonly IActorRef _mediator;

    public TimelineUpdatePublisher()
    {
        // Get the DistributedPubSub mediator
        _mediator = DistributedPubSub.Get(Context.System).Mediator;

        Receive<PublishTimelineUpdate>(msg =>
        {
            // Publish to a topic - reaches all subscribers across all nodes
            _mediator.Tell(new Publish($"timeline:{msg.UserId}", msg.Update));
        });
    }
}

Akka.Hosting Configuration for DistributedPubSub

builder.WithDistributedPubSub(role: null); // Available on all roles, or specify a role

Topic Design Patterns

2. Supervision Strategies

Key Clarification: Supervision is for CHILDREN

A supervision strategy defined on an actor dictates how that actor supervises its children, NOT how the actor itself is supervised.

public class ParentActor : ReceiveActor
{
    // This strategy applies to children of ParentActor, NOT to ParentActor itself
    protected override SupervisorStrategy SupervisorStrategy()
    {
        return new OneForOneStrategy(
            maxNrOfRetries: 10,
            withinTimeRange: TimeSpan.FromSeconds(30),
            decider: ex => ex switch
            {
                ArithmeticException => Directive.Resume,
                NullReferenceException => Directive.Restart,
                ArgumentException => Directive.Stop,
                _ => Directive.Escalate
            });
    }
}

Default Supervision Strategy

The default OneForOneStrategy already includes rate limiting:

• 10 restarts within 1 second = actor is permanently stopped
• This prevents infinite restart loops

You rarely need a custom strategy unless you have specific requirements.

When to Define Custom Supervision

Good reasons:

• Actor throws exceptions indicating irrecoverable state corruption -> Restart
• Actor throws exceptions that should NOT cause restart (expected failures) -> Resume
• Child failures should affect siblings -> Use AllForOneStrategy
• Need different retry limits than the default

Bad reasons:

• "Just to be safe" - the default is already safe
• Don't understand what the actor does - understand it first

3. Error Handling: Supervision vs Try-Catch

When to Use Try-Catch (Most Cases)

Use try-catch when:

• The failure is expected (network timeout, invalid input, external service down)
• You know exactly why the exception occurred
• You can handle it gracefully (retry, return error response, log and continue)
• Restarting would not help (same error would occur again)

public class RssFeedPollerActor : ReceiveActor
{
    public RssFeedPollerActor()
    {
        ReceiveAsync<PollFeed>(async msg =>
        {
            try
            {
                var feed = await _httpClient.GetStringAsync(msg.FeedUrl);
                var items = ParseFeed(feed);
                // Process items...
            }
            catch (HttpRequestException ex)
            {
                // Expected failure - log and schedule retry
                _log.Warning("Feed {Url} unavailable: {Error}", msg.FeedUrl, ex.Message);
                Context.System.Scheduler.ScheduleTellOnce(
                    TimeSpan.FromMinutes(5), Self, msg, Self);
            }
            catch (XmlException ex)
            {
                // Invalid feed format - log and mark as bad
                _log.Error("Feed {Url} has invalid format: {Error}", msg.FeedUrl, ex.Message);
                Sender.Tell(new FeedPollResult.InvalidFormat(msg.FeedUrl));
            }
        });
    }
}

When to Let Supervision Handle It

Let exceptions propagate (trigger supervision) when:

• You have no idea why the exception occurred
• The actor's state might be corrupt
• A restart would help (fresh state, reconnect resources)
• It's a programming error (NullReferenceException, InvalidOperationException from bad logic)

Anti-Pattern: Swallowing Unknown Exceptions

// BAD: Swallowing exceptions hides problems
catch (Exception ex)
{
    _log.Error(ex, "Error processing work");
    // Actor continues with potentially corrupt state
}

// GOOD: Handle known exceptions, let unknown ones propagate
catch (HttpRequestException ex)
{
    // Known, expected failure - handle gracefully
    _log.Warning("HTTP request failed: {Error}", ex.Message);
    Sender.Tell(new WorkResult.TransientFailure());
}
// Unknown exceptions propagate to supervision

4. Props vs DependencyResolver

When to Use Plain Props

Use Props.Create() when:

• Actor doesn't need IServiceProvider or IRequiredActor<T>
• All dependencies can be passed via constructor
• Actor is simple and self-contained

// Simple actor with no DI needs
public static Props Props(PostId postId, IPostWriteStore store)
    => Akka.Actor.Props.Create(() => new PostEngagementActor(postId, store));

When to Use DependencyResolver

Use resolver.Props<T>() when:

• Actor needs IServiceProvider to create scoped services
• Actor uses IRequiredActor<T> to get references to other actors
• Actor has many dependencies that are already in DI container

// Registration with DI
builder.WithActors((system, registry, resolver) =>
{
    var actor = system.ActorOf(resolver.Props<OrderProcessorActor>(), "order-processor");
    registry.Register<OrderProcessorActor>(actor);
});

Remote Deployment Considerations

You almost never need remote deployment. If you're not doing remote deployment (and you probably aren't):

• Props.Create(() => new Actor(...)) with closures is fine
• The "serialization issue" warning doesn't apply

For most applications, use cluster sharding instead of remote deployment - it handles distribution automatically.

5. Work Distribution Patterns

When you have many background jobs (RSS feeds, email sending, etc.), don't process them all at once - this causes thundering herd problems.

Three patterns to solve this:

1. Database-Driven Work Queue - Use FOR UPDATE SKIP LOCKED for natural cross-node distribution
2. Akka.Streams Rate Limiting - Throttle processing within a single node
3. Durable Queue (Outbox Pattern) - Database-backed outbox for reliable processing

See work-distribution-patterns.md for full code samples.

6. Common Mistakes Summary

7. Quick Reference

Communication Pattern Decision Tree

Need to communicate between actors?
├── Same process only? -> EventStream is fine
├── Across cluster nodes?
│   ├── Point-to-point? -> Use ActorSelection or known IActorRef
│   └── Pub/sub? -> Use DistributedPubSub
└── Fire-and-forget to external system? -> Consider outbox pattern

Error Handling Decision Tree

Exception occurred in actor?
├── Expected failure (HTTP timeout, invalid input)?
│   └── Try-catch, handle gracefully, continue
├── State might be corrupt?
│   └── Let supervision restart
├── Unknown cause?
│   └── Let supervision restart
└── Programming error (null ref, bad logic)?
    └── Let supervision restart, fix the bug

Props Decision Tree

Creating actor Props?
├── Actor needs IServiceProvider?
│   └── Use resolver.Props<T>()
├── Actor needs IRequiredActor<T>?
│   └── Use resolver.Props<T>()
├── Simple actor with constructor params?
│   └── Use Props.Create(() => new Actor(...))
└── Remote deployment needed?
    └── Probably not - use cluster sharding instead

8. Cluster/Local Mode Abstractions

For applications that need to run both in clustered production and local/test environments, use abstraction patterns to toggle between implementations:

• AkkaExecutionMode enum - Controls which implementations are used (LocalTest vs Clustered)
• GenericChildPerEntityParent - Mimics sharding behavior locally using the same IMessageExtractor
• IPubSubMediator - Abstracts DistributedPubSub for swappable local/cluster implementations

See cluster-local-abstractions.md for complete implementation code.

9. Actor Logging

Use ILoggingAdapter, Not ILogger

In actors, use ILoggingAdapter from Context.GetLogger() instead of DI-injected ILogger<T>:

public class MyActor : ReceiveActor
{
    private readonly ILoggingAdapter _log = Context.GetLogger();

    public MyActor()
    {
        Receive<MyMessage>(msg =>
        {
            _log.Info("Processing message for user {UserId}", msg.UserId);
            _log.Error(ex, "Failed to process {MessageType}", msg.GetType().Name);
        });
    }
}

Why ILoggingAdapter:

• Integrates with Akka's logging pipeline and supervision
• Supports semantic/structured logging as of v1.5.57
• Method names: Info(), Debug(), Warning(), Error() (not Log* variants)
• No DI required - obtained directly from actor context

Don't inject ILogger into actors - it bypasses Akka's logging infrastructure.

Semantic Logging (v1.5.57+)

// Named placeholders for better log aggregation and querying
_log.Info("Order {OrderId} processed for customer {CustomerId}", order.Id, order.CustomerId);

// Prefer named placeholders over positional
// Good: {OrderId}, {CustomerId}
// Avoid: {0}, {1}

10. Managing Async Operations with CancellationToken

When actors launch async operations via PipeTo, those operations can outlive the actor if not properly managed. Key practices:

• Actor CTS in PostStop - Always cancel and dispose in PostStop()
• New CTS per operation - Cancel previous before starting new work
• Pass token everywhere - EF Core queries, HTTP calls, etc.
• Linked CTS for timeouts - External calls get short timeouts to prevent hanging
• Graceful handling - Distinguish timeout vs shutdown in catch blocks

See async-cancellation-patterns.md for complete implementation code.