Skip to content

Latest commit

 

History

History
278 lines (198 loc) · 13.4 KB

provider_memory.md

File metadata and controls

278 lines (198 loc) · 13.4 KB
Raw

Memory (in-process) Provider for SlimMessageBus

Please read the Introduction before reading this provider documentation.

Introduction

The Memory transport provider can be used for internal communication within the same process. It is the simplest transport provider and does not require any external messaging infrastructure.

Since messages are passed in memory and never persisted, they will be lost if the application process terminates while consuming these messages.

Good use case for in memory communication is:

  • to integrate the domain layer with other application layers via domain events pattern,
  • to implement mediator pattern (when combined with interceptors),
  • to run unit tests against application code that normally runs with an out of process transport provider (Kafka, Azure Service Bus, etc),
  • to start simple messaging without having to provision messaging infrastructure, but when time comes reconfigure SMB to leverage messaging infrastructure.

Configuration

The memory transport is configured using the .WithProviderMemory():

using SlimMessageBus.Host.Memory;

services.AddSlimMessageBus(mbb =>
{
  // Bus configuration happens here (...)
  mbb.WithProviderMemory(); // requires SlimMessageBus.Host.Memory package
});

Serialization

Since messages are passed within the same process, serializing and deserializing them is redundant. Also, disabling serialization gives a performance improvement.

Serialization is disabled by default for memory bus.

Serialization can be disabled or enabled:

services.AddSlimMessageBus(mbb =>
{
   // Bus configuration happens here (...)
   mbb.WithProviderMemory(cfg =>
   {
      // Serialize the domain events instead of passing the same instance across to handlers/consumers
      cfg.EnableMessageSerialization = true
   });
   // Serializer not needed if EnableMessageSerialization = false
   mbb.AddJsonSerializer();
});

When serialization is disabled for in memory passed messages, the exact same object instance send by the producer will be received by the consumer. Therefore state changes on the consumer end will be visible by the producer. Consider making the messages immutable (read only) in that case.

Headers

The headers published to the memory bus are delivered to the consumer. This is managed by the EnableMessageHeaders setting, which is enabled by default.

services.AddSlimMessageBus(mbb =>
{
   mbb.WithProviderMemory(cfg =>
   {
      // Header passing can be disabled when not needed (and to save on memory allocations)
      cfg.EnableMessageHeaders = false;
   });
});

Before version v2.5.1, to enable header passing the serialization had to be enabled.

Virtual Topics

Unlike other transport providers, memory transport does not have true notion of topics (or queues). However, it is still required to use topic names. This is required, so that the bus knows on which virtual topic to deliver the message to, and from what virtual topic to consume from.

The consumer configuration side should use .Topic() to set the virtual topic name:

// declare that OrderSubmittedEvent will be consumed
mbb.Consume<OrderSubmittedEvent>(x => x.Topic(x.MessageType.Name).WithConsumer<OrderSubmittedHandler>());

// alternatively
mbb.Consume<OrderSubmittedEvent>(x => x.Topic("OrderSubmittedEvent").WithConsumer<OrderSubmittedHandler>());

The producer configuration side should use .DefaultTopic() to set the virtual topic name:

mbb.Produce<OrderSubmittedEvent>(x => x.DefaultTopic("OrderSubmittedEvent"));

The virtual topic name can be any string. It helps to connect the relevant producers and consumers together.

Auto Declaration

Since 1.19.1

For bus configuration, we can leverage .AutoDeclareFrom() method to discover all the consumers (IConsumer<T>) and handlers (IRequestHandler<T,R>) types in an assembly and auto declare the respective producers and consumers/handlers in the bus. This can be useful to auto declare all of the domain event handlers in an application layer.

mbb
   .WithProviderMemory()
   .AutoDeclareFrom(Assembly.GetExecutingAssembly());
   // If we want to filter to specific consumer/handler types then we can supply an additional filter:
   //.AutoDeclareFrom(Assembly.GetExecutingAssembly(), consumerTypeFilter: (consumerType) => consumerType.Name.EndsWith("Handler"));

For example, assuming this is the discovered handler type:

public class EchoRequestHandler : IRequestHandler<EchoRequest, EchoResponse>
{
   public Task<EchoResponse> OnHandle(EchoRequest request) { /* ... */ }
}

The bus auto registrations will set-up the producer and handler to the equivalent:

mbb.Produce<EchoRequest>(x => x.DefaultTopic(x.MessageType.Name));
mbb.Handle<EchoRequest, EchoResponse>(x => x.Topic(x.MessageType.Name).WithConsumer<EchoRequestHandler>());

The virtual topic name will be derived from the message type name by default. This can be customized by passing an additional parameter to the AutoDeclareFrom() method.

Using AutoDeclareFrom() to configure the memory bus is recommended, as it will declare the producers and consumers automatically as consumer types are added over time.

Note that it is still required to register (or auto register) the consumer/handler types in the underlying DI (see here).

Polymorphic message support

Since 1.21.0

The polymorphic message types (message that share a common ancestry) are supported by the AutoDeclareFrom():

  • for every consumer / handler implementation found it analyzes the message type inheritance tree,
  • it declares a producer in the bus for the oldest ancestor of the message type hierarchy,
  • it declares a consumer in the bus for the oldest ancestor message type and within that, configures a consumer for derived message type,
  • topic names are derived from the ancestor message type.

Lifecycle

Concurrency and Ordering

The order of message delivery to consumer will match the order of producer.

By default, each consumer processes one message at a time to ensure ordering. To increase the concurrency level use the .Instances(n) setting:

mbb.Consume<PingMessage>(x => x.Instances(2));

When number of concurrent consumer instances > 0 (.Instances(N)) then up to N messages will be processed concurrently (having impact on ordering).

Blocking Publish

Similar to Send<T>(), the Publish<T>() is blocking by default.

It might be expected that the Publish<T>() to be non-blocking (asynchronous) and that the consumer processes the message in the background. However, blocking mode is enabled by default to ensure the consumer has finished processing by the time the method call returns. Often we want all the side effect to finish within the unit of work (ongoing web-request, or external message being handled). That behavior is optimized for domain-events where the side effects are to be executed synchronously within the unit of work.

Asynchronous Publish

Since 2.3.0

To use non-blocking publish use the EnableBlockingPublish property setting:

services.AddSlimMessageBus(mbb =>
{
    mbb
        .WithProviderMemory(cfg =>
        {
            cfg.EnableMessageSerialization = _enableSerialization;
            cfg.EnableBlockingPublish = _enableBlockingPublish;
        })
        .AddServicesFromAssemblyContaining<PingConsumer>()
        .AddJsonSerializer();
});

When the .Publish<T>() is invoked in the non-blocking mode:

  • the consumers will be executed in another async task (in the background),
  • that task cancellation token will be bound to the message bus lifecycle (consumers are stopped, the bus is disposed or application shuts down),
  • the order of message delivery to consumer will match the order of publish,
  • however, when number of concurrent consumer instances > 0 (.Instances(N)) then up to N messages will be processed concurrently (having impact on ordering)
  • the unit of work where the message is published is decoupled from the consumer unit of work, there will be an independent per message scope created in the DI on the consumer side - this allows to scope consumer dependencies to the message unit of work.

Error Handling

The exceptions raised in a handler (IRequestHandler<T, R>) are bubbled up to the sender (await bus.Send(request)). Likewise, the exceptions raised in a consumer (IConsumer<T>) are also bubbled up to the publisher (await bus.Publish<T>(message)) when blocking publish mode is enabled.

In the case of non-blocking publish mode when an exception is raised by a message consumer, the memory transport will log the exception and move on to the next message. For more elaborate behavior a custom error handler should be setup.

There is a memory transport specific error handler interface IMemoryConsumerErrorHandler that will be preferred by the memory transport.

The error handler has to be registered in MSDI for all (or specified) message types:

// Register error handler in MSDI for any message type
services.AddTransient(typeof(IMemoryConsumerErrorHandler<>), typeof(CustomMemoryConsumerErrorHandler<>));

See also the common error handling.

Per-Message DI scope

Unlike stated for the Introduction the memory bus has per-message scoped disabled by default. However, the memory bus consumer/handler would join (enlist in) the already ongoing DI scope. This is desired in scenarios where an external message is being handled as part of the unit of work (Kafka/ASB, etc) or an API HTTP request is being handled - the memory bus consumer/handler will be looked up in the ongoing/current DI scope.

Benchmarks

The project SlimMessageBus.Host.Memory.Benchmark runs basic scenarios for pub/sub and request/response on the memory bus. The benchmark should provide some reference about the speed / performance of the Memory Bus.

  • The test includes a 1M of simple messages being produced.
  • The consumers do not do any logic - we want to test how fast can the messages flow through the bus.
  • The benchmark application uses a real life setup including dependency injection container.
  • There is a variation of the test that captures the overhead for the interceptor pipeline.

Pub/Sub scenario results:

Method messageCount Mean Error StdDev Gen0 Gen1 Gen2 Allocated
PubSub 1000000 1.201 s 0.0582 s 0.0816 s 118000.0000 3000.0000 3000.0000 489.03 MB
PubSubWithConsumerInterceptor 1000000 1.541 s 0.0247 s 0.0219 s 191000.0000 3000.0000 3000.0000 778.95 MB
PubSubWithProducerInterceptor 1000000 1.479 s 0.0094 s 0.0078 s 200000.0000 3000.0000 3000.0000 817.09 MB
PubSubWithPublishInterceptor 1000000 1.511 s 0.0178 s 0.0219 s 200000.0000 3000.0000 3000.0000 817.09 MB

Pub/Sub rate is 832639 messages/s on the tested machine (without interceptors).

Request/Response scenario results:

Method messageCount Mean Error StdDev Gen0 Gen1 Gen2 Allocated
RequestResponse 1000000 2.424 s 0.0351 s 0.0274 s 134000.0000 39000.0000 6000.0000 801.83 MB
ReqRespWithConsumerInterceptor 1000000 3.056 s 0.0608 s 0.0769 s 205000.0000 55000.0000 6000.0000 1229.08 MB
ReqRespWithProducerInterceptor 1000000 2.957 s 0.0517 s 0.0458 s 229000.0000 60000.0000 6000.0000 1374.04 MB
ReqRespWithRequestHandlerInterceptor 1000000 3.422 s 0.0644 s 0.0742 s 217000.0000 58000.0000 6000.0000 1297.74 MB
ReqRespWithSendInterceptor 1000000 2.934 s 0.0285 s 0.0223 s 219000.0000 59000.0000 7000.0000 1305.38 MB

Request/Response rate is 412541 messages/s on the tested machine (without interceptors).

The test results are for the following environment:

BenchmarkDotNet=v0.13.2, OS=Windows 11 (10.0.22621.819)
Intel Core i7-8550U CPU 1.80GHz (Kaby Lake R), 1 CPU, 8 logical and 4 physical cores
.NET SDK=6.0.402
  [Host]     : .NET 6.0.11 (6.0.1122.52304), X64 RyuJIT AVX2
  Job-XKUBHP : .NET 6.0.11 (6.0.1122.52304), X64 RyuJIT AVX2

See the benchmark source here.