Search

Chapter 44. Implementing a Component

download PDF

Abstract

This chapter provides a general overview of the approaches can be used to implement a Apache Camel component.

44.1. Component Architecture

44.1.1. Factory Patterns for a Component

Overview

A Apache Camel component consists of a set of classes that are related to each other through a factory pattern. The primary entry point to a component is the Component object itself (an instance of org.apache.camel.Component type). You can use the Component object as a factory to create Endpoint objects, which in turn act as factories for creating Consumer, Producer, and Exchange objects. These relationships are summarized in Figure 44.1, “Component Factory Patterns”

Figure 44.1. Component Factory Patterns

Component factory patterns

Component

A component implementation is an endpoint factory. The main task of a component implementor is to implement the Component.createEndpoint() method, which is responsible for creating new endpoints on demand.
Each kind of component must be associated with a component prefix that appears in an endpoint URI. For example, the file component is usually associated with the file prefix, which can be used in an endpoint URI like file://tmp/messages/input. When you install a new component in Apache Camel, you must define the association between a particular component prefix and the name of the class that implements the component.

Endpoint

Each endpoint instance encapsulates a particular endpoint URI. Every time Apache Camel encounters a new endpoint URI, it creates a new endpoint instance. An endpoint object is also a factory for creating consumer endpoints and producer endpoints.
Endpoints must implement the org.apache.camel.Endpoint interface. The Endpoint interface defines the following factory methods:
  • createConsumer() and createPollingConsumer()—Creates a consumer endpoint, which represents the source endpoint at the beginning of a route.
  • createProducer()—Creates a producer endpoint, which represents the target endpoint at the end of a route.
  • createExchange()—Creates an exchange object, which encapsulates the messages passed up and down the route.

Consumer

Consumer endpoints consume requests. They always appear at the start of a route and they encapsulate the code responsible for receiving incoming requests and dispatching outgoing replies. From a service-oriented prospective a consumer represents a service.
Consumers must implement the org.apache.camel.Consumer interface. There are a number of different patterns you can follow when implementing a consumer. These patterns are described in Section 44.1.3, “Consumer Patterns and Threading”.

Producer

Producer endpoints produce requests. They always appears at the end of a route and they encapsulate the code responsible for dispatching outgoing requests and receiving incoming replies. From a service-oriented prospective a producer represents a service consumer.
Producers must implement the org.apache.camel.Producer interface. You can optionally implement the producer to support an asynchronous style of processing. See Section 44.1.4, “Asynchronous Processing” for details.

Exchange

Exchange objects encapsulate a related set of messages. For example, one kind of message exchange is a synchronous invocation, which consists of a request message and its related reply.
Exchanges must implement the org.apache.camel.Exchange interface. The default implementation, DefaultExchange, is sufficient for many component implementations. However, if you want to associated extra data with the exchanges or have the exchanges preform additional processing, it can be useful to customize the exchange implementation.

Message

There are two different message slots in an Exchange object:
  • In message—holds the current message.
  • Out message—temporarily holds a reply message.
All of the message types are represented by the same Java object, org.apache.camel.Message. It is not always necessary to customize the message implementation—the default implementation, DefaultMessage, is usually adequate.

44.1.2. Using a Component in a Route

Overview

A Apache Camel route is essentially a pipeline of processors, of org.apache.camel.Processor type. Messages are encapsulated in an exchange object, E, which gets passed from node to node by invoking the process() method. The architecture of the processor pipeline is illustrated in Figure 44.2, “Consumer and Producer Instances in a Route”.

Figure 44.2. Consumer and Producer Instances in a Route

Consumer and Producer instances in a route

Source endpoint

At the start of the route, you have the source endpoint, which is represented by an org.apache.camel.Consumer object. The source endpoint is responsible for accepting incoming request messages and dispatching replies. When constructing the route, Apache Camel creates the appropriate Consumer type based on the component prefix from the endpoint URI, as described in Section 44.1.1, “Factory Patterns for a Component”.

Processors

Each intermediate node in the pipeline is represented by a processor object (implementing the org.apache.camel.Processor interface). You can insert either standard processors (for example, filter, throttler, or delayer) or insert your own custom processor implementations.

Target endpoint

At the end of the route is the target endpoint, which is represented by an org.apache.camel.Producer object. Because it comes at the end of a processor pipeline, the producer is also a processor object (implementing the org.apache.camel.Processor interface). The target endpoint is responsible for sending outgoing request messages and receiving incoming replies. When constructing the route, Apache Camel creates the appropriate Producer type based on the component prefix from the endpoint URI.

44.1.3. Consumer Patterns and Threading

Overview

The pattern used to implement the consumer determines the threading model used in processing the incoming exchanges. Consumers can be implemented using one of the following patterns:

Event-driven pattern

In the event-driven pattern, the processing of an incoming request is initiated when another part of the application (typically a third-party library) calls a method implemented by the consumer. A good example of an event-driven consumer is the Apache Camel JMX component, where events are initiated by the JMX library. The JMX library calls the handleNotification() method to initiate request processing—see Example 47.4, “JMXConsumer Implementation” for details.
Figure 44.3, “Event-Driven Consumer” shows an outline of the event-driven consumer pattern. In this example, it is assumed that processing is triggered by a call to the notify() method.

Figure 44.3. Event-Driven Consumer

message chain using an event-driven consumer
The event-driven consumer processes incoming requests as follows:
  1. The consumer must implement a method to receive the incoming event (in Figure 44.3, “Event-Driven Consumer” this is represented by the notify() method). The thread that calls notify() is normally a separate part of the application, so the consumer's threading policy is externally driven.
    For example, in the case of the JMX consumer implementation, the consumer implements the NotificationListener.handleNotification() method to receive notifications from JMX. The threads that drive the consumer processing are created within the JMX layer.
  2. In the body of the notify() method, the consumer first converts the incoming event into an exchange object, E, and then calls process() on the next processor in the route, passing the exchange object as its argument.

Scheduled poll pattern

In the scheduled poll pattern, the consumer retrieves incoming requests by checking at regular time intervals whether or not a request has arrived. Checking for requests is scheduled automatically by a built-in timer class, the scheduled executor service, which is a standard pattern provided by the java.util.concurrent library. The scheduled executor service executes a particular task at timed intervals and it also manages a pool of threads, which are used to run the task instances.
Figure 44.4, “Scheduled Poll Consumer” shows an outline of the scheduled poll consumer pattern.

Figure 44.4. Scheduled Poll Consumer

Scheduled Poll Consumer
The scheduled poll consumer processes incoming requests as follows:
  1. The scheduled executor service has a pool of threads at its disposal, that can be used to initiate consumer processing. After each scheduled time interval has elapsed, the scheduled executor service attempts to grab a free thread from its pool (there are five threads in the pool by default). If a free thread is available, it uses that thread to call the poll() method on the consumer.
  2. The consumer's poll() method is intended to trigger processing of an incoming request. In the body of the poll() method, the consumer attempts to retrieve an incoming message. If no request is available, the poll() method returns immediately.
  3. If a request message is available, the consumer inserts it into an exchange object and then calls process() on the next processor in the route, passing the exchange object as its argument.

Polling pattern

In the polling pattern, processing of an incoming request is initiated when a third-party calls one of the consumer's polling methods:
  • receive()
  • receiveNoWait()
  • receive(long timeout)
It is up to the component implementation to define the precise mechanism for initiating calls on the polling methods. This mechanism is not specified by the polling pattern.
Figure 44.5, “Polling Consumer” shows an outline of the polling consumer pattern.

Figure 44.5. Polling Consumer

Polling Consumer
The polling consumer processes incoming requests as follows:
  1. Processing of an incoming request is initiated whenever one of the consumer's polling methods is called. The mechanism for calling these polling methods is implementation defined.
  2. In the body of the receive() method, the consumer attempts to retrieve an incoming request message. If no message is currently available, the behavior depends on which receive method was called.
    • receiveNoWait() returns immediately
    • receive(long timeout) waits for the specified timeout interval[3] before returning
    • receive() waits until a message is received
  3. If a request message is available, the consumer inserts it into an exchange object and then calls process() on the next processor in the route, passing the exchange object as its argument.

44.1.4. Asynchronous Processing

Overview

Producer endpoints normally follow a synchronous pattern when processing an exchange. When the preceding processor in a pipeline calls process() on a producer, the process() method blocks until a reply is received. In this case, the processor's thread remains blocked until the producer has completed the cycle of sending the request and receiving the reply.
Sometimes, however, you might prefer to decouple the preceding processor from the producer, so that the processor's thread is released immediately and the process() call does not block. In this case, you should implement the producer using an asynchronous pattern, which gives the preceding processor the option of invoking a non-blocking version of the process() method.
To give you an overview of the different implementation options, this section describes both the synchronous and the asynchronous patterns for implementing a producer endpoint.

Synchronous producer

Figure 44.6, “Synchronous Producer” shows an outline of a synchronous producer, where the preceding processor blocks until the producer has finished processing the exchange.

Figure 44.6. Synchronous Producer

Synchronous Producer
The synchronous producer processes an exchange as follows:
  1. The preceding processor in the pipeline calls the synchronous process() method on the producer to initiate synchronous processing. The synchronous process() method takes a single exchange argument.
  2. In the body of the process() method, the producer sends the request (In message) to the endpoint.
  3. If required by the exchange pattern, the producer waits for the reply (Out message) to arrive from the endpoint. This step can cause the process() method to block indefinitely. However, if the exchange pattern does not mandate a reply, the process() method can return immediately after sending the request.
  4. When the process() method returns, the exchange object contains the reply from the synchronous call (an Out message message).

Asynchronous producer

Figure 44.7, “Asynchronous Producer” shows an outline of an asynchronous producer, where the producer processes the exchange in a sub-thread, and the preceding processor is not blocked for any significant length of time.

Figure 44.7. Asynchronous Producer

Asynchronous Producer
The asynchronous producer processes an exchange as follows:
  1. Before the processor can call the asynchronous process() method, it must create an asynchronous callback object, which is responsible for processing the exchange on the return portion of the route. For the asynchronous callback, the processor must implement a class that inherits from the AsyncCallback interface.
  2. The processor calls the asynchronous process() method on the producer to initiate asynchronous processing. The asynchronous process() method takes two arguments:
    • an exchange object
    • a synchronous callback object
  3. In the body of the process() method, the producer creates a Runnable object that encapsulates the processing code. The producer then delegates the execution of this Runnable object to a sub-thread.
  4. The asynchronous process() method returns, thereby freeing up the processor's thread. The exchange processing continues in a separate sub-thread.
  5. The Runnable object sends the In message to the endpoint.
  6. If required by the exchange pattern, the Runnable object waits for the reply (Out or Fault message) to arrive from the endpoint. The Runnable object remains blocked until the reply is received.
  7. After the reply arrives, the Runnable object inserts the reply (Out message) into the exchange object and then calls done() on the asynchronous callback object. The asynchronous callback is then responsible for processing the reply message (executed in the sub-thread).


[3] The timeout interval is typically specified in milliseconds.
Red Hat logoGithubRedditYoutubeTwitter

Learn

Try, buy, & sell

Communities

About Red Hat Documentation

We help Red Hat users innovate and achieve their goals with our products and services with content they can trust.

Making open source more inclusive

Red Hat is committed to replacing problematic language in our code, documentation, and web properties. For more details, see the Red Hat Blog.

About Red Hat

We deliver hardened solutions that make it easier for enterprises to work across platforms and environments, from the core datacenter to the network edge.

© 2024 Red Hat, Inc.