Reaktive Systeme in Java

1. Einleitung

In diesem Tutorial werden die Grundlagen zum Erstellen reaktiver Systeme in Java mit Spring und anderen Tools und Frameworks erläutert.

Dabei werden wir diskutieren, wie reaktive Programmierung nur ein Treiber für die Schaffung eines reaktiven Systems ist. Dies wird uns helfen, die Gründe für die Erstellung reaktiver Systeme und verschiedene Spezifikationen, Bibliotheken und Standards zu verstehen, die es auf diesem Weg inspiriert hat.

2. Was sind reaktive Systeme?

In den letzten Jahrzehnten hat die Technologielandschaft mehrere Störungen erfahren, die zu einer vollständigen Veränderung der Art und Weise geführt haben, wie wir Wert in der Technologie sehen. Die Computerwelt vor dem Internet hätte sich niemals vorstellen können, wie und mit welchen Mitteln sie unseren heutigen Tag verändern wird.

Angesichts der Reichweite des Internets für die Massen und der sich ständig weiterentwickelnden Erfahrung, die es verspricht, müssen Anwendungsarchitekten auf Trab sein, um ihre Anforderungen zu erfüllen.

Grundsätzlich bedeutet dies, dass wir eine Anwendung niemals so entwerfen können, wie wir es früher getan haben. Eine sehr reaktionsschnelle Anwendung ist kein Luxus mehr, sondern eine Notwendigkeit .

Auch das ist angesichts zufälliger Ausfälle und unvorhersehbarer Belastungen. Das Gebot der Stunde ist nicht nur, das richtige Ergebnis zu erzielen, sondern es auch schnell zu bekommen! Es ist sehr wichtig, die erstaunlichen Benutzererlebnisse zu fördern, die wir versprechen.

Dies schafft die Notwendigkeit eines Architekturstils, der uns reaktive Systeme geben kann.

2.1. Reaktives Manifest

Bereits im Jahr 2013 kam ein Entwicklerteam unter der Leitung von Jonas Boner zusammen, um eine Reihe von Grundprinzipien in einem Dokument zu definieren, das als Reactive Manifesto bekannt ist. Dies war der Grundstein für einen Architekturstil zur Schaffung reaktiver Systeme. Seitdem hat dieses Manifest großes Interesse bei den Entwicklern gefunden.

Grundsätzlich schreibt dieses Dokument das Rezept für ein reaktives System vor, das flexibel, lose gekoppelt und skalierbar ist . Dies macht solche Systeme einfach zu entwickeln, tolerant gegenüber Ausfällen und vor allem sehr reaktionsschnell. Dies ist die Grundlage für unglaubliche Benutzererlebnisse.

Was ist dieses Geheimrezept? Nun, es ist kaum ein Geheimnis! Das Manifest definiert die grundlegenden Merkmale oder Prinzipien eines reaktiven Systems:

  • Als Reaktion : Ein reaktives System eine schnelle und konsistente Reaktionszeit und damit eine gleichbleibende Qualität der Dienstleistung bieten sollte
  • Ausfallsicher : Ein reaktives System sollte bei zufälligen Fehlern durch Replikation und Isolation reaktionsfähig bleiben
  • Elastisch : Ein solches System sollte unter unvorhersehbaren Workloads durch kostengünstige Skalierbarkeit reaktionsfähig bleiben
  • Nachrichtengesteuert : Es sollte sich auf die asynchrone Nachrichtenübertragung zwischen Systemkomponenten stützen

Diese Prinzipien klingen einfach und sinnvoll, sind jedoch in komplexen Unternehmensarchitekturen nicht immer einfacher zu implementieren. In diesem Tutorial entwickeln wir ein Beispielsystem in Java unter Berücksichtigung dieser Prinzipien!

3. Was ist reaktive Programmierung?

Bevor wir fortfahren, ist es wichtig, den Unterschied zwischen reaktiver Programmierung und reaktiven Systemen zu verstehen. Wir verwenden beide Begriffe ziemlich oft und verstehen sie leicht falsch. Wie wir bereits gesehen haben, sind reaktive Systeme das Ergebnis eines bestimmten Architekturstils.

Im Gegensatz dazu ist reaktive Programmierung ein Programmierparadigma, bei dem der Schwerpunkt auf der Entwicklung asynchroner und nicht blockierender Komponenten liegt . Der Kern der reaktiven Programmierung ist ein Datenstrom, den wir beobachten und darauf reagieren können, auch wenn wir Gegendruck ausüben. Dies führt zu einer nicht blockierenden Ausführung und damit zu einer besseren Skalierbarkeit mit weniger Ausführungsthreads.

Dies bedeutet nicht, dass sich reaktive Systeme und reaktive Programmierung gegenseitig ausschließen. In der Tat ist reaktive Programmierung ein wichtiger Schritt zur Realisierung eines reaktiven Systems, aber es ist nicht alles!

3.1. Reaktive Streams

Reactive Streams ist eine Community-Initiative, die bereits im Jahr 2013 gestartet wurde, um einen Standard für die asynchrone Stream-Verarbeitung mit nicht blockierendem Gegendruck bereitzustellen . Ziel war es, eine Reihe von Schnittstellen, Methoden und Protokollen zu definieren, mit denen die erforderlichen Operationen und Entitäten beschrieben werden können.

Seitdem sind mehrere Implementierungen in mehreren Programmiersprachen entstanden, die der Spezifikation für reaktive Streams entsprechen. Dazu gehören Akka Streams, Ratpack und Vert.x, um nur einige zu nennen.

3.2. Reaktive Bibliotheken für Java

Eines der ersten Ziele hinter den reaktiven Streams war es, schließlich als offizielle Java-Standardbibliothek aufgenommen zu werden. Infolgedessen entspricht die Spezifikation für reaktive Streams semantisch der in Java 9 eingeführten Java Flow-Bibliothek.

Abgesehen davon gibt es einige beliebte Möglichkeiten, um reaktive Programmierung in Java zu implementieren:

  • Reaktive Erweiterungen: Im Volksmund als ReactiveX bekannt, bieten sie eine API für die asynchrone Programmierung mit beobachtbaren Streams. Diese sind für mehrere Programmiersprachen und Plattformen verfügbar, einschließlich Java, wo es als RxJava bekannt ist
  • Projektreaktor: Dies ist eine weitere reaktive Bibliothek, die auf der Spezifikation der reaktiven Streams basiert und darauf abzielt, Nichtanwendungen auf der JVM zu erstellen. Es ist auch die Grundlage des reaktiven Stapels im Spring-Ökosystem

4. Eine einfache Anwendung

Für dieses Tutorial entwickeln wir eine einfache Anwendung, die auf der Microservices-Architektur mit einem minimalen Frontend basiert. Die Anwendungsarchitektur sollte genügend Elemente enthalten, um ein reaktives System zu erstellen.

Für unsere Anwendung verwenden wir eine durchgängige reaktive Programmierung und andere Muster und Werkzeuge, um die grundlegenden Eigenschaften eines reaktiven Systems zu erreichen.

4.1. Die Architektur

Zunächst definieren wir eine einfache Anwendungsarchitektur, die nicht unbedingt die Eigenschaften reaktiver Systeme aufweist . Von da an werden wir nacheinander die notwendigen Änderungen vornehmen, um diese Eigenschaften zu erreichen.

Beginnen wir also mit der Definition einer einfachen Architektur:

This is quite a simple architecture that has a bunch of microservices to facilitate a commerce use-case where we can place an order. It also has a frontend for user experience, and all communication happens as REST over HTTP. Moreover, every microservice manages their data in individual databases, a practice known as database-per-service.

We'll go ahead and create this simple application in the following sub-sections. This will be our base to understand the fallacies of this architecture and ways and means to adopt principles and practices so that we can transform this into a reactive system.

4.3. Inventory Microservice

Inventory microservice will be responsible for managing a list of products and their current stock. It will also allow altering the stock as orders are processed. We'll use Spring Boot with MongoDB to develop this service.

Let's begin by defining a controller to expose some endpoints:

@GetMapping public List getAllProducts() { return productService.getProducts(); } @PostMapping public Order processOrder(@RequestBody Order order) { return productService.handleOrder(order); } @DeleteMapping public Order revertOrder(@RequestBody Order order) { return productService.revertOrder(order); }

and a service to encapsulate our business logic:

@Transactional public Order handleOrder(Order order) { order.getLineItems() .forEach(l -> { Product> p = productRepository.findById(l.getProductId()) .orElseThrow(() -> new RuntimeException("Could not find the product: " + l.getProductId())); if (p.getStock() >= l.getQuantity()) { p.setStock(p.getStock() - l.getQuantity()); productRepository.save(p); } else { throw new RuntimeException("Product is out of stock: " + l.getProductId()); } }); return order.setOrderStatus(OrderStatus.SUCCESS); } @Transactional public Order revertOrder(Order order) { order.getLineItems() .forEach(l -> { Product p = productRepository.findById(l.getProductId()) .orElseThrow(() -> new RuntimeException("Could not find the product: " + l.getProductId())); p.setStock(p.getStock() + l.getQuantity()); productRepository.save(p); }); return order.setOrderStatus(OrderStatus.SUCCESS); }

Note that we're persisting the entities within a transaction, which ensures that no inconsistent state results in case of exceptions.

Apart from these, we'll also have to define the domain entities, the repository interface, and a bunch of configuration classes necessary for everything to work properly.

But since these are mostly boilerplate, we'll avoid going through them, and they can be referred to in the GitHub repository provided in the last section of this article.

4.4. Shipping Microservice

The shipping microservice will not be very different either. This will be responsible for checking if a shipment can be generated for the order and create one if possible.

As before we'll define a controller to expose our endpoints, in fact just a single endpoint:

@PostMapping public Order process(@RequestBody Order order) { return shippingService.handleOrder(order); }

and a service to encapsulate the business logic related to order shipment:

public Order handleOrder(Order order) { LocalDate shippingDate = null; if (LocalTime.now().isAfter(LocalTime.parse("10:00")) && LocalTime.now().isBefore(LocalTime.parse("18:00"))) { shippingDate = LocalDate.now().plusDays(1); } else { throw new RuntimeException("The current time is off the limits to place order."); } shipmentRepository.save(new Shipment() .setAddress(order.getShippingAddress()) .setShippingDate(shippingDate)); return order.setShippingDate(shippingDate) .setOrderStatus(OrderStatus.SUCCESS); }

Our simple shipping service is just checking the valid time window to place orders. We'll avoid discussing the rest of the boilerplate code as before.

4.5. Order Microservice

Finally, we'll define an order microservice which will be responsible for creating a new order apart from other things. Interestingly, it'll also play as an orchestrator service where it will communicate with the inventory service and the shipping service for the order.

Let's define our controller with the required endpoints:

@PostMapping public Order create(@RequestBody Order order) { Order processedOrder = orderService.createOrder(order); if (OrderStatus.FAILURE.equals(processedOrder.getOrderStatus())) { throw new RuntimeException("Order processing failed, please try again later."); } return processedOrder; } @GetMapping public List getAll() { return orderService.getOrders(); }

And, a service to encapsulate the business logic related to orders:

public Order createOrder(Order order) { boolean success = true; Order savedOrder = orderRepository.save(order); Order inventoryResponse = null; try { inventoryResponse = restTemplate.postForObject( inventoryServiceUrl, order, Order.class); } catch (Exception ex) { success = false; } Order shippingResponse = null; try { shippingResponse = restTemplate.postForObject( shippingServiceUrl, order, Order.class); } catch (Exception ex) { success = false; HttpEntity deleteRequest = new HttpEntity(order); ResponseEntity deleteResponse = restTemplate.exchange( inventoryServiceUrl, HttpMethod.DELETE, deleteRequest, Order.class); } if (success) { savedOrder.setOrderStatus(OrderStatus.SUCCESS); savedOrder.setShippingDate(shippingResponse.getShippingDate()); } else { savedOrder.setOrderStatus(OrderStatus.FAILURE); } return orderRepository.save(savedOrder); } public List getOrders() { return orderRepository.findAll(); }

The handling of orders where we're orchestrating calls to inventory and shipping services is far from ideal. Distributed transactions with multiple microservices is a complex topic in itself and beyond the scope of this tutorial.

However, we'll see later in this tutorial how a reactive system can avoid the need for distributed transactions to a certain extent.

As before, we'll not go through the rest of the boilerplate code. However, this can be referenced in the GitHub repo.

4.6. Front-end

Let's also add a user interface to make the discussion complete. The user interface will be based on Angular and will be a simple single-page application.

We'll need to create a simple component in Angular to handle create and fetch orders. Of specific importance is the part where we call our API to create the order:

createOrder() { let headers = new HttpHeaders({'Content-Type': 'application/json'}); let options = {headers: headers} this.http.post('//localhost:8080/api/orders', this.form.value, options) .subscribe( (response) => { this.response = response }, (error) => { this.error = error } ) }

The above code snippet expects order data to be captured in a form and available within the scope of the component. Angular offers fantastic support for creating simple to complex forms using reactive and template-driven forms.

Also important is the part where we get previously created orders:

getOrders() { this.previousOrders = this.http.get(''//localhost:8080/api/orders'') }

Please note that the Angular HTTP module is asynchronous in nature and hence returns RxJS Observables. We can handle the response in our view by passing them through an async pipe:

Your orders placed so far:

  • Order ID: {{ order.id }}, Order Status: {{order.orderStatus}}, Order Message: {{order.responseMessage}}

Of course, Angular will require templates, styles, and configurations to work, but these can be referred to in the GitHub repository. Please note that we have bundled everything in a single component here, which is ideally not something we should do.

But, for this tutorial, those concerns are not in scope.

4.7. Deploying the Application

Now that we've created all individual parts of the application, how should we go about deploying them? Well, we can always do this manually. But we should be careful that it can soon become tedious.

For this tutorial, we'll use Docker Compose to build and deploy our application on a Docker Machine. This will require us to add a standard Dockerfile in each service and create a Docker Compose file for the entire application.

Let's see how this docker-compose.yml file looks:

version: '3' services: frontend: build: ./frontend ports: - "80:80" order-service: build: ./order-service ports: - "8080:8080" inventory-service: build: ./inventory-service ports: - "8081:8081" shipping-service: build: ./shipping-service ports: - "8082:8082"

This is a fairly standard definition of services in Docker Compose and does not require any special attention.

4.8. Problems With This Architecture

Now that we have a simple application in place with multiple services interacting with each other, we can discuss the problems in this architecture. There are what we'll try to address in the following sections and eventually get to the state where we would have transformed our application into a reactive system!

While this application is far from a production-grade software and there are several issues, we'll focus on the issues that pertain to the motivations for reactive systems:

  • Failure in either inventory service or shipping service can have a cascading effect
  • The calls to external systems and database are all blocking in nature
  • The deployment cannot handle failures and fluctuating loads automatically

5. Reactive Programming

Blocking calls in any program often result in critical resources just waiting for things to happen. These include database calls, calls to web services, and file system calls. If we can free up threads of execution from this waiting and provide a mechanism to circle back once results are available, it will yield much better resource utilization.

This is what adopting the reactive programming paradigm does for us. While it's possible to switch over to a reactive library for many of these calls, it may not be possible for everything. For us, fortunately, Spring makes it much easier to use reactive programming with MongoDB and REST APIs:

Spring Data Mongo has support for reactive access through the MongoDB Reactive Streams Java Driver. It provides ReactiveMongoTemplate and ReactiveMongoRepository, both of which have extensive mapping functionality.

Spring WebFlux provides the reactive-stack web framework for Spring, enabling non-blocking code and Reactive Streams backpressure. It leverages the Reactor as its reactive library. Further, it provides WebClient for performing HTTP requests with Reactive Streams backpressure. It uses Reactor Netty as the HTTP client library.

5.1. Inventory Service

We'll begin by changing our endpoints to emit reactive publishers:

@GetMapping public Flux getAllProducts() { return productService.getProducts(); }
@PostMapping public Mono processOrder(@RequestBody Order order) { return productService.handleOrder(order); } @DeleteMapping public Mono revertOrder(@RequestBody Order order) { return productService.revertOrder(order); }

Obviously, we'll have to make necessary changes to the service as well:

@Transactional public Mono handleOrder(Order order) { return Flux.fromIterable(order.getLineItems()) .flatMap(l -> productRepository.findById(l.getProductId())) .flatMap(p -> { int q = order.getLineItems().stream() .filter(l -> l.getProductId().equals(p.getId())) .findAny().get() .getQuantity(); if (p.getStock() >= q) { p.setStock(p.getStock() - q); return productRepository.save(p); } else { return Mono.error(new RuntimeException("Product is out of stock: " + p.getId())); } }) .then(Mono.just(order.setOrderStatus("SUCCESS"))); } @Transactional public Mono revertOrder(Order order) { return Flux.fromIterable(order.getLineItems()) .flatMap(l -> productRepository.findById(l.getProductId())) .flatMap(p -> { int q = order.getLineItems().stream() .filter(l -> l.getProductId().equals(p.getId())) .findAny().get() .getQuantity(); p.setStock(p.getStock() + q); return productRepository.save(p); }) .then(Mono.just(order.setOrderStatus("SUCCESS"))); }

5.2. Shipping Service

Similarly, we'll change the endpoint of our shipping service:

@PostMapping public Mono process(@RequestBody Order order) { return shippingService.handleOrder(order); }

And, corresponding changes in the service to leverage reactive programming:

public Mono handleOrder(Order order) { return Mono.just(order) .flatMap(o -> { LocalDate shippingDate = null; if (LocalTime.now().isAfter(LocalTime.parse("10:00")) && LocalTime.now().isBefore(LocalTime.parse("18:00"))) { shippingDate = LocalDate.now().plusDays(1); } else { return Mono.error(new RuntimeException("The current time is off the limits to place order.")); } return shipmentRepository.save(new Shipment() .setAddress(order.getShippingAddress()) .setShippingDate(shippingDate)); }) .map(s -> order.setShippingDate(s.getShippingDate()) .setOrderStatus(OrderStatus.SUCCESS)); }

5.3. Order Service

We'll have to make similar changes in the endpoints of the order service:

@PostMapping public Mono create(@RequestBody Order order) { return orderService.createOrder(order) .flatMap(o -> { if (OrderStatus.FAILURE.equals(o.getOrderStatus())) { return Mono.error(new RuntimeException("Order processing failed, please try again later. " + o.getResponseMessage())); } else { return Mono.just(o); } }); } @GetMapping public Flux getAll() { return orderService.getOrders(); }

The changes to service will be more involved as we'll have to make use of Spring WebClient to invoke the inventory and shipping reactive endpoints:

public Mono createOrder(Order order) { return Mono.just(order) .flatMap(orderRepository::save) .flatMap(o -> { return webClient.method(HttpMethod.POST) .uri(inventoryServiceUrl) .body(BodyInserters.fromValue(o)) .exchange(); }) .onErrorResume(err -> { return Mono.just(order.setOrderStatus(OrderStatus.FAILURE) .setResponseMessage(err.getMessage())); }) .flatMap(o -> { if (!OrderStatus.FAILURE.equals(o.getOrderStatus())) { return webClient.method(HttpMethod.POST) .uri(shippingServiceUrl) .body(BodyInserters.fromValue(o)) .exchange(); } else { return Mono.just(o); } }) .onErrorResume(err -> { return webClient.method(HttpMethod.POST) .uri(inventoryServiceUrl) .body(BodyInserters.fromValue(order)) .retrieve() .bodyToMono(Order.class) .map(o -> o.setOrderStatus(OrderStatus.FAILURE) .setResponseMessage(err.getMessage())); }) .map(o -> { if (!OrderStatus.FAILURE.equals(o.getOrderStatus())) { return order.setShippingDate(o.getShippingDate()) .setOrderStatus(OrderStatus.SUCCESS); } else { return order.setOrderStatus(OrderStatus.FAILURE) .setResponseMessage(o.getResponseMessage()); } }) .flatMap(orderRepository::save); } public Flux getOrders() { return orderRepository.findAll(); }

This kind of orchestration with reactive APIs is no easy exercise and often error-prone as well as hard to debug. We'll see how this can be simplified in the next section.

5.4. Front-end

Now, that our APIs are capable of streaming events as they occur, it's quite natural that we should be able to leverage that in our front-end as well. Fortunately, Angular supports EventSource, the interface for Server-Sent Events.

Let's see how can we pull and process all our previous orders as a stream of events:

getOrderStream() { return Observable.create((observer) => { let eventSource = new EventSource('//localhost:8080/api/orders') eventSource.onmessage = (event) => { let json = JSON.parse(event.data) this.orders.push(json) this._zone.run(() => { observer.next(this.orders) }) } eventSource.onerror = (error) => { if(eventSource.readyState === 0) { eventSource.close() this._zone.run(() => { observer.complete() }) } else { this._zone.run(() => { observer.error('EventSource error: ' + error) }) } } }) }

6. Message-Driven Architecture

The first problem we're going to address is related to service-to-service communication. Right now, these communications are synchronous, which presents several problems. These include cascading failures, complex orchestration, and distributed transactions to name a few.

An obvious way to solve this problem is to make these communications asynchronous. A message broker for facilitating all service-to-service communication can do the trick for us. We'll use Kafka as our message broker and Spring for Kafka to produce and consume messages:

We'll use a single topic to produce and consume order messages with different order statuses for services to react.

Let's see how each service needs to change.

6.1. Inventory Service

Let's begin by defining the message producer for our inventory service:

@Autowired private KafkaTemplate kafkaTemplate; public void sendMessage(Order order) { this.kafkaTemplate.send("orders", order); }

Next, we'll have to define a message consumer for inventory service to react to different messages on the topic:

@KafkaListener(topics = "orders", groupId = "inventory") public void consume(Order order) throws IOException { if (OrderStatus.RESERVE_INVENTORY.equals(order.getOrderStatus())) { productService.handleOrder(order) .doOnSuccess(o -> { orderProducer.sendMessage(order.setOrderStatus(OrderStatus.INVENTORY_SUCCESS)); }) .doOnError(e -> { orderProducer.sendMessage(order.setOrderStatus(OrderStatus.INVENTORY_FAILURE) .setResponseMessage(e.getMessage())); }).subscribe(); } else if (OrderStatus.REVERT_INVENTORY.equals(order.getOrderStatus())) { productService.revertOrder(order) .doOnSuccess(o -> { orderProducer.sendMessage(order.setOrderStatus(OrderStatus.INVENTORY_REVERT_SUCCESS)); }) .doOnError(e -> { orderProducer.sendMessage(order.setOrderStatus(OrderStatus.INVENTORY_REVERT_FAILURE) .setResponseMessage(e.getMessage())); }).subscribe(); } }

This also means that we can safely drop some of the redundant endpoints from our controller now. These changes are sufficient to achieve asynchronous communication in our application.

6.2. Shipping Service

The changes in shipping service are relatively similar to what we did earlier with the inventory service. The message producer is the same, and the message consumer is specific to shipping logic:

@KafkaListener(topics = "orders", groupId = "shipping") public void consume(Order order) throws IOException { if (OrderStatus.PREPARE_SHIPPING.equals(order.getOrderStatus())) { shippingService.handleOrder(order) .doOnSuccess(o -> { orderProducer.sendMessage(order.setOrderStatus(OrderStatus.SHIPPING_SUCCESS) .setShippingDate(o.getShippingDate())); }) .doOnError(e -> { orderProducer.sendMessage(order.setOrderStatus(OrderStatus.SHIPPING_FAILURE) .setResponseMessage(e.getMessage())); }).subscribe(); } }

We can safely drop all the endpoints in our controller now as we no longer need them.

6.3. Order Service

The changes in order service will be a little more involved as this is where we were doing all the orchestration earlier.

Nevertheless, the message producer remains unchanged, and message consumer takes on order service-specific logic:

@KafkaListener(topics = "orders", groupId = "orders") public void consume(Order order) throws IOException { if (OrderStatus.INITIATION_SUCCESS.equals(order.getOrderStatus())) { orderRepository.findById(order.getId()) .map(o -> { orderProducer.sendMessage(o.setOrderStatus(OrderStatus.RESERVE_INVENTORY)); return o.setOrderStatus(order.getOrderStatus()) .setResponseMessage(order.getResponseMessage()); }) .flatMap(orderRepository::save) .subscribe(); } else if ("INVENTORY-SUCCESS".equals(order.getOrderStatus())) { orderRepository.findById(order.getId()) .map(o -> { orderProducer.sendMessage(o.setOrderStatus(OrderStatus.PREPARE_SHIPPING)); return o.setOrderStatus(order.getOrderStatus()) .setResponseMessage(order.getResponseMessage()); }) .flatMap(orderRepository::save) .subscribe(); } else if ("SHIPPING-FAILURE".equals(order.getOrderStatus())) { orderRepository.findById(order.getId()) .map(o -> { orderProducer.sendMessage(o.setOrderStatus(OrderStatus.REVERT_INVENTORY)); return o.setOrderStatus(order.getOrderStatus()) .setResponseMessage(order.getResponseMessage()); }) .flatMap(orderRepository::save) .subscribe(); } else { orderRepository.findById(order.getId()) .map(o -> { return o.setOrderStatus(order.getOrderStatus()) .setResponseMessage(order.getResponseMessage()); }) .flatMap(orderRepository::save) .subscribe(); } }

The consumer here is merely reacting to order messages with different order statuses. This is what gives us the choreography between different services.

Lastly, our order service will also have to change to support this choreography:

public Mono createOrder(Order order) { return Mono.just(order) .flatMap(orderRepository::save) .map(o -> { orderProducer.sendMessage(o.setOrderStatus(OrderStatus.INITIATION_SUCCESS)); return o; }) .onErrorResume(err -> { return Mono.just(order.setOrderStatus(OrderStatus.FAILURE) .setResponseMessage(err.getMessage())); }) .flatMap(orderRepository::save); }

Note that this is far simpler than the service we had to write with reactive endpoints in the last section. Asynchronous choreography often results in far simpler code, although it does come at the cost of eventual consistency and complex debugging and monitoring. As we may guess, our front-end will no longer get the final status of the order immediately.

7. Container Orchestration Service

The last piece of the puzzle that we want to solve is related to deployment.

What we want in the application is ample redundancy and a tendency to scale up or down depending upon the need automatically.

We've already achieved containerization of services through Docker and are managing dependencies between them through Docker Compose. While these are fantastic tools in their own right, they do not help us to achieve what we want.

Hence, we need a container orchestration service that can take care of redundancy and scalability in our application. While there are several options, one of the popular ones includes Kubernetes. Kubernetes provides us with a cloud vendor-agnostic way to achieve highly scalable deployments of containerized workloads.

Kubernetes wraps containers like Docker into Pods, which are the smallest unit of deployment. Further, we can use Deployment to describe the desired state declaratively.

Deployment creates ReplicaSets, which internally is responsible for bringing up the pods. We can describe a minimum number of identical pods that should be running at any point in time. This provides redundancy and hence high availability.

Let's see how can we define a Kubernetes deployment for our applications:

apiVersion: apps/v1 kind: Deployment metadata: name: inventory-deployment spec: replicas: 3 selector: matchLabels: name: inventory-deployment template: metadata: labels: name: inventory-deployment spec: containers: - name: inventory image: inventory-service-async:latest ports: - containerPort: 8081 --- apiVersion: apps/v1 kind: Deployment metadata: name: shipping-deployment spec: replicas: 3 selector: matchLabels: name: shipping-deployment template: metadata: labels: name: shipping-deployment spec: containers: - name: shipping image: shipping-service-async:latest ports: - containerPort: 8082 --- apiVersion: apps/v1 kind: Deployment metadata: name: order-deployment spec: replicas: 3 selector: matchLabels: name: order-deployment template: metadata: labels: name: order-deployment spec: containers: - name: order image: order-service-async:latest ports: - containerPort: 8080

Here we're declaring our deployment to maintain three identical replicas of pods at any time. While this is a good way to add redundancy, it may not be sufficient for varying loads. Kubernetes provides another resource known as the Horizontal Pod Autoscaler which can scale the number of pods in a deployment based on observed metrics like CPU utilization.

Please note that we have just covered the scalability aspects of the application hosted on a Kubernetes cluster. This does not necessarily imply that the underlying cluster itself is scalable. Creating a high availability Kubernetes cluster is a non-trivial task and beyond the scope of this tutorial.

8. Resulting Reactive System

Now that we've made several improvements in our architecture, it's perhaps time to evaluate this against the definition of a Reactive System. We'll keep the evaluation against the four characteristics of a Reactive Systems we discussed earlier in the tutorial:

  • Responsive: The adoption of the reactive programming paradigm should help us achieve end-to-end non-blocking and hence a responsive application
  • Resilient: Kubernetes deployment with ReplicaSet of the desired number of pods should provide resilience against random failures
  • Elastic: Kubernetes cluster and resources should provide us the necessary support to be elastic in the face of unpredictable loads
  • Message-Driven: Having all service-to-service communication handled asynchronously through a Kafka broker should help us here

While this looks quite promising, it's far from over. To be honest, the quest for a truly reactive system should be a continuous exercise of improvements. We can never preempt all that can fail in a highly complex infrastructure, where our application is just a small part.

A reactive system thus will demand reliability from every part that makes the whole. Right from the physical network to infrastructure services like DNS, they all should fall in line to help us achieve the end goal.

Often, it may not be possible for us to manage and provide the necessary guarantees for all these parts. And this is where a managed cloud infrastructure helps alleviate our pain. We can choose from a host of services like IaaS (Infeastrure-as-a-Service), BaaS (Backend-as-a-Service), and PaaS (Platform-as-a-Service) to delegate the responsibilities to external parties. This leaves us with the responsibility of our application as far as possible.

9. Conclusion

In this tutorial, we went through the basics of reactive systems and how does it compare with reactive programming. We created a simple application with multiple microservices and highlighted the problems we intend to solve with a reactive system.

Darüber hinaus haben wir reaktive Programmierung, nachrichtenbasierte Architektur und Container-Orchestrierungsdienst in die Architektur eingeführt, um ein reaktives System zu realisieren.

Zuletzt haben wir die resultierende Architektur besprochen und wie sie eine Reise in Richtung des reaktiven Systems bleibt! Dieses Tutorial führt uns nicht in alle Tools, Frameworks oder Muster ein, die uns beim Erstellen eines reaktiven Systems helfen können, sondern führt uns in die Reise ein.

Wie üblich finden Sie den Quellcode für diesen Artikel auf GitHub.