Serverless and Data Streaming

Introduction
Claim Service
Claim Processor
Stream Processor
ClaimsCount Dashboard

Introduction

There are many benefits to following clean and layered approaches to application architecture such as hexagonal arcthitecture. When application core is decoupled from transport details,the core functionality can be exposed in a manner to suit any business need, for e.g. as a REST endpoint or via asynchronous event listeners. It also means codebase that is testable by design.

In the Java world, SpringFramework makes this very easy through Spring Cloud Function project. The idea is our application can be thought of as functions that take an input and produce output. The framework then provides support to deploy these functions as a REST service or via event listeners in a runtime environment of choice - Serverless, Kubernetes or stand alone event listeners.

The sample application illustrates how Spring Cloud Function can be used to write business logic as java functions and deploy anywhere for e.g., to a Serverless platform or as a REST endpoints or asynchronous Kafka event listeners. It also shows how the solution can easility be extended to implement an event streaming architecture, finishing off with an http end point serving Server Sent Events for real time notifications to a Web application.

The sample application implements a very simple Claim Processing solution where Customers can submit claim requests that will then be processed according to the business rules resulting in customer getting paid or the request being rejected.

The application consists of three main components.

ClaimsService
ClaimsProcessor
StreamProcessor

Overall architecture diagram:

ClaimRequest submitted to ClaimService -> produces ClaimCreatedEvent -> ClaimProcessor picks the ClaimCreatedEvent to go through fraud check process to produce PaymentEvent which is processed by payment processor for the final settlement.

Claim Service

The application exposes a http endpoint for users to submit their claim request. The service is implemented using Spring Cloud Function following hexagonal architecture principles.

At the core, the service is a function that accepts ClaimRequest to produce a ClaimResponse output. For simplicity, ClaimRequest has following attributes:

{
   firstName: string,
   lastName: string,
   email: string,
   claimAmount: number,
   claimType: string
}

and ClaimResponse consists of

{
   status: {
      code: string,
      messate: string
   },
   correlationId: string
}

Service publishes event on a Kafka topic upon receipt of customer claim request.

Running Locally for Development

Run kafka locally
- using Docker Compose cd docker && docker-compose up (ensure that docker daemon is running) It starts kafka with image bitnami/kafka:latest which is kafka without zookeeper.

The application is a spring cloud function and can be started locally using the following command: ./gradlew bootRun To test, issuing following POST request,

POST http://localhost:8080/submitClaimRequest
Content-Type: application/json

{
  "firstName": "Joe",
  "lastName": "Blogs",
  "amount": 20,
  "claimType": "CLAIM_A",
  "email": "test@user.com"
}

will produce following response:

{
  "status" : {
    "code": "UNDER_REVIEW",
    "message": "Application received"
  },
  "correlationId": "7d979021-5cc8-478d-8990-9d3d7c23b20a"
}

Running on local K8s Cluster

Pre-requisites

Local Kubernetes

k3s is lightweight distribution of Kubernetes perfect for resource constrained environments and a good choice for local development environment. And k3d can be used to run k3s locally on Docker. Also ensure kubectl is installed to connect to the cluster. Create a local cluster with name dev k3d cluster create dev kubectl get nodes to list the nodes in the cluster.

Kafka

We can start with hosted kafka solution, or deploy our own Kafka broker in the cluster. For a hosted solution https://www.confluent.io/get-started/ is one option to get up and running quickly.

Deployment

build image using buildpack ./gradlew bootBuildImage, which will build a local docker image claimservice:0.0.1-SNAPSHOT. [!IMPORTANT] Ensure that the thin jar plugin related config is commented out from build.gradle.kts to ensure that the image is built using self-contained fat jar
import the image into the cluster so that it is accessible to the deployments k3d image import claimservice:0.0.1-SNAPSHOT -c {cluster-name}
cd k8s
kubectl create -f secrets.yaml (for kafka credentials to connect to Confluent hosted broker)
kubectl create -f deployment.yaml
kubectl create -f service-node-port.yaml (to create a node port service for local testing)

The service can now be accessed at POST http://localhost:30000/submitClaimRequest

POST http://localhost:30000/submitClaimRequest
Content-Type: application/json
{
  "firstName": "Joe",
  "lastName": "Blogs",
  "amount": 20,
  "claimType": "",
  "email": "test@user.com"
}

Running on AWS Lambda

SpringCloudFunction provides adapters for various serverless platforms. For AWS Lambda, we need the spring cloud function adapter for aws and the thin jar plugin to produce the shaded jar for deployment.

implementation("org.springframework.cloud:spring-cloud-function-adapter-aws")
implementation("org.springframework.boot.experimental:spring-boot-thin-gradle-plugin:1.0.29.RELEASE")

and the following declarations in the build.gradle.kts

tasks.assemble{
	dependsOn("shadowJar")
}

tasks.withType<ShadowJar> {
    archiveClassifier.set("aws")
	dependencies {
		exclude("org.springframework.cloud:spring-cloud-function-web")
	}
	// Required for Spring
	mergeServiceFiles()
	append("META-INF/spring.handlers")
	append("META-INF/spring.schemas")
	append("META-INF/spring.tooling")
	transform(PropertiesFileTransformer::class.java) {
		paths.add("META-INF/spring.factories")
		mergeStrategy = "append"
	}
}

tasks.withType<Jar> {
	manifest {
		attributes["Start-Class"] = "com.arun.claimservice.ClaimserviceApplication"
	}
}

We also need a request handler to process incoming requests and Spring provides a generic request handler for this: org.springframework.cloud.function.adapter.aws.FunctionInvoker::handleRequest

Following environment variables need to be defined for this function example - spring_profiles_active, KAFKA_USR, KAFKA_PASSWORD

Running on TAS

Running on KNative

Running on OpenShift

Running on AWS EKS

Prerequisites

Create an AWS EKS cluster by following instructions ...

Claims Processor

Claims Processor consists of functions that pick up the submitted claims for fraud check and then finalise the payment. Following functions carry out fraud checks and processing of the payments. These functions are invoked by kafka message listeners defined in KafkaConfig.

fun performFraudCheck(fraudCheckService: FraudCheckService): (ClaimCreatedEvent) -> FraudCheckResult {
    return {
       fraudCheckService.performFraudCheck(it)
    }
}

fun performPayment(paymentService: PaymentService): (FraudCheckResult) -> PaymentResult {
    return {
        paymentService.performPayment(it)
    }
}

Stream Processor

In the first part, implemented a stream that listens to the CLAIMS topic for new claim requests, aggregates them by claim type and publishes claim count by type on a new topic. Secondly implemented a REST endpoint via spring cloud function to serve events as they are published on this new topic using reactor Sink.

Implement Stream Topology

Stream topology is implemented in

    @Bean
    fun kStream(builder: StreamsBuilder) {
        claimsStream(builder)
    }

The stream topology above calculates number of claims by type in every five-minute window.

REST API to serve from Stream

We can expose a REST api to serve the claim count as processed by the stream. For this we need to first define a sink for the topic event-count where the claim count events are being published.

SINK set up

First define a sink

    @Bean
    fun claimCountSink(): Sinks.Many<ClaimCount> {
        val replaySink = Sinks.many().multicast().onBackpressureBuffer<ClaimCount>()
        return replaySink
    }

And then use this sink for the topic event-topic where claim count events are published by the stream.

@KafkaListener(
    topics = ["event-count"], groupId = "CONSUMER_ONE",
    containerFactory = "claimCountKafkaListenerContainerFactory"
)
class ClaimCountListener(private val sink: Sinks.Many<ClaimCount>) {
    private val logger = LoggerFactory.getLogger(ClaimCountListener::class.java)

    @KafkaHandler
    fun processMessage(claimCount: ClaimCount) {
        logger.info("Received claim count $claimCount")
        sink.emitNext(claimCount, { s,e -> true})
    }
}

Spring Cloud Function set up

Following function exposes retrieveClaimCount endpoint which will publish claim count events as server sent events.

    @Bean
    fun retrieveClaimCount(claimCountSinks: Sinks.Many<ClaimCount>): () -> Flux<ClaimCount> {
        return retrieveClaimCountFunction(claimCountSinks)
    }

Claim Count Dashboard (UI)

The dashboard uses EventSource to streaming event counts from the API to the browser which is then displayed in a line chart.

import {useEffect} from "react";

const eventSource = new EventSource(`http://localhost:8081/retrieveClaimCount`);
...
eventSource.onmessage = m => {
...
    // display events in the chart
}

The react app claims-ui can be started by running npm run dev and navigate to http://localhost:3000/claimCount

Scheduler to generate test events

The function taking KStream as input is not getting invoked, but the others with bytearray inputs are invoked and can also be chained together to feed output from one as input to the other.

achalise/cloudnative-serverless-streaming