Spring Cloud Function 2018-06-04 Mark Fisher, Dave Syer Introduction Spring Cloud Function is a project with the following high-level goals: Promote the implementation of business logic via functions. Decouple the development lifecycle of business logic from any specific runtime target so that the same code can run as a web endpoint, a stream processor, or a task. Support a uniform programming model across serverless providers, as well as the ability to run standalone (locally or in a PaaS). Enable Spring Boot features (auto-configuration, dependency injection, metrics) on serverless providers. It abstracts away all of the transport details and infrastructure, allowing the developer to keep all the familiar tools and processes, and focus firmly on business logic. Here’s a complete, executable, testable Spring Boot application (implementing a simple string manipulation): @SpringBootApplication public class Application { @Bean public Function<Flux<String>, Flux<String>> uppercase() { return flux -> flux.map(value -> value.toUpperCase()); } public static void main(String[] args) { SpringApplication.run(Application.class, args); } } It’s just a Spring Boot application, so it can be built, run and tested, locally and in a CI build, the same way as any other Spring Boot application. The Function is from java.util and Flux is a Reactive Streams Publisher from Project Reactor. The function can be accessed over HTTP or messaging. Spring Cloud Function has 4 main features: Wrappers for @Beans of type Function, Consumer and Supplier, exposing them to the outside world as either HTTP endpoints and/or message stream listeners/publishers with RabbitMQ, Kafka etc. Compiling strings which are Java function bodies into bytecode, and then turning them into @Beans that can be wrapped as above. Deploying a JAR file containing such an application context with an isolated classloader, so that you can pack them together in a single JVM. Adapters for AWS Lambda, Azure, Apache OpenWhisk and possibly other "serverless" service providers. Spring Cloud is released under the non-restrictive Apache 2.0 license. If you would like to contribute to this section of the documentation or if you find an error, please find the source code and issue trackers in the project at github. Getting Started Build from the command line (and "install" the samples): $ ./mvnw clean install (If you like to YOLO add -DskipTests.) Run one of the samples, e.g. $ java -jar spring-cloud-function-samples/function-sample/target/*.jar This runs the app and exposes its functions over HTTP, so you can convert a string to uppercase, like this: $ curl -H "Content-Type: text/plain" localhost:8080/uppercase -d Hello HELLO You can convert multiple strings (a Flux<String>) by separating them with new lines $ curl -H "Content-Type: text/plain" localhost:8080/uppercase -d 'Hello > World' HELLOWORLD (You can use QJ in a terminal to insert a new line in a literal string like that.) Building and Running a Function The sample @SpringBootApplication above has a function that can be decorated at runtime by Spring Cloud Function to be an HTTP endpoint, or a Stream processor, for instance with RabbitMQ, Apache Kafka or JMS. The @Beans can be Function, Consumer or Supplier (all from java.util), and their parametric types can be String or POJO. A Function is exposed as a Spring Cloud Stream Processor if spring-cloud-function-stream is on the classpath. A Consumer is also exposed as a Stream Sink and a Supplier translates to a Stream Source. HTTP endpoints are exposed if the Stream binder is spring-cloud-stream-binder-servlet. Functions can be of Flux<String> or Flux<Pojo> and Spring Cloud Function takes care of converting the data to and from the desired types, as long as it comes in as plain text or (in the case of the POJO) JSON. TBD: support for Flux<Message<Pojo>> and maybe plain Pojo types (Fluxes implied and implemented by the framework). Functions can be grouped together in a single application, or deployed one-per-jar. It’s up to the developer to choose. An app with multiple functions can be deployed multiple times in different "personalities", exposing different functions over different physical transports. Dynamic Compilation There is a sample app that uses the function compiler to create a function from a configuration property. The vanilla "function-sample" also has that feature. And there are some examples that you can run to see the compilation happening at run time. To run these examples, change into the scripts directory: cd scripts Also, start a RabbitMQ server locally (e.g. execute rabbitmq-server).
Start the Function Registry Service: ./function-registry.sh
Register a Function: ./registerFunction.sh -n uppercase -f "f->f.map(s->s.toString().toUpperCase())"
Run a REST Microservice using that Function: ./web.sh -f uppercase -p 9000 curl -H "Content-Type: text/plain" -H "Accept: text/plain" localhost:9000/uppercase -d foo
Register a Supplier: ./registerSupplier.sh -n words -f "()->Flux.just(\"foo\",\"bar\")"
Run a REST Microservice using that Supplier: ./web.sh -s words -p 9001 curl -H "Accept: application/json" localhost:9001/words
Register a Consumer: ./registerConsumer.sh -n print -t String -f "System.out::println"
Run a REST Microservice using that Consumer: ./web.sh -c print -p 9002 curl -X POST -H "Content-Type: text/plain" -d foo localhost:9002/print
Run Stream Processing Microservices: First register a streaming words supplier: ./registerSupplier.sh -n wordstream -f "()->Flux.interval(Duration.ofMillis(1000)).map(i->\"message-\"+i)" Then start the source (supplier), processor (function), and sink (consumer) apps (in reverse order): ./stream.sh -p 9103 -i uppercaseWords -c print ./stream.sh -p 9102 -i words -f uppercase -o uppercaseWords ./stream.sh -p 9101 -s wordstream -o words The output will appear in the console of the sink app (one message per second, converted to uppercase): MESSAGE-0 MESSAGE-1 MESSAGE-2 MESSAGE-3 MESSAGE-4 MESSAGE-5 MESSAGE-6 MESSAGE-7 MESSAGE-8 MESSAGE-9 ...
Function Catalog and Flexible Function Signatures One of the main features of Spring Cloud Function is to adapt and support a range of type signatures for user-defined functions. So users can supply a bean of type Function<String,String>, for instance, and the FunctionCatalog will wrap it into a Function<Flux<String>,Flux<String>>. Users don’t normally have to care about the FunctionCatalog at all, but it is useful to know what kind of functions are supported in user code. Generally speaking users can expect that if they write a function for a plain old Java type (or primitive wrapper), then the function catalog will wrap it to a Flux of the same type. If the user writes a function using Message (from spring-messaging) it will receive and transmit headers from any adapter that supports key-value metadata (e.g. HTTP headers). Here are the details. User Function Catalog Registration Function<S,T> Function<Flux<S>, Flux<T>> Function<Message<S>,Message<T>> Function<Flux<Message<S>>, Flux<Message<T>>> Function<Flux<S>, Flux<T>> Function<Flux<S>, Flux<T>> (pass through) Supplier<T> Supplier<Flux<T>> Supplier<Flux<T>> Supplier<Flux<T>> Consumer<T> Function<Flux<T>, Mono<Void>> Consumer<Message<T>> Function<Flux<Message<T>>, Mono<Void>> Consumer<Flux<T>> Consumer<Flux<T>> Consumer is a little bit special because it has a void return type, which implies blocking, at least potentially. Most likely you will not need to write Consumer<Flux<?>>, but if you do need to do that, remember to subscribe to the input flux. If you declare a Consumer of a non publisher type (which is normal), it will be converted to a function that returns a publisher, so that it can be subscribed to in a controlled way. A function catalog can contain a Supplier and a Function (or Consumer) with the same name (like a GET and a POST to the same resource). It can even contain a Consumer<Flux<>> with the same name as a Function, but it cannot contain a Consumer<T> and a Function<T,S> with the same name when T is not a Publisher because the consumer would be converted to a Function and only one of them can be registered. Standalone Web Applications The spring-cloud-function-web module has autoconfiguration that activates when it is included in a Spring Boot web application (with MVC support). There is also a spring-cloud-starter-function-web to collect all the optional dependnecies in case you just want a simple getting started experience. With the web configurations activated your app will have an MVC endpoint (on "/" by default, but configurable with spring.cloud.function.web.path) that can be used to access the functions in the application context. The supported content types are plain text and JSON. Method Path Request Response Status GET /{supplier} - Items from the named supplier 200 OK POST /{consumer} JSON object or text Mirrors input and pushes request body into consumer 202 Accepted POST /{consumer} JSON array or text with new lines Mirrors input and pushes body into consumer one by one 202 Accepted POST /{function} JSON object or text The result of applying the named function 200 OK POST /{function} JSON array or text with new lines The result of applying the named function 200 OK GET /{function}/{item} - Convert the item into an object and return the result of applying the function 200 OK As the table above shows the behaviour of the endpoint depends on the method and also the type of incoming request data. When the incoming data is single valued, and the target function is declared as obviously single valued (i.e. not returning a collection or Flux), then the response will also contain a single value. For multi-valued responses the client can ask for a server-sent event stream by sending `Accept: text/event-stream". If there is only one function (consumer etc.) then the name in the path is optional. Composite functions can be addressed using pipes or commas to separate function names (pipes are legal in URL paths, but a bit awkward to type on the command line). Functions and consumers that are declared with input and output in Message<?> will see the request headers on the input messages, and the output message headers will be converted to HTTP headers. When POSTing text the response format might be different with Spring Boot 2.0 and older versions, depending on the content negotiation (provide content type and accpt headers for the best results). Standalone Streaming Applications To send or receive messages from a broker (such as RabbitMQ or Kafka) you can use the spring-cloud-function-stream adapter. Add the adapter to your classpath along with the appropriate binder from Spring Cloud Stream. The adapter will bind to the message broker as a Processor (input and output streams) unless the user explicitly disables one or the other using spring.cloud.function.stream.{source,sink}.enabled=false. An incoming message is routed to a function (or consumer). If there is only one, then the choice is obvious. If there are multiple functions that can accept an incoming message, the message is inspected to see if there is a stream_routekey header containing the name of a function. Routing headers or function names can be composed using a comma- or pipe-separated name. The header is also added to outgoing messages from a supplier. Messages with no route key can be routed exclusively to a function or consumer by specifying spring.cloud.function.stream.{processor,sink}.name. If a single function cannot be identified to process an incoming message there will be an error, unless you set spring.cloud.function.stream.shared=true, in which case such messages will be sent to all compatible functions. A single supplier can be chosen for output messages from a supplier (if more than one is available) using the spring.cloud.function.stream.source.name. some binders will fail on startup if the message broker is not available and the function catalog contains suppliers that immediately produce messages when accessed. You can switch off the automatic publishing from suppliers on startup using the spring.cloud.function.strean.supplier.enabled=false flag. Deploying a Packaged Function Spring Cloud Function provides a "deployer" library that allows you to launch a jar file (or exploded archive, or set of jar files) with an isolated class loader and expose the functions defined in it. This is quite a powerful tool that would allow you to, for instance, adapt a function to a range of different input-output adapters without changing the target jar file. Serverless platforms often have this kind of feature built in, so you could see it as a building block for a function invoker in such a platform (indeed the Riff Java function invoker uses this library). The standard entry point of the API is the Spring configuration annotation @EnableFunctionDeployer. If that is used in a Spring Boot application the deployer kicks in and looks for some configuration to tell it where to find the function jar. At a minimum the user has to provide a function.location which is a URL or resource location for the archive containing the functions. It can optionally use a maven: prefix to locate the artifact via a dependency lookup (see FunctionProperties for complete details). A Spring Boot application is bootstrapped from the jar file, using the MANIFEST.MF to locate a start class, so that a standard Spring Boot fat jar works well, for example. If the target jar can be launched successfully then the result is a function registered in the main application’s FunctionCatalog. The registered function can be applied by code in the main application, even though it was created in an isolated class loader (by deault). Serverless Platform Adapters As well as being able to run as a standalone process, a Spring Cloud Function application can be adapted to run one of the existing serverless platforms. In the project there are adapters for AWS Lambda, Azure, and Apache OpenWhisk. The Oracle Fn platform has its own Spring Cloud Function adapter. And Riff supports Java functions and its Java Function Invoker acts natively is an adapter for Spring Cloud Function jars.
AWS Lambda The AWS adapter takes a Spring Cloud Function app and converts it to a form that can run in AWS Lambda.
Introduction The adapter has a couple of generic request handlers that you can use. The most generic is SpringBootStreamHandler, which uses a Jackson ObjectMapper provided by Spring Boot to serialize and deserialize the objects in the function. There is also a SpringBootRequestHandler which you can extend, and provide the input and output types as type parameters (enabling AWS to inspect the class and do the JSON conversions itself). If your app has more than one @Bean of type Function etc. then you can choose the one to use by configuring function.name (e.g. as FUNCTION_NAME environment variable in AWS). The functions are extracted from the Spring Cloud FunctionCatalog (searching first for Function then Consumer and finally Supplier).
Notes on JAR Layout You don’t need the Spring Cloud Function Web or Stream adapter at runtime in Lambda, so you might need to exclude those before you create the JAR you send to AWS. A Lambda application has to be shaded, but a Spring Boot standalone application does not, so you can run the same app using 2 separate jars (as per the sample here). The sample app creates 2 jar files, one with an aws classifier for deploying in Lambda, and one executable (thin) jar that includes spring-cloud-function-web at runtime. Spring Cloud Function will try and locate a "main class" for you from the JAR file manifest, using the Start-Class attribute (which will be added for you by the Spring Boot tooling if you use the starter parent). If there is no Start-Class in your manifest you can use an environment variable MAIN_CLASS when you deploy the function to AWS.
Build ./mvnw -U clean package
Upload Build the sample under spring-cloud-function-samples/function-sample-aws and upload the -aws jar file to Lambda. The handler can be example.Handler or org.springframework.cloud.function.adapter.aws.SpringBootStreamHandler (FQN of the class, not a method reference, although Lambda does accept method references). Using the AWS command line tools it looks like this: aws lambda create-function --function-name Uppercase --role arn:aws:iam::[USERID]:role/service-role/[ROLE] --zip-file fileb://function-sample-aws/target/function-sample-aws-1.0.0.BUILD-SNAPSHOT-aws.jar --handler org.springframework.cloud.function.adapter.aws.SpringBootStreamHandler --description "Spring Cloud Function Adapter Example" --runtime java8 --region us-east-1 --timeout 30 --memory-size 1024 --publish The input type for the function in the AWS sample is a Foo with a single property called "value". So you would need this to test it: { "value": "test" }
Azure Functions The Azure adapter bootstraps a Spring Cloud Function context and channels function calls from the Azure framework into the user functions, using Spring Boot configuration where necessary. Azure Functions has quite a unique, but invasive programming model, involving annotations in user code that are specific to the platform. The Spring Cloud Function Azure adapter trades the convenience of these annotations for portability of the function implementations. Instead of using the annotations you have to write some JSON by hand (at least for now) to guide the platform to call the right methods in the adapter. The adapter has a generic http request handler that you can use. There is a AzureSpringBootRequestHandler which you must extend, and provide the input and output types as type parameters (enabling Azure to inspect the class and do the JSON conversions itself). If your app has more than one @Bean of type Function etc. then you can choose the one to use by configuring function.name. The functions are extracted from the Spring Cloud FunctionCatalog.
Notes on JAR Layout You don’t need the Spring Cloud Function Web at runtime in Azure, so you need to exclude this before you create the JAR you deploy to Azure. A function application on Azure has to be shaded, but a Spring Boot standalone application does not, so you can run the same app using 2 separate jars (as per the sample here). The sample app creates the shaded jar file, with an azure classifier for deploying in Azure.
JSON Configuration The Azure tooling needs to find some JSON configuration files to tell it how to deploy and integrate the function (e.g. which Java class to use as the entry point, and which triggers to use). Those files can be created with the Maven plugin for a non-Spring function, but the tooling doesn’t work yet with the adapter in its current form. There is an example function.json in the sample which hooks the function up as an HTTP endpoint: { "scriptFile" : "../function-sample-azure-1.0.0.BUILD-SNAPSHOT-azure.jar", "entryPoint" : "example.FooHandler.execute", "bindings" : [ { "type" : "httpTrigger", "name" : "foo", "direction" : "in", "authLevel" : "anonymous", "methods" : [ "get", "post" ] }, { "type" : "http", "name" : "$return", "direction" : "out" } ], "disabled" : false }
Build ./mvnw -U clean package
Running the sample You can run the sample locally, just like the other Spring Cloud Function samples: and curl -H "Content-Type: text/plain" localhost:8080/function -d '{"value": "hello foobar"}'. You will need the az CLI app and some node.js fu (see https://docs.microsoft.com/en-us/azure/azure-functions/functions-create-first-java-maven for more detail). To deploy the function on Azure runtime: $ az login $ mvn azure-functions:deploy On another terminal try this: curl https://<azure-function-url-from-the-log>/api/uppercase -d '{"value": "hello foobar!"}'. Please ensure that you use the right URL for the function above. Alternatively you can test the function in the Azure Dashboard UI (click on the function name, go to the right hand side and click "Test" and to the bottom right, "Run"). The input type for the function in the Azure sample is a Foo with a single property called "value". So you need this to test it with something like below: { "value": "foobar" }
Apache Openwhisk The OpenWhisk adapter is in the form of an executable jar that can be used in a a docker image to be deployed to Openwhisk. The platform works in request-response mode, listening on port 8080 on a specific endpoint, so the adapter is a simple Spring MVC application.
Quick Start Implement a POF (be sure to use the functions package): package functions; import java.util.function.Function; public class Uppercase implements Function<String, String> { public String apply(String input) { return input.toUpperCase(); } } Install it into your local Maven repository: ./mvnw clean install Create a function.properties file that provides its Maven coordinates. For example: dependencies.function: com.example:pof:0.0.1-SNAPSHOT Copy the openwhisk runner JAR to the working directory (same directory as the properties file): cp spring-cloud-function-adapters/spring-cloud-function-adapter-openwhisk/target/spring-cloud-function-adapter-openwhisk-1.0.0.BUILD-SNAPSHOT.jar runner.jar Generate a m2 repo from the --thin.dryrun of the runner JAR with the above properties file: java -jar -Dthin.root=m2 runner.jar --thin.name=function --thin.dryrun Use the following Dockerfile: FROM openjdk:8-jdk-alpine VOLUME /tmp COPY m2 /m2 ADD runner.jar . ADD function.properties . ENV JAVA_OPTS="" ENTRYPOINT [ "java", "-Djava.security.egd=file:/dev/./urandom", "-jar", "runner.jar", "--thin.root=/m2", "--thin.name=function", "--function.name=uppercase"] EXPOSE 8080
you could use a Spring Cloud Function app, instead of just a jar with a POF in it, in which case you would have to change the way the app runs in the container so that it picks up the main class as a source file. For example, you could change the ENTRYPOINT above and add --spring.main.sources=com.example.SampleApplication.
Build the Docker image: docker build -t [username/appname] . Push the Docker image: docker push [username/appname] Use the OpenWhisk CLI (e.g. after vagrant ssh) to create the action: wsk action create example --docker [username/appname] Invoke the action: wsk action invoke example --result --param payload foo { "result": "FOO" }