Introducing Java Functions for Spring Cloud Stream Applications - Part 1

Last week Spring posted Introducing Java Functions for Spring Cloud Stream Applications - Part 0
to announce the release of Spring Cloud Stream applications 2020.0.0-M2.
Here, explore function composition, one of the more powerful features enabled by the function oriented architecture presented in Part 0. If you haven’t had a chance to read Part 0, now would be a great time!

Function Composition

Function composition has a solid theoretical foundation in mathematics and computer science.
In practical terms, it is a way to join a sequence of functions to create a more complex function.

Let’s look at a simple example using Java functions. We have two functions, reverse and upper.
Each accepts a String as input and produces a String as output. We can compose them using the built-in andThen method. The composite function is itself a Function<String, String>.
If you run this, it will print ESREVER.

Function<String, String> reverse = s -> new StringBuilder(s).reverse().toString();
Function<String, String> upper = String::toUpperCase;
Function<String, String> reverseUpper = reverse.andThen(upper);
System.out.println(reverseUpper.apply("reverse"));

Tip
in addition to andThen, java.util.Function includes compose which first applies the argument (b) and then applies a to the result.
Thus, a.compose(b).apply(s) is equivalent to a.apply(b.apply(s)).
Function Composition in Spring Cloud Function
Spring Cloud Function includes some great features to take composing functions to another level.

Declarative Composition

If we define our functions from the above example as Spring beans,

@Bean
Function<String, String> reverse() {
    return s -> new StringBuilder(s).reverse().toString();
}

@Bean
Function<String, String> upper() {
    return String::toUpperCase;
}

we can compose these functions using the spring.cloud.function.definition property spring.cloud.function.definition=upper|reverse

Here | is a composition operator which results in an auto-configured bean implementing the composite function, along with related resources to let you seamlessly invoke the composite function.

Composition With Supplier and Consumer
Spring Cloud Function extends native Java Function composition to support composition with Supplier and Consumer.

This follows from concepts which are implicitly true:

A Function composed with a Consumer is a Consumer

A Supplier composed with a Function is a Supplier

A Supplier composed with a Consumer is a valid processing model (with no inputs or outputs, this form of composition does not map to a functional interface, but is analogous to Runnable)

As we shall see, Spring Cloud Stream Applications employ these concepts to great effect.

Type Conversion
When using function composition, we have to consider compatible argument types.
Using native Java composition, we can do compose a Function<Integer,String> with a Function<String, Integer> into a Function<Integer, Integer> :

Function<Integer, String> intToStr = String::valueOf;
Function<String, Integer> doubleit = i -> Integer.parseInt(i) * 2;
Function<Integer, Integer> composite = intToStr.andThen(doubleit);
composite.apply(10);
When running a Spring application, Spring Cloud Function uses Spring’s standard type conversion support to coerce function arguments as needed.
Given the following Function bean definitions, the function definition intToStr|doubleit works as expected, converting the String to an Integer.

@Bean
Function<Integer, Integer> doubleit() {
    return i -> i * 2;
}

@Bean
Function<Integer, String> intToStr() {
return String::valueOf;
}

In addition to converting primitives, Spring functions can convert between Message and POJO, JSON String and POJO, and more.
For example, the following functions can be composed in either order:

@Bean
Function<Integer, Integer> doubleit() {
    return i -> i * 2;
}

@Bean
Function<Integer, Message<String>> convertIntMessage() {
    return i -> MessageBuilder.withPayload(String.valueOf(i)).build();
}

Function Composition in Spring Cloud Stream
Spring Cloud Stream 3.x builds on Spring Cloud Function to fully support a functional programming model. The fundamental premise of Spring Cloud Stream is that it enables a function to execute in a distributed environment. The binder binds the input(s) and output(s) of a function packaged in a Spring Boot application, to configured message broker destinations so that the output produced by one function is consumed as the input of another remotely running function. We can think of a data streaming pipeline as just a distributed composition of functional components.

To illustrate this, a typical Spring Cloud Stream pipeline like

source | processor1 | processor2 | processor3 | sink
is logically equivalent to

supplier | function1 | function2 | function3 | sink
This idea leads to some interesting architectural choices since we can use function composition to combine some or all of these components into a single application.

For example we can implement the sequence of three processors as a single application, let’s call it composed-processor, packaging function1, function2, and function3,and composed by spring.cloud.function.definition=function1|function2|function3. Now the pipeline can be deployed as:

source | composed-processor | sink
Even simpler, we can create a composed-source to do all the processing within the source:

composed-source | sink
As always, there is no right answer here. There are always trade-offs to consider:

Function composition results in less deployments. This reduces cost, latency, operational complexity, and so on.

Individual deployments are loosely coupled and can scale independently.

The message broker provides guaranteed delivery. When a simple stateless application goes down and is restarted, it can continue where it left off, processing the pending results of the previous processing step.

A single application that performs complex processing is harder to reason about and keeps intermediate processing results in memory, or possibly in an interim data store. When a stateful application fails, it can lead to inconsistent state, making recovery harder.

If these trade-offs look familiar, it’s because they are pretty much the same as any microservice vs monolith debate. In the end, do what works best for you.

Function Composition with Prepackaged Source Applications
In some cases, function composition is a no-brainer. From the start, we have provided pre-packaged processors to perform simple transformations, or filtering using SpEL. The legacy architecture required a separate processor when using the prepackaged sources or sinks. A common complaint from users was “why do I need to deploy a separate application just to evaluate a SpEL expression?” To address this, we initially introduced a form of support for function composition in an earlier release. To use this feature with the prepackaged applications required forking them to modify the code or the build dependencies to provide the functions.

The current release provides function composition out of the box for all of the prepackaged sources. Specifically, a source can now be composed with prepackaged functions to perform any of the following locally:

execute SpEL transformations

enrich message headers

filter events

produce task launch requests

As an example, we can compose the time source with a header enricher and filter with configuration properties and run it as a standalone Spring boot application:

java -jar target/time-source-rabbit-3.0.0-SNAPSHOT.jar
--spring.cloud.stream.bindings.output.destination=even       --spring.cloud.function.definition=timeSupplier|headerEnricherFunction|filterFunction
--header.enricher.headers=seconds=T(java.lang.Integer).valueOf(payload .substring(payload.length() - 2))
--filter.function.expression=headers[seconds]%2==0
This will publish the time, such as `07/16/20 16:43:48, every other second whenever the number of seconds is even, to the configured destination even.

Here we are using a prepackaged time source for RabbitMQ, binding the output to a topic exchange named even. The binder will create the exchange if it does not exist. The function definition extends the supplier to extract the seconds, convert it to an integer and store it in the seconds message header and then filter on the value of the header. Only even values pass the filter.

Task Launch Requests
In 2018, we introduced a reference architecture for running file ingest with Spring Cloud Data Flow and Spring Batch. To do this, we forked the sftp source as sftp-dataflow, specifically to implement a prepackaged source that produces task launch requests. The task launch request is a simple value object, rendered as JSON, and consumed by the tasklauncher-sink. The sink acts as a client to Data Flow to launch a batch application per the request. We initially chose sftp since it is the most commonly used protocol for file processing. However, we realized that the same pattern can be applied to any source. We now can do this with function composition. Along with the standard sftp source , we can trigger a task launch from ftp, file, s3, and so one. Even the time source can be used to launch a task at regular intervals.

This somewhat contrived example produces task launch requests:

java -jar target/time-source-rabbit-3.0.0-SNAPSHOT.jar
--spring.cloud.stream.bindings.output.destination=time-test
--spring.cloud.stream.function.definition=timeSupplier|spelFunction|headerEnricherFunction|taskLaunchRequestFunction
--spel.function.expression=payload.length()
--header.enricher.headers=task-id=payload*2
--task.launch.request.task-name-expression="'task-'+headers['task-id']
The payload, as JSON, is {"args":[],"deploymentProps":{},"name":"task-34"}

Function composition with user written code
In reality, when users develop a Spring Cloud Stream pipeline, they are likely to select a source and sink from our prepackaged Spring Cloud Stream Applications. Processors are typically user-written code, implementing specific business logic. If you are writing a processor, or want to extend a source or sink, any of the functions are available to you. Since we publish the functions as separate artifacts, you can simply include them in your dependencies. You can either use declarative composition, as shown above, or you can inject them into your code and invoke them programmatically. Of course, you can easily integrate your own functions as well.