MicroProfile Reactive Streams Operators Specification

The java.util.streams API provides functionality necessary for manipulating streams in memory introducing functional programming into the Java language. However, manipulating potentially infinite asynchronous streams has some very different requirements which the java.util.streams API is not suitable for.

Reactive Streams is a specification for asynchronous streaming between different libraries and/or technologies, which is included in JDK9 as the java.util.concurrent.Flow spec. Reactive Streams itself however is an SPI for library and technology vendors to implement, it is not intended that application developers implement it as the semantics are very complex.

Commonly, Reactive Streams requires more than just plumbing publishers to subscribers, the stream typically needs to be manipulated in some way, such as applying operations such as map, filter and flatMap. Neither Reactive Streams, nor the JDK, provides an API for doing these manipulations. Since users are not meant to implement Reactive Streams themselves, this means the only way to do these manipulations currently is to depend on a third party library providing operators, such as Akka Streams, RxJava or Reactor.

This API seeks to fill that gap, so that MicroProfile application developers can manipulate Reactive Streams without bringing in a third party dependency. By itself, this API is not interesting to MicroProfile, but with the addition of other Reactive features, such as the MicroProfile Reactive Messaging proposal, it is essential.

There are a number of different approaches to handling the cross cutting concerns relating to actually running streams. These include how and whether context is propagated, concurrency models, buffering, hooks for monitoring and remoting. Different implementations of Reactive Streams offer different approaches based on varying opinions on how these cross cutting concerns should be treated. For this reason, this API provides an underlying SPI to allow different engines to be plugged in to actually run the streams. Implementors of this spec will provide a default implementation, however users can select to use a custom implementation if they need.

This specification started as a proposal for the JDK, but after discussions there, it was decided that the JDK was not the right place to incubate such an API. It is the intention that once incubated in MicroProfile, this specification will be put to the JDK again as a proposal.

This specification aims to define a set of operators to manipulate Reactive Streams. The proposed API is voluntarily close to the java.utils.Stream API.

Reactive Streams has been included in the JDK9 as the java.util.concurrent.Flow API, which contains the Publisher, Subscriber, Subscription and Processor interfaces as nested interfaces of Flow. MicroProfile however is not ready to move to a baseline requirement for JDK9 or above.

For this reason, this proposal uses the JDK6 compatible org.reactivestreams API, which provides identical Publisher, Subscriber, Subscription and Processor interfaces as members of the org.reactivestreams package. This dependency contains nothing but those interfaces.

It has been discussed that MicroProfile could copy those interfaces itself, so as to not add this dependency, however this would most likely frustrate use of Reactive Streams in MicroProfile. There is a large ecosystem built around the org.reactivestreams interfaces. If application developers wanted to leverage that ecosystem in their application, they would have to write adapters to bridge the two APIs. Given that the org.reactivestreams dependency is a jar that contains only these interfaces, identical to the interfaces in JDK9, there doesn’t seem to be any value in copying them again into MicroProfile.

To facilitate an eventual migration to the JDK9 interfaces, once MicroProfile adopts JDK9 or later as a baseline JDK version, all methods that pass org.reactivestreams interfaces to the user (either as a return value, or by virtue of a user providing a callback to the method to receive it) will have Rs added to their name. For example, getSubscriber will be called getRsSubscriber. This will allow new methods to be added in future that return java.util.concurrent.Flow interfaces, without the Rs in the name, allowing the existing Rs methods to be kept for a limited time for backwards compatibility. Methods that accept a org.reactivestreams interface do not need to be given the same treatment, as support for the JDK9 interfaces can be added by overloading them, with backwards compatibility being maintained (see reactive approach for MicroProfile).

Each builder contains zero or more stages. There are three different shapes of stages:

Stream stages can be built into graphs, using the builders. There are four different shapes of graphs that can be built:

While building a stream, the stream may change shape during its construction. For example, a publisher may be collected into a List of elements. When this happens, the stream becomes a closed graph, since there is no longer an outlet, but just a result, the result being the List of elements:

Here’s an example of a more complex situation where a PublisherBuilder is plumbed to a SubscriberBuilder, producing a CompletionRunner:

When MicroProfile specifications provide an API that uses Reactive Streams, it is intended that application developers can return and pass the builder interfaces directly to the MicroProfile APIs. In many cases, application developers will not need to run the streams themselves. However, should they need to run the streams directly themselves, they can do so by using the streams build or run methods. PublisherBuilder, SubscriberBuilder and ProcessorBuilder all provide a build method that returns a Publisher, CompletionSubscriber and Processor respectively, while CompletionRunner, since it actually runs the stream, provides a run method that returns a CompletionStage.

The CompletionSubscriber class is so named because, where a CompletionStage is a stage of asynchronous computation that completes with a value or an error, a CompletionSubscriber is subscriber to an asynchronous stream that completes with a value or an error.

The build and run methods both provide a zero arg variant, which uses the default Reactive Streams engine provided by the platform, as well as a overload that takes a ReactiveStreamsEngine, allowing application developers to use a custom engine when they please.

Within a single stage in a single stream, implementations must guarantee a happens-before relationship between any invocation of all callbacks supplied to that stage.

Between different stages of a single stream, it is recommended that a happens-before relationship exists between callbacks of different stages, however this is not required, and end users of the API must not depend on this. Implementations are expected to implement this for performance reasons, this is also known as operator fusing. Implementations may decide, for whatever reason, not to fuse stages in certain circumstances, when this is done, it is said that there is an asynchronous boundary between the two stages.

When Reactive Streams interfaces, that is, Publisher, Subscriber or Processor, are wrapped in a stage, implementations may place an asynchronous boundary between that stage and its neighbors if they choose. Whether implementations do or don’t place an asynchronous boundary there, they must conform to the Reactive Streams specification with regards to memory visibility.

Exceptions thrown by user supplied callbacks must be caught and propagated downstream, unless otherwise handled by an error handling stage.

User exceptions must not be thrown by the build or run methods of the builders. For example, if a user supplies an Iterable as a source for a PublisherBuilder, and the iterator() method is invoked synchronously from the build method on the publisher, and it throws an exception, this exception must be caught, and propagated through the stream, not thrown from the build method.

An exception to this is the callbacks provided in user supplied Reactive Streams, ie Publisher, Subscriber and Processor. Since these interfaces are specified not to throw any exceptions when the consumer is adhering to the spec, implementations may assume that these interfaces won’t throw exceptions, and the behavior of an implementation should these interfaces throw exceptions is unspecified.

In some cases, exceptions may be wrapped, for example, CompletableFuture wraps exceptions in a CompletionException. It is recommended that implementations do not wrap exceptions if they don’t need to, but rather that they propagate them downstream as is.

Errors may be eagerly propagated by stages if an implementation chooses to do this. Such propagation can aid with fast failure of stages, to ensure things like connection failures do not wait excessively to discover that they have failed.

A consequence of this is that errors can in certain circumstances overtake elements. For example, the flatMapCompletionStage stage may receive an error while an element is being asynchronously processed. This error may immediately be propagated downstream, which will mean the eventual redemption of the CompletionStage returned by the mapper function for the stage will be ignored, and the element it is redeemed with will be dropped.

When a user does not desire this behavior, and wants to guarantee that stages don’t drop elements when upstream errors arrive, they may insert an error recovery stage, such as onErrorResume, before the stage that they don’t want to drop elements from. They will then need to implement an out of band mechanism to propagate the error, such as wrapping it in an element, should they wish to handle it downstream.

Implementations must assume that any PublisherBuilder, SubscriberBuilder or ProcessorBuilder supplied to them, or any stages within, potentially hold resources, and must be cleaned up when the stream shuts down. For example, a concat stage accepts two `PublisherBuilder’s. If the stream is cancelled before the first is finished, the second must still be built, and then cancelled too.

Reactive Streams does not allow null elements. Hence, any user callbacks that return null as an element to be emitted by a stage must cause the stage to fail with a NullPointerException.

If a MicroProfile container provides an implementation of MicroProfile Reactive Streams Operators, then it must make an application scoped ReactiveStreamsEngine available for injection.

This specification places no requirements on the propagation of CDI context, or what context(s) should be active when user supplied callbacks are executed during the running of a stream.

Other specifications that use this specification may require that implementations make certain context’s to be active when user callbacks are executed. In this case, it is expected that such specifications will have the responsibility of running the streams.

For example, a hypothetical WebSocket specification may allow user code to return a ProcessorBuilder to handle messages:

In this case, the implementation of that hypothetical WebSocket specification is responsible for running the ProcessorBuilder returned by the user, and that specification may require that the engine it uses to run it makes the request context active in callbacks, such as the map callback above. Since the implementation of that specification is in control of which engine is used to run the processor, this requirement can be made by that specification. It is the responsibility of that implementation (and the MicroProfile container that pulls these implementations together) to ensure that the MicroProfile Reactive Streams Operators implementation used is able to support CDI context propagation.

In contrast, if a user is responsible for executing a stream, like in the following hypothetical example:

Then there is no clear way for the container to control how the engine used there, which in this case would be loaded through the Java ServiceLoader API, would propagate context. For this reason, any cases where users are running their own streams are not expected to require any CDI context propagation, and it is recommended that specifications favour APIs where the container runs the stream, not the user, to facilitate context propagation.

This just shows the fluency of the API. It wouldn’t make sense to actually do the below in practice, since the JDK8 streams API itself is better for working with in memory streams.

This shows how common collection types can be converted to a Publisher of the elements in the collection.

The above example shows a very simple conversion of a List to a Publisher, of course other operations can be done on the elements before building the Publisher, in this case we go on to transform each object to a line in a CSV file, and then represent it as a stream of bytes.

This shows building a subscriber for a byte stream, such as for the JDK9 HttpClient API. It assumes another library has provided a Reactive Streams Processor that parses byte streams into streams of objects.

This shows building a processor, for example, a message library may require processing messages, and then emitting an ACK identifier so that each handled element can be acknowledged as handled.

A library may provide a Reactive Streams publisher that the application developer needs to consume. This shows how that can be done.

A library may provide a subscriber to feed a connection. This shows how that subscriber can be fed.

The goal of the Reactive Stream Operators specification is to define building blocks to enable the implementation of asynchronous processing of stream of data. On the other hand, the Java Stream API provides a synchronous approach to compute a result by analyzing data conveyed in a stream. Because of this asynchronous vs. synchronous processing, the terminal stages (such as collect, findFirst…​) define by this API return CompletableStage<T> and not T. Indeed, only when the result has been computed the returned CompletableStage is completed. As an example, here is the two versions of the same processing:

The asynchronous vs. synchronous difference also means that the error propagation works differently. In the Java Streams API, the processing can be wrapped in a try/catch construct. In the asynchronous case, the error is propagated into the returned future. In the example above, the function passed to the whenComplete stage receives the result as well as the failure (if any). If the processing throws an exception, the function can react by looking at the err parameter.

The Reactive Streams specification is intrinsically sequential. So none of the parallel processing ability from the Java Stream API are supported. As a consequence, the API does not provide a parallel​() method. Also, operations like findAny are not provided as the behavior would be equivalent to the provided findFirst method.

There are two parts to the TCK, an API TCK, and an SPI TCK. In addition, the whole TCK can either be run directly against an SPI implementation, or against a container with an Arquillian adapter.

The TCK uses TestNG for testing, and contains many test classes. These test classes are brought together using TestNG @Factory annotated methods, so that running the TCK can be done by simply running a single test class that is a factory for the rest of the test classes.

Additionally, this TCK also uses the Reactive Streams TCK. The Reactive Streams TCK is used to test Reactive Streams Publishers, Subscribers and Processors to ensure that they conform to the Reactive Streams spec. Wherever possible, this TCK will use it to verify the Publishers, Subscribers and Processors built by the API. For example, the returned Processor built by building a simple map stage is run through the Reactive Streams Processor TCK verification.

It should be noted that the Reactive Streams TCK is structured in such a way that each numbered requirement in the Reactive Streams specification has a test, even if that requirement is untestable by the TCK, or if its optional. In the case where requirements are optional or untested, the Reactive Streams TCK skips the test. Consequently, when running this TCK, there are a significant number of skipped tests.