Reactive Streams is a specification that determines the minimum set of interfaces required to build the asynchronous processing of a large volume of unbounded data. It is a specification aimed at JVM and JavaScript runtime. The main goal of the Reactive Streams specification is to standardize the exchange of stream data across an asynchronous boundary of applications. The API consists of the following four interfaces:

1. Publisher: The publisher is responsible for the generation of an unbounded number of asynchronous events and pushing those events to the associated subscribers.

2. Subscriber: The subscriber is a consumer of the events published by a publisher. The subscriber gets events for subscription, data, completion, and error. It can choose to perform actions on any of them.

3. Subscription: A subscription is a shared context between the publisher and subscriber, for the purpose of mediating the data exchange between the two. The subscription is available with the subscriber only, and enables it to control the flow of events from the publisher. The subscription becomes invalid if there is an error or a completion. A subscriber can also cancel the subscriptions, in order to close its stream.

4. Processor: The processor represents a stage of data processing between a subscriber and a publisher. Consequently, it is bound by both of them. The processor has to obey the contract between the publisher and the subscriber. If there is an error, it must propagate it back to the subscriber.

While there are only four interfaces, there are around 30 rules that govern the data exchange between the publisher and the subscriber. These rules are based on the two principles covered in the following sections.

Asynchronous execution refers to the ability to execute tasks without having to wait to finish previously executed tasks first. The execution model decouples tasks, so that each of them can be performed simultaneously, utilizing the available hardware.

The Reactive Streams API delivers events in an asynchronous manner. A publisher can generate event data in a synchronous blocking manner. On the other hand, each of the on-event handlers can process the events in a synchronously blocking manner. However, event publishing must occur asynchronously. It must not be blocked by the subscriber while processing events.

A subscriber can control events in its queue to avoid any overruns. It can also request more events if there is additional capacity. Backpressure enforces the publisher to bound the event queues according to the subscriber. Furthermore, a subscriber can ask to receive one element at a time, building a stop-and-wait protocol. It can also ask for multiple elements. On the other hand, a publisher can apply the appropriate buffers to hold non-delivered events, or it can just start to drop events if the production rate is more than the consumption rate.

It is important to note that the Reactive Streams API is aimed at the flow of events between different systems. Unlike ReactiveX, it does not provide any operators to perform transformations. The API has been adopted as a part of the java.util.concurrent.Flow package in JDK 9.

David Karnok, a veteran of various reactive projects like Rxjava and Reactor, has categorized the evolution of reactive libraries into the following generations.

The zero generation revolves around the java.util.observable interface and the related callbacks. It essentially uses the observable design pattern for reactive development. It lacks the necessary support of composition, operators, and backpressure.

The first generation represents Erik Mejer's attempt to address reactive issues by building Rx.NET. This referred to implementations in the form of the IObserver and IObservable interfaces. The overall design was synchronous and lacked backpressure.

The first generation deficiencies of backpressure and synchronous handling were handled in the second generation APIs. This generation refers to the first implementations of Reactive Extensions, such as RxJava 1.X and Akka.

The third generation refers to the Reactive Streams specification, which enables library implementors to be compatible with each other and compose sequences, cancellations, and backpressure across boundaries. It also enables an end user to switch between implementations at their own will.

The fourth generation refers to the fact that reactive operators can be combined in an external or internal fashion, leading to performance optimization. A fourth generation reactive API looks like a third generation, but internally, the operators have changed significantly to yield intended benefits. Reactor 3.0 and RxJava 2.x belong to this generation.

The fifth generation refers to a future work, in which there will be a need for bidirectional reactive I/O operations over the streams.