Spring Cloud Data Flow’s architectural style is different than other Stream and Batch processing platforms. For example in Apache Spark, Apache Flink, and Google Cloud Dataflow applications run on a dedicated compute engine cluster. The nature of the compute engine gives these platforms a richer environment for performing complex calculations on the data as compared to Spring Cloud Data Flow, but it introduces complexity of another execution environment that is often not needed when creating data centric applications. That doesn’t mean you cannot do real time data computations when using Spring Cloud Data Flow. Refer to the analytics section which describes the integration of Redis to handle common counting based use-cases as well as the RxJava integration for functional API driven analytics use-cases, such as time-sliding-window and moving-average among others.

Similarly, Apache Storm, Hortonworks DataFlow and Spring Cloud Data Flow’s predecessor, Spring XD, use a dedicated application execution cluster, unique to each product, that determines where your code should execute on the cluster and perform health checks to ensure that long lived applications are restarted if they fail. Often, framework specific interfaces are required to be used in order to correctly “plug in” to the cluster’s execution framework.

As we discovered during the evolution of Spring XD, the rise of multiple container frameworks in 2015 made creating our own runtime a duplication of efforts. There is no reason to build your own resource management mechanics, when there are multiple runtime platforms that offer this functionality already. Taking these considerations into account is what made us shift to the current architecture where we delegate the execution to popular runtimes, runtimes that you may already be using for other purposes. This is an advantage in that it reduces the cognitive distance for creating and managing data centric applications as many of the same skills used for deploying other end-user/web applications are applicable.

Spring Cloud Stream is most closely integrated with Spring Integration’s imperative "event at a time" programming model. This means you write code that handles a single event callback. For example,

In this case the String payload of a message coming on the input channel, is handed to the log method. The @EnableBinding annotation is what is used to tie together the input channel to the external middleware.

However, Spring Cloud Stream can support other programming styles. The use of reactive APIs where incoming and outgoing data is handled as continuous data flows and it defines how each individual message should be handled. You can also use operators that describe functional transformations from inbound to outbound data flows. The upcoming versions will support Apache Kafka’s KStream API in the programming model.

The Stream DSL describes linear sequences of data flowing through the system. For example, in the stream definition http | transformer | cassandra, each pipe symbol connects the application on the left to the one on the right. Named channels can be used for routing and to fan out data to multiple messaging destinations.

Taps can be used to ‘listen in’ to the data that if flowing across any of the pipe symbols. Taps can be used as sources for new streams with an in independent life cycle.

For an application that will consume events, Spring Cloud stream exposes a concurrency setting that controls the size of a thread pool used for dispatching incoming messages. See the Consumer properties documentation for more information.

A common pattern in stream processing is to partition the data as it moves from one application to the next. Partitioning is a critical concept in stateful processing, for either performance or consistency reasons, to ensure that all related data is processed together. For example, in a time-windowed average calculation example, it is important that all measurements from any given sensor are processed by the same application instance. Alternatively, you may want to cache some data related to the incoming events so that it can be enriched without making a remote procedure call to retrieve the related data.

Spring Cloud Data Flow supports partitioning by configuring Spring Cloud Stream’s output and input bindings. Spring Cloud Stream provides a common abstraction for implementing partitioned processing use cases in a uniform fashion across different types of middleware. Partitioning can thus be used whether the broker itself is naturally partitioned (e.g., Kafka topics) or not (e.g., RabbitMQ). The following image shows how data could be partitioned into two buckets, such that each instance of the average processor application consumes a unique set of data.

To use a simple partitioning strategy in Spring Cloud Data Flow, you only need set the instance count for each application in the stream and a partitionKeyExpression producer property when deploying the stream. The partitionKeyExpression identifies what part of the message will be used as the key to partition data in the underlying middleware. An ingest stream can be defined as http | averageprocessor | cassandra (Note that the Cassandra sink isn’t shown in the diagram above). Suppose the payload being sent to the http source was in JSON format and had a field called sensorId. Deploying the stream with the shell command stream deploy ingest --propertiesFile ingestStream.properties where the contents of the file ingestStream.properties are

will deploy the stream such that all the input and output destinations are configured for data to flow through the applications but also ensure that a unique set of data is always delivered to each averageprocessor instance. In this case the default algorithm is to evaluate payload.sensorId % partitionCount where the partitionCount is the application count in the case of RabbitMQ and the partition count of the topic in the case of Kafka.

Please refer to [passing_stream_partition_properties] for additional strategies to partition streams during deployment and how they map onto the underlying Spring Cloud Stream Partitioning properties.

Also note, that you can’t currently scale partitioned streams. Read the section Scaling at runtime for more information.

Streams are composed of applications that use the Spring Cloud Stream library as the basis for communicating with the underlying messaging middleware product. Spring Cloud Stream also provides an opinionated configuration of middleware from several vendors, in particular providing persistent publish-subscribe semantics.

The Binder abstraction in Spring Cloud Stream is what connects the application to the middleware. There are several configuration properties of the binder that are portable across all binder implementations and some that are specific to the middleware.

For consumer applications there is a retry policy for exceptions generated during message handling. The retry policy is configured using the common consumer properties maxAttempts, backOffInitialInterval, backOffMaxInterval, and backOffMultiplier. The default values of these properties will retry the callback method invocation 3 times and wait one second for the first retry. A backoff multiplier of 2 is used for the second and third attempts.

When the number of retry attempts has exceeded the maxAttempts value, the exception and the failed message will become the payload of a message and be sent to the application’s error channel. By default, the default message handler for this error channel logs the message. You can change the default behavior in your application by creating your own message handler that subscribes to the error channel.

Spring Cloud Stream also supports a configuration option for both Kafka and RabbitMQ binder implementations that will send the failed message and stack trace to a dead letter queue. The dead letter queue is a destination and its nature depends on the messaging middleware (e.g in the case of Kafka it is a dedicated topic). To enable this for RabbitMQ set the consumer properties republishtoDlq and autoBindDlq and the producer property autoBindDlq to true when deploying the stream. To always apply these producer and consumer properties when deploying streams, configure them as common application properties when starting the Data Flow server.

Additional messaging delivery guarantees are those provided by the underlying messaging middleware that is chosen for the application for both producing and consuming applications. Refer to the Kafka Consumer and Producer and Rabbit Consumer and Producer documentation for more details. You will find extensive declarative support for all the native QOS options.

The Data Flow Server uses an embedded servlet container and exposes REST endpoints for creating, deploying, undeploying, and destroying streams and tasks, querying runtime state, analytics, and the like. The Data Flow Server is implemented using Spring’s MVC framework and the Spring HATEOAS library to create REST representations that follow the HATEOAS principle.

Each Data Flow Server executable jar targets a single runtime by delegating to the implementation of the deployer Service Provider Interface found on the classpath.

We provide a Data Flow Server executable jar that targets a single runtime. The Data Flow server delegates to the implementation of the deployer Service Provider Interface found on the classpath. In the current version, there are no endpoints specific to a target runtime, but may be available in future releases as a convenience to access runtime specific features

While we provide a server executable for each of the target runtimes you can also create your own customized server application using Spring Initialzr. This let’s you add or remove functionality relative to the executable jar we provide. For example, adding additional security implementations, custom endpoints, or removing Task or Analytics REST endpoints. You can also enable or disable some features through the use of feature toggles.

The Data Flow Server executable jars support basic http, LDAP(S), File-based, and OAuth 2.0 authentication to access its endpoints. Refer to the security section for more information.

Authorization via groups is planned for a future release.

The target runtimes supported by Data Flow all have the ability to restart a long lived application should it fail. Spring Cloud Data Flow sets up whatever health probe is required by the runtime environment when deploying the application.

The collective state of all applications that comprise the stream is used to determine the state of the stream. If an application fails, the state of the stream will change from ‘deployed’ to ‘partial’.

Each target runtime lets you control the amount of memory, disk and CPU that is allocated to each application. These are passed as properties in the deployment manifest using key names that are unique to each runtime. Refer to the each platforms server documentation for more information.

When deploying a stream, you can set the instance count for each individual application that comprises the stream. Once the stream is deployed, each target runtime lets you control the target number of instances for each individual application. Using the APIs, UIs, or command line tools for each runtime, you can scale up or down the number of instances as required. Future work will provide a portable command in the Data Flow Server to perform this operation.

Currently, this is not supported with the Kafka binder (based on the 0.8 simple consumer at the time of the release), as well as partitioned streams, for which the suggested workaround is redeploying the stream with an updated number of instances. Both cases require a static consumer set up based on information about the total instance count and current instance index, a limitation intended to be addressed in future releases. For example, Kafka 0.9 and higher provides good infrastructure for scaling applications dynamically and will be available as an alternative to the current Kafka 0.8 based binder in the near future. One specific concern regarding scaling partitioned streams is the handling of local state, which is typically reshuffled as the number of instances is changed. This is also intended to be addressed in the future versions, by providing first class support for local state management.

Application versioning, that is upgrading or downgrading an application from one version to another, is not directly supported by Spring Cloud Data Flow. You must rely on specific target runtime features to perform these operational tasks.

The roadmap for Spring Cloud Data Flow will deploy applications that are compatible with Spinnaker to manage the complete application lifecycle. This also includes automated canary analysis backed by application metrics. Portable commands in the Data Flow server to trigger pipelines in Spinnaker are also planned.

You can access the server log by using the following command (just supply the name of pod for the server):

The streams apps are deployed with teh stream name followed by the name of the app and for processors and sinks there is also an instance index appended.

To see details for a specifc app deployment you can use (just supply the name of pod for the app):

For the application logs use:

If you would like to tail a log you can use:

Tasks are launched as bare pods without a replication controller. The pods remain after the tasks complete and this gives you an opportunity to review the logs.

To review the task logs use:

You have two options to delete completed pods. You can delete them manually once they are no longer needed.

To delete the task pod use:

You can also use the Data Flow shell command task execution cleanup command to remove the completed pod for a task execution.

First we need to determine the ID for the task execution:

Next we issue the command to cleanup the execution artifacts (the completed pod):

The bulk import applications page provides numerous options for defining and importing a set of applications in one go. For bulk import the application definitions are expected to be expressed in a properties style:

<type>.<name> = <coordinates>

For example:

task.timestamp=maven://org.springframework.cloud.task.app:timestamp-task:1.2.0.RELEASE

processor.transform=maven://org.springframework.cloud.stream.app:transform-processor-rabbit:1.2.0.RELEASE

At the top of the bulk import page an Uri can be specified that points to a properties file stored elsewhere, it should contain properties formatted as above. Alternatively, using the textbox labeled Apps as Properties it is possible to directly list each property string. Finally, if the properties are stored in a local file the Select Properties File option will open a local file browser to select the file. After setting your definitions via one of these routes, click Import.

At the bottom of the page there are quick links to the property files for common groups of stream apps and task apps. If those meet your needs, simply select your appropriate variant (rabbit, kafka, docker, etc) and click the Import action on those lines to immediately import all those applications.

Apps encapsulate a unit of work into a reusable component. Within the Data Flow runtime environment Apps allow users to create definitions for Streams as well as Tasks. Consequently, the Apps tab within the Tasks section allows users to create Task definitions.

On this screen you can perform the following actions:

This page lists the Data Flow Task definitions and provides actions to launch or destroy those tasks. It also provides a shortcut operation to define one or more tasks using simple textual input, indicated by the bulk define tasks button.

This page lists the Batch Job Executions and provides the option to restart or stop a specific job execution, provided the operation is available. Furthermore, you have the option to view the Job execution details.

The list of Job Executions also shows the state of the underlying Job Definition. Thus, if the underlying definition has been deleted, deleted will be shown.

Spring Cloud Data Flow builds upon Spring Cloud Deployer SPI and the platform specific dataflow-server uses the respective SPI implementations. Specifically, if we were to troubleshoot deployment specific issues; such as the network errors, it’d be useful to enable the DEBUG logs at the underlying deployer and the libraries used by it.

The streaming applications in Spring Cloud Data Flow are Spring Boot applications and they can be independently setup with logging configurations.

For instance, if you’d have to troubleshoot the header and payload specifics that are being passed around source, processor and sink channels, you’d be deploying the stream with the following options.

(where, org.springframework.integration is the global package for everything Spring Integration related, which is responsible for messaging channels)

These properties can also be specified via deployment properties when deploying the stream.

If you have custom Spring XD modules, you’d have to refactor them to use Spring Cloud Stream and Spring Cloud Task annotations, with updated dependencies and built as normal Spring Boot "applications".

Terminology wise, in Spring Cloud Data Flow, the message bus implementation is commonly referred to as binders.

A Task by definition, is any application that does not run forever, including Spring Batch jobs, and they end/stop at some point. Task applications can be majorly used for on-demand use-cases such as database migration, machine learning, scheduled operations etc. Using Spring Cloud Task, users can build Spring Batch jobs as microservice applications.

The Admin-UI is now renamed as Dashboard. The URI for accessing the Dashboard is changed from localhost:9393/admin-ui to localhost:9393/dashboard

Spring Cloud Data Flow comes with a significantly simplified architecture. In fact, when compared with Spring XD, there are less peripherals that are necessary to operationalize Spring Cloud Data Flow.

To support centralized and consistent management of an application’s configuration properties, Spring Cloud Config client libraries have been included into the Spring Cloud Data Flow server as well as the Spring Cloud Stream applications provided by the Spring Cloud Stream App Starters. You can also pass common application properties to all streams when the Data Flow Server starts.

Spring Cloud Data Flow is a Spring Boot application. Depending on the platform of your choice, you can download the respective release uber-jar and deploy/push it to the runtime platform (cloud foundry, apache yarn, kubernetes, or apache mesos). For example, if you’re running Spring Cloud Data Flow on Cloud Foundry, you’d download the Cloud Foundry server implementation and do a cf push as explained in the reference guide.

The hdfs-sink application builds upon Spring Hadoop 2.4.0 release, so this application is compatible with following Hadoop distributions.

Spring Cloud Data Flow can be deployed and used with Apche YARN in two different ways.

Let’s review some use-cases to compare and contrast the differences between Spring XD and Spring Cloud Data Flow.

There is a "full" profile that will generate documentation. You can build just the documentation by executing

$ ./mvnw clean package -DskipTests -P full -pl spring-cloud-dataflow-server-kubernetes-docs -am

If you don’t have an IDE preference we would recommend that you use Spring Tools Suite or Eclipse when working with the code. We use the m2eclipe eclipse plugin for maven support. Other IDEs and tools should also work without issue.

Before we accept a non-trivial patch or pull request we will need you to sign the contributor’s agreement. Signing the contributor’s agreement does not grant anyone commit rights to the main repository, but it does mean that we can accept your contributions, and you will get an author credit if we do. Active contributors might be asked to join the core team, and given the ability to merge pull requests.

None of these is essential for a pull request, but they will all help. They can also be added after the original pull request but before a merge.