Batch Data Processing with CDAP using Workflow

Source Code Repository: Source code (and other resources) for this guide are available at the CDAP Guides GitHub repository.

The Workflow system in Cask Data Application Platform (CDAP) allows specifying, executing, scheduling, and monitoring complex series of jobs and tasks. In this guide, you will learn how it can be used to execute the MapReduce programs in parallel based on the evaluation of conditions.

What You Will Build

This guide will take you through building a CDAP application that uses ingested raw purchase events (of the form '<name> bought <n> <item>s for $<price>', which are parsed using a primitive parser for sentences) to compute in parallel the total purchases made by each customer along with the total purchases made for each product.

You will:

  • Build PurchaseEventParser, a MapReduce program to parse the raw purchase events and create Purchase objects from them;
  • Build PurchaseCounterByCustomer, a MapReduce program to count the purchases made per customer;
  • Build PurchaseCounterByProduct, a MapReduce program to count the purchases made per product;
  • Build PurchaseWorkflow, a Workflow which will first execute the MapReduce program PurchaseEventParser. If the predicate PurchaseEventVerifier, which uses the MapReduce counters emitted by the PEP to determine data quality, evaluates to true, the workflow will in parallel execute the MapReduce program PurchaseCounterByCustomer and PurchaseCounterByProduct otherwise, it will execute the action ProblemLogger;
  • Use Datasets to persist results of the MapReduce programs; and
  • Build a Service to serve the results via HTTP.

Let’s Build It!

The following sections will guide you through building an application from scratch. If you are interested in deploying and running the application right away, you can clone its source code from this GitHub repository. In that case, feel free to skip the next two sections and jump right to the Build and Run Application section.

Application Design

The application will assume that the purchase events are ingested into a Stream. The events can be ingested into a Stream continuously in real time or in batches; whichever way, it doesn’t affect the ability of the MapReduce programs to consume them.

The PurchaseWorkflow encapsulates the set of MapReduce programs, which extracts the required information from the raw purchase events and computes the total purchases made by each customer and total purchases made for each product in a specific time range. The results of the computation are persisted in Datasets.

Finally, the application contains a Service that exposes an HTTP endpoint to access the data stored in the Datasets.



The first step is to construct our application structure. We will use a standard Maven project structure for all of the source code files:


The CDAP application is identified by the PurchaseWorkflowApp class. This class extends an AbstractApplication, and overrides the configure method to define all of the application components:

public class PurchaseWorkflowApp extends AbstractApplication {
  public void configure() {
    setDescription("Application describing the Workflow");

    addStream(new Stream("purchaseEvents"));

    addMapReduce(new PurchaseEventParser());
    addMapReduce(new PurchaseCounterByCustomer());
    addMapReduce(new PurchaseCounterByProduct());
    addWorkflow(new PurchaseWorkflow());

                       .setDescription("Schedule execution every 1 hour")
                       .createTimeSchedule("0 * * * *"),

    addService(new PurchaseResultService());

    createDataset("purchaseRecords", KeyValueTable.class);
    createDataset("customerPurchases", KeyValueTable.class);
    createDataset("productPurchases", KeyValueTable.class);

The PurchaseWorkflowApp application defines a new Stream where purchase events are ingested. Once the data is ingested, the events can be processed in real time or batch. In our application, we will process the events in batch using the PurchaseWorkflow program and compute the total purchases made by each customer and the total purchases made for each product in a specific time range. We will use three MapReduce programs PurchaseEventParser, PurchaseCounterByCustomer, and PurchaseCounterByProduct to apply different processing on the purchase events and the Workflow PurchaseWorkflow to connect these MapReduce programs.

The result of the Workflow execution is persisted into Datasets; the application uses the createDataset method to define the Dataset. We use three datasets: purchaseRecords to store the valid parsed purchase events; customerPurchases to store the total purchases made by each customer; and productPurchases to store the total purchases made for each product. The Purchase class defines the type used to store the parsed purchase events.

The application also adds a custom Workflow action ProblemLogger. When a Workflow executes a custom action, it invokes the run method in the action. In ProblemLogger, we only add a log statement; however it could be customized to send emails to the concerned parties.

The PurchaseWorkflow is scheduled to execute every hour.

Finally, the application adds a service for querying the results from the Datasets.

Let's take a closer look at the Workflow.

The PurchaseWorkflow extends an AbstractWorkflow class and overrides the configure method:

public class PurchaseWorkflow extends AbstractWorkflow {
  protected void configure() {
    setDescription("Workflow to parse the purchase events and count the revenue per customer and per product");


    condition(new PurchaseEventVerifier())
      .addAction(new ProblemLogger())

In the configure method we specify the topology for connecting the programs which will run as a part of the Workflow execution. As the first action in the PurchaseWorkflow, we add the MapReduce program PurchaseEventParser. This program will parse raw purchase events (using a primitive sentence parser) and create Purchase objects from them.

After that, we add a condition in the Workflow, which takes a predicate PurchaseEventVerifier. If the predicate evaluates to true, we fork the execution of the Workflow into two parallel branches. One branch executes the PurchaseCounterByCustomer MapReduce program, while the other executes the PurchaseCounterByProduct MapReduce program.

If the predicate evaluates to false, then actions in the otherwise section will be executed. We have added a single custom action, ProblemLogger to the otherwise section as an example of what is possible.

Lets take a closer look at the predicate PurchaseEventVerifier.

public class PurchaseEventVerifier implements Predicate<WorkflowContext> {

  private static final String TASK_COUNTER_GROUP_NAME = "org.apache.hadoop.mapreduce.TaskCounter";

  public boolean apply(WorkflowContext workflowContext) {
    if (workflowContext == null) {
      return false;

    WorkflowToken token = workflowContext.getToken();
    if (token == null) {
      return false;

    Value mapInputRecords = token.get(TASK_COUNTER_GROUP_NAME + "." + MAP_INPUT_RECORDS_COUNTER_NAME,
    Value mapOutputRecords = token.get(TASK_COUNTER_GROUP_NAME + "." + MAP_OUTPUT_RECORDS_COUNTER_NAME,
    if (mapInputRecords != null && mapOutputRecords != null) {
      // Return true if at least 80% of the records were successfully parsed and emitted
      // by previous map job
      return (mapOutputRecords.getAsLong() >= (mapInputRecords.getAsLong() * 80/100));
    return false;

PurchaseEventVerifier needs to be a public class which implements the interface Predicate<WorkflowContext>. The apply method in the predicate takes WorkflowContext as a parameter. The Hadoop counters emitted by the previous MapReduce program (in our case PurchaseEventParser) can be retrieved in this method using the workflowContext object. We query for the number of input records to the mappers and the number of records emitted by the mappers. If at least 80% of the records were successfully parsed and emitted as Purchase by the mappers, the method returns true and the fork in the Workflow will be executed. If the method returns false, the otherwise section in the condition is executed, which contains the ProblemLogger custom action.

Build and Run Application

The PurchaseWorkflowApp can be built and packaged using the Apache Maven command:

$ mvn clean package

Note that the remaining commands assume that the script is available on your PATH. If this is not the case, please add it:

$ export PATH=$PATH:<CDAP home>/bin

If you haven't already started a standalone CDAP installation, start it with the command:

$ start

We can then deploy the application to the standalone CDAP installation:

$ load artifact target/cdap-workflow-guide-<version>.jar
$ create app PurchaseWorkflowApp cdap-workflow-guide <version> user

Next, we will send some sample purchase events into the stream for processing:

$ send stream purchaseEvents '"bob bought 3 apples for $30"'
$ send stream purchaseEvents '"joe bought 1 apple for $100"'
$ send stream purchaseEvents '"joe bought 10 pineapples for $20"'
$ send stream purchaseEvents '"cat bought 3 bottles for $12"'
$ send stream purchaseEvents '"cat bought 2 pops for $14"'

We can now start the Workflow to process the events that were ingested:

$ start workflow PurchaseWorkflowApp.PurchaseWorkflow

The Workflow will take a couple of minutes to execute.

We can then start the PurchaseResultService and query the processed results:

$ start service PurchaseWorkflowApp.PurchaseResultService
  • Retrieve the purchase records for customer joe:

    $ curl http://localhost:10000/v3/namespaces/default/apps/PurchaseWorkflowApp/services/PurchaseResultService/methods/purchaserecords/joe

    Example output:

  • Retrieve the total purchases made by customer joe:

    $ curl http://localhost:10000/v3/namespaces/default/apps/PurchaseWorkflowApp/services/PurchaseResultService/methods/purchases/customers/joe

    Example output:

  • Retrieve the total purchases made for product apple:

    $ curl http://localhost:10000/v3/namespaces/default/apps/PurchaseWorkflowApp/services/PurchaseResultService/methods/purchases/products/apple

    Example output:


You have now seen how to write a Workflow to connect different MapReduce programs and run them in parallel based on a condition.

Share and Discuss!

Have a question? Discuss at the CDAP User Mailing List.


Copyright © 2015 Cask Data, Inc.

Licensed under the Apache License, Version 2.0 (the "License"); you may not use this file except in compliance with the License. You may obtain a copy of the License at

Unless required by applicable law or agreed to in writing, software distributed under the License is distributed on an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. See the License for the specific language governing permissions and limitations under the License.