Basic API Concepts

Flink programs are regular programs that implement transformations on distributed collections (e.g., filtering, mapping, updating state, joining, grouping, defining windows, aggregating). Collections are initially created from sources (e.g., by reading files, kafka, or from local collections). Results are returned via sinks, which may for example write the data to (distributed) files, or to standard output (for example the command line terminal). Flink programs run in a variety of contexts, standalone, or embedded in other programs. The execution can happen in a local JVM, or on clusters of many machines.

Depending on the type of data sources, i.e. bounded or unbounded sources you would either write a batch program or a streaming program where the DataSet API is used for the former and the DataStream API is used for the latter. This guide will introduce the basic concepts that are common to both APIs but please see our Streaming Guide and Batch Guide for concrete information about writing programs with each API.

NOTE: When showing actual examples of how the APIs can be used we will use StreamingExecutionEnvironment and the DataStream API. The concepts are exactly the same in the DataSet API, just replace by ExecutionEnvironment and DataSet.

• Linking with Flink
• DataSet and DataStream
• Anatomy of a Flink Program
• Lazy Evaluation
• Specifying Keys
• Specifying Transformation Functions
• Supported Data Types
• Execution Configuration
• Program Packaging and Distributed Execution
• Accumulators & Counters
• Parallel Execution
  • Operator Level
  • Execution Environment Level
  • Client Level
  • System Level
• Execution Plans

Linking with Flink

To write programs with Flink, you need to include the Flink library corresponding to your programming language in your project.

The simplest way to do this is to use one of the quickstart scripts: either for Java or for Scala. They create a blank project from a template (a Maven Archetype), which sets up everything for you. To manually create the project, you can use the archetype and create a project by calling:

mvn archetype:generate \
    -DarchetypeGroupId=org.apache.flink \
    -DarchetypeArtifactId=flink-quickstart-java \
    -DarchetypeVersion=1.1.5

mvn archetype:generate \
    -DarchetypeGroupId=org.apache.flink \
    -DarchetypeArtifactId=flink-quickstart-scala \
    -DarchetypeVersion=1.1.5

The archetypes are working for stable releases and preview versions (-SNAPSHOT).

If you want to add Flink to an existing Maven project, add the following entry to your dependencies section in the pom.xml file of your project:

<!-- Use this dependency if you are using the DataStream API -->
<dependency>
  <groupId>org.apache.flink</groupId>
  <artifactId>flink-streaming-java_2.10</artifactId>
  <version>1.1.5</version>
</dependency>
<!-- Use this dependency if you are using the DataSet API -->
<dependency>
  <groupId>org.apache.flink</groupId>
  <artifactId>flink-java</artifactId>
  <version>1.1.5</version>
</dependency>
<dependency>
  <groupId>org.apache.flink</groupId>
  <artifactId>flink-clients_2.10</artifactId>
  <version>1.1.5</version>
</dependency>

<!-- Use this dependency if you are using the DataStream API -->
<dependency>
  <groupId>org.apache.flink</groupId>
  <artifactId>flink-streaming-scala_2.10</artifactId>
  <version>1.1.5</version>
</dependency>
<!-- Use this dependency if you are using the DataSet API -->
<dependency>
  <groupId>org.apache.flink</groupId>
  <artifactId>flink-scala_2.10</artifactId>
  <version>1.1.5</version>
</dependency>
<dependency>
  <groupId>org.apache.flink</groupId>
  <artifactId>flink-clients_2.10</artifactId>
  <version>1.1.5</version>
</dependency>

import org.apache.flink.api.scala._

import org.apache.flink.api.scala.createTypeInformation

Scala Dependency Versions

Because Scala 2.10 binary is not compatible with Scala 2.11 binary, we provide multiple artifacts to support both Scala versions.

Starting from the 0.10 line, we cross-build all Flink modules for both 2.10 and 2.11. If you want to run your program on Flink with Scala 2.11, you need to add a _2.11 suffix to the artifactId values of the Flink modules in your dependencies section.

If you are looking for building Flink with Scala 2.11, please check build guide.

Hadoop Dependency Versions

If you are using Flink together with Hadoop, the version of the dependency may vary depending on the version of Hadoop (or more specifically, HDFS) that you want to use Flink with. Please refer to the downloads page for a list of available versions, and instructions on how to link with custom versions of Hadoop.

In order to link against the latest SNAPSHOT versions of the code, please follow this guide.

The flink-clients dependency is only necessary to invoke the Flink program locally (for example to run it standalone for testing and debugging). If you intend to only export the program as a JAR file and run it on a cluster, you can skip that dependency.

Back to top

DataSet and DataStream

Flink has the special classes DataSet and DataStream to represent data in a program. You can think of them as immutable collections of data that can contain duplicates. In the case of DataSet the data is finite while for a DataStream the number of elements can be unbounded.

These collections differ from regular Java collections in some key ways. First, they are immutable, meaning that once they are created you cannot add or remove elements. You can also not simply inspect the elements inside.

A collection is initially created by adding a source in a Flink program and new collections are derived from these by transforming them using API methods such as map, filter and so on.

Anatomy of a Flink Program

Flink program programs look like regular programs that transform collections of data. Each program consists of the same basic parts:

1. Obtain an execution environment,
2. Load/create the initial data,
3. Specify transformations on this data,
4. Specify where to put the results of your computations,
5. Trigger the program execution

getExecutionEnvironment()

createLocalEnvironment()

createRemoteEnvironment(String host, int port, String... jarFiles)

final StreamExecutionEnvironment env = StreamExecutionEnvironment.getExecutionEnvironment();

DataStream<String> text = env.readTextFile("file:///path/to/file");

DataStream<String> input = ...;

DataStream<Integer> parsed = input.map(new MapFunction<String, Integer>() {
    @Override
    public Integer map(String value) {
        return Integer.parseInt(value);
    }
});

writeAsText(String path)

print()

getExecutionEnvironment()

createLocalEnvironment()

createRemoteEnvironment(host: String, port: Int, jarFiles: String*)

val env = StreamExecutionEnvironment.getExecutionEnvironment()

val text: DataStream[String] = env.readTextFile("file:///path/to/file")

val input: DataSet[String] = ...

val mapped = input.map { x => x.toInt }

writeAsText(path: String)

print()

Once you specified the complete program you need to trigger the program execution by calling execute() on the StreamExecutionEnvironment. Depending on the type of the ExecutionEnvironment the execution will be triggered on your local machine or submit your program for execution on a cluster.

The execute() method is returning a JobExecutionResult, this contains execution times and accumulator results.

Please see the Streaming Guide for information about streaming data sources and sink and for more in-depths information about the supported transformations on DataStream.

Check out the Batch Guide for information about batch data sources and sink and for more in-depths information about the supported transformations on DataSet.

Back to top

Lazy Evaluation

All Flink programs are executed lazily: When the program’s main method is executed, the data loading and transformations do not happen directly. Rather, each operation is created and added to the program’s plan. The operations are actually executed when the execution is explicitly triggered by an execute() call on the execution environment. Whether the program is executed locally or on a cluster depends on the type of execution environment

The lazy evaluation lets you construct sophisticated programs that Flink executes as one holistically planned unit.

Back to top

Specifying Keys

Some transformations (join, coGroup, keyBy, groupBy) require that a key be defined on a collection of elements. Other transformations (Reduce, GroupReduce, Aggregate, Windows) allow data being grouped on a key before they are applied.

A DataSet is grouped as

DataSet<...> input = // [...]
DataSet<...> reduced = input
  .groupBy(/*define key here*/)
  .reduceGroup(/*do something*/);

while a key can be specified on a DataStream using

DataStream<...> input = // [...]
DataStream<...> windowed = input
  .key(/*define key here*/)
  .window(/*window specification*/);

The data model of Flink is not based on key-value pairs. Therefore, you do not need to physically pack the data set types into keys and values. Keys are “virtual”: they are defined as functions over the actual data to guide the grouping operator.

NOTE: In the following discussion we will use the DataStream API and keyBy. For the DataSet API you just have to replace by DataSet and groupBy.

Define keys for Tuples

The simplest case is grouping Tuples on one or more fields of the Tuple:

DataStream<Tuple3<Integer,String,Long>> input = // [...]
KeyedStream<Tuple3<Integer,String,Long> keyed = input.keyBy(0)

val input: DataStream[(Int, String, Long)] = // [...]
val keyed = input.keyBy(0)

The tuples is grouped on the first field (the one of Integer type).

DataStream<Tuple3<Integer,String,Long>> input = // [...]
KeyedStream<Tuple3<Integer,String,Long> keyed = input.keyBy(0,1)

val input: DataSet[(Int, String, Long)] = // [...]
val grouped = input.groupBy(0,1)

Here, we group the tuples on a composite key consisting of the first and the second field.

A note on nested Tuples: If you have a DataStream with a nested tuple, such as:

DataStream<Tuple3<Tuple2<Integer, Float>,String,Long>> ds;

Specifying keyBy(0) will cause the system to use the full Tuple2 as a key (with the Integer and Float being the key). If you want to “navigate” into the nested Tuple2, you have to use field expression keys which are explained below.

Define keys using Field Expressions

You can use String-based field expressions to reference nested fields and define keys for grouping, sorting, joining, or coGrouping.

Field expressions make it very easy to select fields in (nested) composite types such as Tuple and POJO types.

// some ordinary POJO (Plain old Java Object)
public class WC {
  public String word;
  public int count;
}
DataStream<WC> words = // [...]
DataStream<WC> wordCounts = words.keyBy("word").window(/*window specification*/);

public static class WC {
  public ComplexNestedClass complex; //nested POJO
  private int count;
  // getter / setter for private field (count)
  public int getCount() {
    return count;
  }
  public void setCount(int c) {
    this.count = c;
  }
}
public static class ComplexNestedClass {
  public Integer someNumber;
  public float someFloat;
  public Tuple3<Long, Long, String> word;
  public IntWritable hadoopCitizen;
}

// some ordinary POJO (Plain old Java Object)
class WC(var word: String, var count: Int) {
  def this() { this("", 0L) }
}
val words: DataStream[WC] = // [...]
val wordCounts = words.keyBy("word").window(/*window specification*/)

// or, as a case class, which is less typing
case class WC(word: String, count: Int)
val words: DataStream[WC] = // [...]
val wordCounts = words.keyBy("word").reduce(/*window specification*/)

class WC(var complex: ComplexNestedClass, var count: Int) {
  def this() { this(null, 0) }
}

class ComplexNestedClass(
    var someNumber: Int,
    someFloat: Float,
    word: (Long, Long, String),
    hadoopCitizen: IntWritable) {
  def this() { this(0, 0, (0, 0, ""), new IntWritable(0)) }
}

Define keys using Key Selector Functions

An additional way to define keys are “key selector” functions. A key selector function takes a single element as input and returns the key for the element. The key can be of any type and be derived from arbitrary computations.

The following example shows a key selector function that simply returns the field of an object:

// some ordinary POJO
public class WC {public String word; public int count;}
DataStream<WC> words = // [...]
KeyedStream<WC> kyed = words
  .keyBy(new KeySelector<WC, String>() {
     public String getKey(WC wc) { return wc.word; }
   });

// some ordinary case class
case class WC(word: String, count: Int)
val words: DataStream[WC] = // [...]
val keyed = words.keyBy( _.word )

Back to top

Specifying Transformation Functions

Most transformations require user-defined functions. This section lists different ways of how they can be specified

class MyMapFunction implements MapFunction<String, Integer> {
  public Integer map(String value) { return Integer.parseInt(value); }
});
data.map(new MyMapFunction());

data.map(new MapFunction<String, Integer> () {
  public Integer map(String value) { return Integer.parseInt(value); }
});

data.filter(s -> s.startsWith("http://"));

data.reduce((i1,i2) -> i1 + i2);

class MyMapFunction implements MapFunction<String, Integer> {
  public Integer map(String value) { return Integer.parseInt(value); }
});

class MyMapFunction extends RichMapFunction<String, Integer> {
  public Integer map(String value) { return Integer.parseInt(value); }
});

data.map(new MyMapFunction());

data.map (new RichMapFunction<String, Integer>() {
  public Integer map(String value) { return Integer.parseInt(value); }
});

val data: DataSet[String] = // [...]
data.filter { _.startsWith("http://") }

val data: DataSet[Int] = // [...]
data.reduce { (i1,i2) => i1 + i2 }
// or
data.reduce { _ + _ }

data.map { x => x.toInt }

class MyMapFunction extends RichMapFunction[String, Int] {
  def map(in: String):Int = { in.toInt }
})

data.map(new MyMapFunction())

data.map (new RichMapFunction[String, Int] {
  def map(in: String):Int = { in.toInt }
})

Rich functions provide, in addition to the user-defined function (map, reduce, etc), four methods: open, close, getRuntimeContext, and setRuntimeContext. These are useful for parameterizing the function (see Passing Parameters to Functions), creating and finalizing local state, accessing broadcast variables (see Broadcast Variables, and for accessing runtime information such as accumulators and counters (see Accumulators and Counters, and information on iterations (see Iterations).

Back to top

Supported Data Types

Flink places some restrictions on the type of elements that can be in a DataSet or DataStream. The reason for this is that the system analyzes the types to determine efficient execution strategies.

There are six different categories of data types:

1. Java Tuples and Scala Case Classes
2. Java POJOs
3. Primitive Types
4. Regular Classes
5. Values
6. Hadoop Writables
7. Special Types

Tuples and Case Classes

DataStream<Tuple2<String, Integer>> wordCounts = env.fromElements(
    new Tuple2<String, Integer>("hello", 1),
    new Tuple2<String, Integer>("world", 2));

wordCounts.map(new MapFunction<Tuple2<String, Integer>, Integer>() {
    @Override
    public String map(Tuple2<String, Integer> value) throws Exception {
        return value.f1;
    }
});

wordCounts.keyBy(0); // also valid .keyBy("f0")

case class WordCount(word: String, count: Int)
val input = env.fromElements(
    WordCount("hello", 1),
    WordCount("world", 2)) // Case Class Data Set

input.keyBy("word")// key by field expression "word"

val input2 = env.fromElements(("hello", 1), ("world", 2)) // Tuple2 Data Set

input2.keyBy(0, 1) // key by field positions 0 and 1

POJOs

Java and Scala classes are treated by Flink as a special POJO data type if they fulfill the following requirements:

• The class must be public.

• It must have a public constructor without arguments (default constructor).

• All fields are either public or must be accessible through getter and setter functions. For a field called foo the getter and setter methods must be named getFoo() and setFoo().

• The type of a field must be supported by Flink. At the moment, Flink uses Avro to serialize arbitrary objects (such as Date).

Flink analyzes the structure of POJO types, i.e., it learns about the fields of a POJO. As a result POJO types are easier to use than general types. Moreover, Flink can process POJOs more efficiently than general types.

The following example shows a simple POJO with two public fields.

public class WordWithCount {

    public String word;
    public int count;

    public WordWithCount() {}

    public WordWithCount(String word, int count) {
        this.word = word;
        this.count = count;
    }
}

DataStream<WordWithCount> wordCounts = env.fromElements(
    new WordWithCount("hello", 1),
    new WordWithCount("world", 2));

wordCounts.keyBy("word"); // key by field expression "word"

class WordWithCount(var word: String, var count: Int) {
    def this() {
      this(null, -1)
    }
}

val input = env.fromElements(
    new WordWithCount("hello", 1),
    new WordWithCount("world", 2)) // Case Class Data Set

input.keyBy("word")// key by field expression "word"

Primitive Types

Flink supports all Java and Scala primitive types such as Integer, String, and Double.

General Class Types

Flink supports most Java and Scala classes (API and custom). Restrictions apply to classes containing fields that cannot be serialized, like file pointers, I/O streams, or other native resources. Classes that follow the Java Beans conventions work well in general.

All classes that are not identified as POJO types (see POJO requirements above) are handled by Flink as general class types. Flink treats these data types as black boxes and is not able to access their their content (i.e., for efficient sorting). General types are de/serialized using the serialization framework Kryo.

Values

Value types describe their serialization and deserialization manually. Instead of going through a general purpose serialization framework, they provide custom code for those operations by means of implementing the org.apache.flinktypes.Value interface with the methods read and write. Using a Value type is reasonable when general purpose serialization would be highly inefficient. An example would be a data type that implements a sparse vector of elements as an array. Knowing that the array is mostly zero, one can use a special encoding for the non-zero elements, while the general purpose serialization would simply write all array elements.

The org.apache.flinktypes.CopyableValue interface supports manual internal cloning logic in a similar way.

Flink comes with pre-defined Value types that correspond to basic data types. (ByteValue, ShortValue, IntValue, LongValue, FloatValue, DoubleValue, StringValue, CharValue, BooleanValue). These Value types act as mutable variants of the basic data types: Their value can be altered, allowing programmers to reuse objects and take pressure off the garbage collector.

Hadoop Writables

You can use types that implement the org.apache.hadoop.Writable interface. The serialization logic defined in the write()and readFields() methods will be used for serialization.

Special Types

You can use special types, including Scala’s Either, Option, and Try. The Java API has its own custom implementation of Either. Similarly to Scala’s Either, it represents a value of one two possible types, Left or Right. Either can be useful for error handling or operators that need to output two different types of records.

Type Erasure & Type Inference

Note: This Section is only relevant for Java.

The Java compiler throws away much of the generic type information after compilation. This is known as type erasure in Java. It means that at runtime, an instance of an object does not know its generic type any more. For example, instances of DataStream<String> and DataStream<Long> look the same to the JVM.

Flink requires type information at the time when it prepares the program for execution (when the main method of the program is called). The Flink Java API tries to reconstruct the type information that was thrown away in various ways and store it explicitly in the data sets and operators. You can retrieve the type via DataStream.getType(). The method returns an instance of TypeInformation, which is Flink’s internal way of representing types.

The type inference has its limits and needs the “cooperation” of the programmer in some cases. Examples for that are methods that create data sets from collections, such as ExecutionEnvironment.fromCollection(), where you can pass an argument that describes the type. But also generic functions like MapFunction<I, O> may need extra type information.

The ResultTypeQueryable interface can be implemented by input formats and functions to tell the API explicitly about their return type. The input types that the functions are invoked with can usually be inferred by the result types of the previous operations.

Execution Configuration

The StreamExecutionEnvironment also contains the ExecutionConfig which allows to set job specific configuration values for the runtime.

StreamExecutionEnvironment env = StreamExecutionEnvironment.getExecutionEnvironment();
ExecutionConfig executionConfig = env.getConfig();

val env = StreamExecutionEnvironment.getExecutionEnvironment
var executionConfig = env.getConfig

The following configuration options are available: (the default is bold)

• enableClosureCleaner() / disableClosureCleaner(). The closure cleaner is enabled by default. The closure cleaner removes unneeded references to the surrounding class of anonymous functions inside Flink programs. With the closure cleaner disabled, it might happen that an anonymous user function is referencing the surrounding class, which is usually not Serializable. This will lead to exceptions by the serializer.

• getParallelism() / setParallelism(int parallelism) Set the default parallelism for the job.

• getNumberOfExecutionRetries() / setNumberOfExecutionRetries(int numberOfExecutionRetries) Sets the number of times that failed tasks are re-executed. A value of zero effectively disables fault tolerance. A value of -1 indicates that the system default value (as defined in the configuration) should be used.

• getExecutionRetryDelay() / setExecutionRetryDelay(long executionRetryDelay) Sets the delay in milliseconds that the system waits after a job has failed, before re-executing it. The delay starts after all tasks have been successfully been stopped on the TaskManagers, and once the delay is past, the tasks are re-started. This parameter is useful to delay re-execution in order to let certain time-out related failures surface fully (like broken connections that have not fully timed out), before attempting a re-execution and immediately failing again due to the same problem. This parameter only has an effect if the number of execution re-tries is one or more.

• getExecutionMode() / setExecutionMode(). The default execution mode is PIPELINED. Sets the execution mode to execute the program. The execution mode defines whether data exchanges are performed in a batch or on a pipelined manner.

• enableForceKryo() / disableForceKryo. Kryo is not forced by default. Forces the GenericTypeInformation to use the Kryo serializer for POJOS even though we could analyze them as a POJO. In some cases this might be preferable. For example, when Flink’s internal serializers fail to handle a POJO properly.

• enableForceAvro() / disableForceAvro(). Avro is not forced by default. Forces the Flink AvroTypeInformation to use the Avro serializer instead of Kryo for serializing Avro POJOs.

• enableObjectReuse() / disableObjectReuse() By default, objects are not reused in Flink. Enabling the object reuse mode will instruct the runtime to reuse user objects for better performance. Keep in mind that this can lead to bugs when the user-code function of an operation is not aware of this behavior.

• enableSysoutLogging() / disableSysoutLogging() JobManager status updates are printed to System.out by default. This setting allows to disable this behavior.

• getGlobalJobParameters() / setGlobalJobParameters() This method allows users to set custom objects as a global configuration for the job. Since the ExecutionConfig is accessible in all user defined functions, this is an easy method for making configuration globally available in a job.

• addDefaultKryoSerializer(Class<?> type, Serializer<?> serializer) Register a Kryo serializer instance for the given type.

• addDefaultKryoSerializer(Class<?> type, Class<? extends Serializer<?>> serializerClass) Register a Kryo serializer class for the given type.

• registerTypeWithKryoSerializer(Class<?> type, Serializer<?> serializer) Register the given type with Kryo and specify a serializer for it. By registering a type with Kryo, the serialization of the type will be much more efficient.

• registerKryoType(Class<?> type) If the type ends up being serialized with Kryo, then it will be registered at Kryo to make sure that only tags (integer IDs) are written. If a type is not registered with Kryo, its entire class-name will be serialized with every instance, leading to much higher I/O costs.

• registerPojoType(Class<?> type) Registers the given type with the serialization stack. If the type is eventually serialized as a POJO, then the type is registered with the POJO serializer. If the type ends up being serialized with Kryo, then it will be registered at Kryo to make sure that only tags are written. If a type is not registered with Kryo, its entire class-name will be serialized with every instance, leading to much higher I/O costs.

Note that types registered with registerKryoType() are not available to Flink’s Kryo serializer instance.

• disableAutoTypeRegistration() Automatic type registration is enabled by default. The automatic type registration is registering all types (including sub-types) used by usercode with Kryo and the POJO serializer.

• setTaskCancellationInterval(long interval) Sets the the interval (in milliseconds) to wait between consecutive attempts to cancel a running task. When a task is canceled a new thread is created which periodically calls interrupt() on the task thread, if the task thread does not terminate within a certain time. This parameter refers to the time between consecutive calls to interrupt() and is set by default to 30000 milliseconds, or 30 seconds.

The RuntimeContext which is accessible in Rich* functions through the getRuntimeContext() method also allows to access the ExecutionConfig in all user defined functions.

Back to top

Program Packaging and Distributed Execution

As described earlier, Flink programs can be executed on clusters by using a remote environment. Alternatively, programs can be packaged into JAR Files (Java Archives) for execution. Packaging the program is a prerequisite to executing them through the command line interface.

Packaging Programs

To support execution from a packaged JAR file via the command line or web interface, a program must use the environment obtained by StreamExecutionEnvironment.getExecutionEnvironment(). This environment will act as the cluster’s environment when the JAR is submitted to the command line or web interface. If the Flink program is invoked differently than through these interfaces, the environment will act like a local environment.

To package the program, simply export all involved classes as a JAR file. The JAR file’s manifest must point to the class that contains the program’s entry point (the class with the public main method). The simplest way to do this is by putting the main-class entry into the manifest (such as main-class: org.apache.flinkexample.MyProgram). The main-class attribute is the same one that is used by the Java Virtual Machine to find the main method when executing a JAR files through the command java -jar pathToTheJarFile. Most IDEs offer to include that attribute automatically when exporting JAR files.

Packaging Programs through Plans

Additionally, we support packaging programs as Plans. Instead of defining a progam in the main method and calling execute() on the environment, plan packaging returns the Program Plan, which is a description of the program’s data flow. To do that, the program must implement the org.apache.flink.api.common.Program interface, defining the getPlan(String...) method. The strings passed to that method are the command line arguments. The program’s plan can be created from the environment via the ExecutionEnvironment#createProgramPlan() method. When packaging the program’s plan, the JAR manifest must point to the class implementing the org.apache.flinkapi.common.Program interface, instead of the class with the main method.

Summary

The overall procedure to invoke a packaged program is as follows:

1. The JAR’s manifest is searched for a main-class or program-class attribute. If both attributes are found, the program-class attribute takes precedence over the main-class attribute. Both the command line and the web interface support a parameter to pass the entry point class name manually for cases where the JAR manifest contains neither attribute.

2. If the entry point class implements the org.apache.flinkapi.common.Program, then the system calls the getPlan(String...) method to obtain the program plan to execute.

3. If the entry point class does not implement the org.apache.flinkapi.common.Program interface, the system will invoke the main method of the class.

Back to top

Accumulators & Counters

Accumulators are simple constructs with an add operation and a final accumulated result, which is available after the job ended.

The most straightforward accumulator is a counter: You can increment it using the Accumulator.add(V value) method. At the end of the job Flink will sum up (merge) all partial results and send the result to the client. Accumulators are useful during debugging or if you quickly want to find out more about your data.

Flink currently has the following built-in accumulators. Each of them implements the Accumulator interface.

• IntCounter, LongCounter and DoubleCounter: See below for an example using a counter.
• Histogram: A histogram implementation for a discrete number of bins. Internally it is just a map from Integer to Integer. You can use this to compute distributions of values, e.g. the distribution of words-per-line for a word count program.

How to use accumulators:

First you have to create an accumulator object (here a counter) in the user-defined transformation function where you want to use it.

private IntCounter numLines = new IntCounter();

Second you have to register the accumulator object, typically in the open() method of the rich function. Here you also define the name.

getRuntimeContext().addAccumulator("num-lines", this.numLines);

You can now use the accumulator anywhere in the operator function, including in the open() and close() methods.

this.numLines.add(1);

The overall result will be stored in the JobExecutionResult object which is returned from the execute() method of the execution environment (currently this only works if the execution waits for the completion of the job).

myJobExecutionResult.getAccumulatorResult("num-lines")

All accumulators share a single namespace per job. Thus you can use the same accumulator in different operator functions of your job. Flink will internally merge all accumulators with the same name.

A note on accumulators and iterations: Currently the result of accumulators is only available after the overall job ended. We plan to also make the result of the previous iteration available in the next iteration. You can use Aggregators to compute per-iteration statistics and base the termination of iterations on such statistics.

Custom accumulators:

To implement your own accumulator you simply have to write your implementation of the Accumulator interface. Feel free to create a pull request if you think your custom accumulator should be shipped with Flink.

You have the choice to implement either Accumulator or SimpleAccumulator.

Accumulator<V,R> is most flexible: It defines a type V for the value to add, and a result type R for the final result. E.g. for a histogram, V is a number and R is a histogram. SimpleAccumulator is for the cases where both types are the same, e.g. for counters.

Back to top

Parallel Execution

This section describes how the parallel execution of programs can be configured in Flink. A Flink program consists of multiple tasks (transformations/operators, data sources, and sinks). A task is split into several parallel instances for execution and each parallel instance processes a subset of the task’s input data. The number of parallel instances of a task is called its parallelism.

The parallelism of a task can be specified in Flink on different levels.

Operator Level

The parallelism of an individual operator, data source, or data sink can be defined by calling its setParallelism() method. For example, like this:

final StreamExecutionEnvironment env = StreamExecutionEnvironment.getExecutionEnvironment();

DataStream<String> text = [...]
DataStream<Tuple2<String, Integer>> wordCounts = text
    .flatMap(new LineSplitter())
    .keyBy(0)
    .timeWindow(Time.seconds(5))
    .sum(1).setParallelism(5);

wordCounts.print();

env.execute("Word Count Example");

val env = StreamExecutionEnvironment.getExecutionEnvironment

val text = [...]
val wordCounts = text
    .flatMap{ _.split(" ") map { (_, 1) } }
    .keyBy(0)
    .timeWindow(Time.seconds(5))
    .sum(1).setParallelism(5)
wordCounts.print()

env.execute("Word Count Example")

Execution Environment Level

As mentioned here Flink programs are executed in the context of an execution environment. An execution environment defines a default parallelism for all operators, data sources, and data sinks it executes. Execution environment parallelism can be overwritten by explicitly configuring the parallelism of an operator.

The default parallelism of an execution environment can be specified by calling the setParallelism() method. To execute all operators, data sources, and data sinks with a parallelism of 3, set the default parallelism of the execution environment as follows:

final StreamExecutionEnvironment env = StreamExecutionEnvironment.getExecutionEnvironment();
env.setParallelism(3);

DataStream<String> text = [...]
DataStream<Tuple2<String, Integer>> wordCounts = [...]
wordCounts.print();

env.execute("Word Count Example");

val env = StreamExecutionEnvironment.getExecutionEnvironment
env.setParallelism(3)

val text = [...]
val wordCounts = text
    .flatMap{ _.split(" ") map { (_, 1) } }
    .keyBy(0)
    .timeWindow(Time.seconds(5))
    .sum(1)
wordCounts.print()

env.execute("Word Count Example")

Client Level

The parallelism can be set at the Client when submitting jobs to Flink. The Client can either be a Java or a Scala program. One example of such a Client is Flink’s Command-line Interface (CLI).

For the CLI client, the parallelism parameter can be specified with -p. For example:

./bin/flink run -p 10 ../examples/*WordCount-java*.jar

In a Java/Scala program, the parallelism is set as follows:

try {
    PackagedProgram program = new PackagedProgram(file, args);
    InetSocketAddress jobManagerAddress = RemoteExecutor.getInetFromHostport("localhost:6123");
    Configuration config = new Configuration();

    Client client = new Client(jobManagerAddress, config, program.getUserCodeClassLoader());

    // set the parallelism to 10 here
    client.run(program, 10, true);

} catch (ProgramInvocationException e) {
    e.printStackTrace();
}

try {
    PackagedProgram program = new PackagedProgram(file, args)
    InetSocketAddress jobManagerAddress = RemoteExecutor.getInetFromHostport("localhost:6123")
    Configuration config = new Configuration()

    Client client = new Client(jobManagerAddress, new Configuration(), program.getUserCodeClassLoader())

    // set the parallelism to 10 here
    client.run(program, 10, true)

} catch {
    case e: Exception => e.printStackTrace
}

System Level

A system-wide default parallelism for all execution environments can be defined by setting the parallelism.default property in ./conf/flink-conf.yaml. See the Configuration documentation for details.

Back to top

Execution Plans

Depending on various parameters such as data size or number of machines in the cluster, Flink’s optimizer automatically chooses an execution strategy for your program. In many cases, it can be useful to know how exactly Flink will execute your program.

Plan Visualization Tool

Flink comes packaged with a visualization tool for execution plans. The HTML document containing the visualizer is located under tools/planVisualizer.html. It takes a JSON representation of the job execution plan and visualizes it as a graph with complete annotations of execution strategies.

The following code shows how to print the execution plan JSON from your program:

final ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();

...

System.out.println(env.getExecutionPlan());

val env = ExecutionEnvironment.getExecutionEnvironment

...

println(env.getExecutionPlan())

To visualize the execution plan, do the following:

1. Open planVisualizer.html with your web browser,
2. Paste the JSON string into the text field, and
3. Press the draw button.

After these steps, a detailed execution plan will be visualized.

Web Interface

Flink offers a web interface for submitting and executing jobs. The interface is part of the JobManager’s web interface for monitoring, per default running on port 8081. Job submission via this interfaces requires that you have set jobmanager.web.submit.enable: true in flink-conf.yaml.

You may specify program arguments before the job is executed. The plan visualization enables you to show the execution plan before executing the Flink job.

Back to top