Flink DataSet API Programming Guide

DataSet programs in Flink are regular programs that implement transformations on data sets (e.g., filtering, mapping, joining, grouping). The data sets are initially created from certain sources (e.g., by reading files, 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.

Please refer to the DataStream API overview for an introduction to the basic concepts of the Flink API. That overview is for the DataStream API but the basic concepts of the two APIs are the same.

In order to create your own Flink DataSet program, we encourage you to start with the anatomy of a Flink Program and gradually add your own transformations. The remaining sections act as references for additional operations and advanced features.

Note: Starting with Flink 1.12 the DataSet has been soft deprecated. We recommend that you use the DataStream API with BATCH execution mode. The linked section also outlines cases where it makes sense to use the DataSet API but those cases will become rarer as development progresses and the DataSet API will eventually be removed. Please also see FLIP-131 for background information on this decision.

• Example Program
• DataSet Transformations
• Specifying Keys
• Data Sources
  • Read Compressed Files
• Data Sinks
• Iteration Operators
• Operating on data objects in functions
  • Object-Reuse Disabled (DEFAULT)
  • Object-Reuse Enabled
• Debugging
  • Local Execution Environment
  • Collection Data Sources and Sinks
• Semantic Annotations
• Broadcast Variables
• Distributed Cache
• Passing Parameters to Functions

Example Program

The following program is a complete, working example of WordCount. You can copy & paste the code to run it locally. You only have to include the correct Flink’s library into your project (see Section Linking with Flink) and specify the imports. Then you are ready to go!

public class WordCountExample {
    public static void main(String[] args) throws Exception {
        final ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
        DataSet<String> text = env.fromElements(
            "Who's there?",
            "I think I hear them. Stand, ho! Who's there?");
        DataSet<Tuple2<String, Integer>> wordCounts = text
            .flatMap(new LineSplitter())
            .groupBy(0)
            .sum(1);
        wordCounts.print();
    }
    public static class LineSplitter implements FlatMapFunction<String, Tuple2<String, Integer>> {
        @Override
        public void flatMap(String line, Collector<Tuple2<String, Integer>> out) {
            for (String word : line.split(" ")) {
                out.collect(new Tuple2<String, Integer>(word, 1));
            }
        }
    }
}

import org.apache.flink.api.scala._
object WordCount {
  def main(args: Array[String]) {
    val env = ExecutionEnvironment.getExecutionEnvironment
    val text = env.fromElements(
      "Who's there?",
      "I think I hear them. Stand, ho! Who's there?")
    val counts = text.flatMap { _.toLowerCase.split("\\W+") filter { _.nonEmpty } }
      .map { (_, 1) }
      .groupBy(0)
      .sum(1)
    counts.print()
  }
}

DataSet Transformations

Data transformations transform one or more DataSets into a new DataSet. Programs can combine multiple transformations into sophisticated assemblies.

This section gives a brief overview of the available transformations. The transformations documentation has a full description of all transformations with examples.

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

data.flatMap(new FlatMapFunction<String, String>() {
  public void flatMap(String value, Collector<String> out) {
    for (String s : value.split(“ “)) {
      out.collect(s);
    }
  }
});

data.mapPartition(new MapPartitionFunction<String, Long>() {
  public void mapPartition(Iterable<String> values, Collector<Long> out) {
    long c = 0;
    for (String s : values) {
      c++;
    }
    out.collect(c);
  }
});

data.filter(new FilterFunction<Integer>() {
  public boolean filter(Integer value) { return value > 1000; }
});

data.reduce(new ReduceFunction<Integer> {
  public Integer reduce(Integer a, Integer b) { return a + b; }
});

data.reduceGroup(new GroupReduceFunction<Integer, Integer> {
  public void reduce(Iterable<Integer> values, Collector<Integer> out) {
    int prefixSum = 0;
    for (Integer i : values) {
      prefixSum += i;
      out.collect(prefixSum);
    }
  }
});

Dataset<Tuple3<Integer, String, Double>> input = // […]
DataSet<Tuple3<Integer, String, Double>> output = input.aggregate(SUM, 0).and(MIN, 2);

Dataset<Tuple3<Integer, String, Double>> input = // […]
DataSet<Tuple3<Integer, String, Double>> output = input.sum(0).andMin(2);

data.distinct();

result = input1.join(input2)
               .where(0)       // key of the first input (tuple field 0)
               .equalTo(1);    // key of the second input (tuple field 1)

// This executes a join by broadcasting the first data set
// using a hash table for the broadcast data
result = input1.join(input2, JoinHint.BROADCAST_HASH_FIRST)
               .where(0).equalTo(1);

input1.leftOuterJoin(input2) // rightOuterJoin or fullOuterJoin for right or full outer joins
      .where(0)              // key of the first input (tuple field 0)
      .equalTo(1)            // key of the second input (tuple field 1)
      .with(new JoinFunction<String, String, String>() {
          public String join(String v1, String v2) {
             // NOTE:
             // - v2 might be null for leftOuterJoin
             // - v1 might be null for rightOuterJoin
             // - v1 OR v2 might be null for fullOuterJoin
          }
      });

data1.coGroup(data2)
     .where(0)
     .equalTo(1)
     .with(new CoGroupFunction<String, String, String>() {
         public void coGroup(Iterable<String> in1, Iterable<String> in2, Collector<String> out) {
           out.collect(…);
         }
      });

DataSet<Integer> data1 = // […]
DataSet<String> data2 = // […]
DataSet<Tuple2<Integer, String>> result = data1.cross(data2);

DataSet<String> data1 = // […]
DataSet<String> data2 = // […]
DataSet<String> result = data1.union(data2);

DataSet<String> in = // […]
DataSet<String> result = in.rebalance()
                           .map(new Mapper());

DataSet<Tuple2<String,Integer>> in = // […]
DataSet<Integer> result = in.partitionByHash(0)
                            .mapPartition(new PartitionMapper());

DataSet<Tuple2<String,Integer>> in = // […]
DataSet<Integer> result = in.partitionByRange(0)
                            .mapPartition(new PartitionMapper());

DataSet<Tuple2<String,Integer>> in = // […]
DataSet<Integer> result = in.partitionCustom(partitioner, key)
                            .mapPartition(new PartitionMapper());

DataSet<Tuple2<String,Integer>> in = // […]
DataSet<Integer> result = in.sortPartition(1, Order.ASCENDING)
                            .mapPartition(new PartitionMapper());

DataSet<Tuple2<String,Integer>> in = // […]
// regular data set
DataSet<Tuple2<String,Integer>> result1 = in.first(3);
// grouped data set
DataSet<Tuple2<String,Integer>> result2 = in.groupBy(0)
                                            .first(3);
// grouped-sorted data set
DataSet<Tuple2<String,Integer>> result3 = in.groupBy(0)
                                            .sortGroup(1, Order.ASCENDING)
                                            .first(3);

The following transformations are available on data sets of Tuples:

DataSet<Tuple3<Integer, Double, String>> in = // […]
DataSet<Tuple2<String, Integer>> out = in.project(2,0);

DataSet<Tuple3<Integer, Double, String>> in = // […]
// a DataSet with a single tuple with minimum values for the Integer and String fields.
DataSet<Tuple3<Integer, Double, String>> out = in.minBy(0, 2);
// a DataSet with one tuple for each group with the minimum value for the Double field.
DataSet<Tuple3<Integer, Double, String>> out2 = in.groupBy(2)
                                                  .minBy(1);

data.map { x => x.toInt }

data.flatMap { str => str.split(“ “) }

data.mapPartition { in => in map { (, 1) } }

data.filter {  > 1000 }

data.reduce {  +  }

data.reduceGroup { elements => elements.sum }

val input: DataSet[(Int, String, Double)] = // […]
val output: DataSet[(Int, String, Double)] = input.aggregate(SUM, 0).aggregate(MIN, 2)

val input: DataSet[(Int, String, Double)] = // […]
val output: DataSet[(Int, String, Double)] = input.sum(0).min(2)

data.distinct()

// In this case tuple fields are used as keys. “0” is the join field on the first tuple
// “1” is the join field on the second tuple.
val result = input1.join(input2).where(0).equalTo(1)

// This executes a join by broadcasting the first data set
// using a hash table for the broadcast data
val result = input1.join(input2, JoinHint.BROADCAST_HASH_FIRST)
                   .where(0).equalTo(1)

val joined = left.leftOuterJoin(right).where(0).equalTo(1) {
   (left, right) =>
     val a = if (left == null) “none” else left._1
     (a, right)
  }

data1.coGroup(data2).where(0).equalTo(1)

val data1: DataSet[Int] = // […]
val data2: DataSet[String] = // […]
val result: DataSet[(Int, String)] = data1.cross(data2)

data.union(data2)

val data1: DataSet[Int] = // […]
val result: DataSet[(Int, String)] = data1.rebalance().map(…)

val in: DataSet[(Int, String)] = // […]
val result = in.partitionByHash(0).mapPartition { … }

val in: DataSet[(Int, String)] = // […]
val result = in.partitionByRange(0).mapPartition { … }

val in: DataSet[(Int, String)] = // […]
val result = in
  .partitionCustom(partitioner, key).mapPartition { … }

val in: DataSet[(Int, String)] = // […]
val result = in.sortPartition(1, Order.ASCENDING).mapPartition { … }

val in: DataSet[(Int, String)] = // […]
// regular data set
val result1 = in.first(3)
// grouped data set
val result2 = in.groupBy(0).first(3)
// grouped-sorted data set
val result3 = in.groupBy(0).sortGroup(1, Order.ASCENDING).first(3)

The following transformations are available on data sets of Tuples:

val in: DataSet[(Int, Double, String)] = // […]
// a data set with a single tuple with minimum values for the Int and String fields.
val out: DataSet[(Int, Double, String)] = in.minBy(0, 2)
// a data set with one tuple for each group with the minimum value for the Double field.
val out2: DataSet[(Int, Double, String)] = in.groupBy(2)
                                             .minBy(1)

Extraction from tuples, case classes and collections via anonymous pattern matching, like the following:

val data: DataSet[(Int, String, Double)] = // [...]
data.map {
  case (id, name, temperature) => // [...]
}

is not supported by the API out-of-the-box. To use this feature, you should use a Scala API extension.

The parallelism of a transformation can be defined by setParallelism(int) while name(String) assigns a custom name to a transformation which is helpful for debugging. The same is possible for Data Sources and Data Sinks.

withParameters(Configuration) passes Configuration objects, which can be accessed from the open() method inside the user function.

Specifying Keys

Some transformations (join, coGroup, groupBy) require that a key be defined on a collection of elements. Other transformations (Reduce, GroupReduce, Aggregate) 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*/);

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.

Define keys for Tuples

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

DataSet<Tuple3<Integer,String,Long>> input = // [...]
UnsortedGrouping<Tuple3<Integer,String,Long>,Tuple> keyed = input.groupBy(0)

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

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

DataSet<Tuple3<Integer,String,Long>> input = // [...]
UnsortedGrouping<Tuple3<Integer,String,Long>,Tuple> keyed = input.groupBy(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 DataSet with a nested tuple, such as:

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

Specifying groupBy(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.

In the example below, we have a WC POJO with two fields “word” and “count”. To group by the field word, we just pass its name to the groupBy() function.

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

Field Expression Syntax:

• Select POJO fields by their field name. For example "user" refers to the “user” field of a POJO type.

• Select Tuple fields by their field name or 0-offset field index. For example "f0" and "5" refer to the first and sixth field of a Java Tuple type, respectively.

• You can select nested fields in POJOs and Tuples. For example "user.zip" refers to the “zip” field of a POJO which is stored in the “user” field of a POJO type. Arbitrary nesting and mixing of POJOs and Tuples is supported such as "f1.user.zip" or "user.f3.1.zip".

• You can select the full type using the "*" wildcard expressions. This does also work for types which are not Tuple or POJO types.

Field Expression Example:

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;
}

These are valid field expressions for the example code above:

• "count": The count field in the WC class.

• "complex": Recursively selects all fields of the field complex of POJO type ComplexNestedClass.

• "complex.word.f2": Selects the last field of the nested Tuple3.

• "complex.hadoopCitizen": Selects the Hadoop IntWritable type.

In the example below, we have a WC POJO with two fields “word” and “count”. To group by the field word, we just pass its name to the groupBy() function.

// some ordinary POJO (Plain old Java Object)
class WC(var word: String, var count: Int) {
  def this() { this("", 0L) }
}
val words: DataSet[WC] = // [...]
val wordCounts = words.groupBy("word")
// or, as a case class, which is less typing
case class WC(word: String, count: Int)
val words: DataSet[WC] = // [...]
val wordCounts = words.groupBy("word")

Field Expression Syntax:

• Select POJO fields by their field name. For example "user" refers to the “user” field of a POJO type.

• Select Tuple fields by their 1-offset field name or 0-offset field index. For example "_1" and "5" refer to the first and sixth field of a Scala Tuple type, respectively.

• You can select nested fields in POJOs and Tuples. For example "user.zip" refers to the “zip” field of a POJO which is stored in the “user” field of a POJO type. Arbitrary nesting and mixing of POJOs and Tuples is supported such as "_2.user.zip" or "user._4.1.zip".

• You can select the full type using the "_" wildcard expressions. This does also work for types which are not Tuple or POJO types.

Field Expression Example:

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)) }
}

These are valid field expressions for the example code above:

• "count": The count field in the WC class.

• "complex": Recursively selects all fields of the field complex of POJO type ComplexNestedClass.

• "complex.word._3": Selects the last field of the nested Tuple3.

• "complex.hadoopCitizen": Selects the Hadoop IntWritable type.

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 deterministic 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;}
DataSet<WC> words = // [...]
UnsortedGrouping<WC> keyed = words
  .groupBy(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: DataSet[WC] = // [...]
val keyed = words.groupBy( _.word )

Data Sources

Data sources create the initial data sets, such as from files or from Java collections. The general mechanism of creating data sets is abstracted behind an InputFormat. Flink comes with several built-in formats to create data sets from common file formats. Many of them have shortcut methods on the ExecutionEnvironment.

File-based:

• readTextFile(path) / TextInputFormat - Reads files line wise and returns them as Strings.

• readTextFileWithValue(path) / TextValueInputFormat - Reads files line wise and returns them as StringValues. StringValues are mutable strings.

• readCsvFile(path) / CsvInputFormat - Parses files of comma (or another char) delimited fields. Returns a DataSet of tuples or POJOs. Supports the basic java types and their Value counterparts as field types.

• readFileOfPrimitives(path, Class) / PrimitiveInputFormat - Parses files of new-line (or another char sequence) delimited primitive data types such as String or Integer.

• readFileOfPrimitives(path, delimiter, Class) / PrimitiveInputFormat - Parses files of new-line (or another char sequence) delimited primitive data types such as String or Integer using the given delimiter.

Collection-based:

• fromCollection(Collection) - Creates a data set from a Java.util.Collection. All elements in the collection must be of the same type.

• fromCollection(Iterator, Class) - Creates a data set from an iterator. The class specifies the data type of the elements returned by the iterator.

• fromElements(T ...) - Creates a data set from the given sequence of objects. All objects must be of the same type.

• fromParallelCollection(SplittableIterator, Class) - Creates a data set from an iterator, in parallel. The class specifies the data type of the elements returned by the iterator.

• generateSequence(from, to) - Generates the sequence of numbers in the given interval, in parallel.

Generic:

• readFile(inputFormat, path) / FileInputFormat - Accepts a file input format.

• createInput(inputFormat) / InputFormat - Accepts a generic input format.

Examples

ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
// read text file from local files system
DataSet<String> localLines = env.readTextFile("file:///path/to/my/textfile");
// read text file from an HDFS running at nnHost:nnPort
DataSet<String> hdfsLines = env.readTextFile("hdfs://nnHost:nnPort/path/to/my/textfile");
// read a CSV file with three fields
DataSet<Tuple3<Integer, String, Double>> csvInput = env.readCsvFile("hdfs:///the/CSV/file")
                           .types(Integer.class, String.class, Double.class);
// read a CSV file with five fields, taking only two of them
DataSet<Tuple2<String, Double>> csvInput = env.readCsvFile("hdfs:///the/CSV/file")
                               .includeFields("10010")  // take the first and the fourth field
                           .types(String.class, Double.class);
// read a CSV file with three fields into a POJO (Person.class) with corresponding fields
DataSet<Person>> csvInput = env.readCsvFile("hdfs:///the/CSV/file")
                         .pojoType(Person.class, "name", "age", "zipcode");
// read a file from the specified path of type SequenceFileInputFormat
DataSet<Tuple2<IntWritable, Text>> tuples =
 env.createInput(HadoopInputs.readSequenceFile(IntWritable.class, Text.class, "hdfs://nnHost:nnPort/path/to/file"));
// creates a set from some given elements
DataSet<String> value = env.fromElements("Foo", "bar", "foobar", "fubar");
// generate a number sequence
DataSet<Long> numbers = env.generateSequence(1, 10000000);
// Read data from a relational database using the JDBC input format
DataSet<Tuple2<String, Integer> dbData =
    env.createInput(
      JdbcInputFormat.buildJdbcInputFormat()
                     .setDrivername("org.apache.derby.jdbc.EmbeddedDriver")
                     .setDBUrl("jdbc:derby:memory:persons")
                     .setQuery("select name, age from persons")
                     .setRowTypeInfo(new RowTypeInfo(BasicTypeInfo.STRING_TYPE_INFO, BasicTypeInfo.INT_TYPE_INFO))
                     .finish()
    );
// Note: Flink's program compiler needs to infer the data types of the data items which are returned
// by an InputFormat. If this information cannot be automatically inferred, it is necessary to
// manually provide the type information as shown in the examples above.

Configuring CSV Parsing

Flink offers a number of configuration options for CSV parsing:

• types(Class ... types) specifies the types of the fields to parse. It is mandatory to configure the types of the parsed fields. In case of the type class Boolean.class, “True” (case-insensitive), “False” (case-insensitive), “1” and “0” are treated as booleans.

• lineDelimiter(String del) specifies the delimiter of individual records. The default line delimiter is the new-line character '\n'.

• fieldDelimiter(String del) specifies the delimiter that separates fields of a record. The default field delimiter is the comma character ','.

• includeFields(boolean ... flag), includeFields(String mask), or includeFields(long bitMask) defines which fields to read from the input file (and which to ignore). By default the first n fields (as defined by the number of types in the types() call) are parsed.

• parseQuotedStrings(char quoteChar) enables quoted string parsing. Strings are parsed as quoted strings if the first character of the string field is the quote character (leading or tailing whitespaces are not trimmed). Field delimiters within quoted strings are ignored. Quoted string parsing fails if the last character of a quoted string field is not the quote character or if the quote character appears at some point which is not the start or the end of the quoted string field (unless the quote character is escaped using ‘’). If quoted string parsing is enabled and the first character of the field is not the quoting string, the string is parsed as unquoted string. By default, quoted string parsing is disabled.

• ignoreComments(String commentPrefix) specifies a comment prefix. All lines that start with the specified comment prefix are not parsed and ignored. By default, no lines are ignored.

• ignoreInvalidLines() enables lenient parsing, i.e., lines that cannot be correctly parsed are ignored. By default, lenient parsing is disabled and invalid lines raise an exception.

• ignoreFirstLine() configures the InputFormat to ignore the first line of the input file. By default no line is ignored.

Recursive Traversal of the Input Path Directory

For file-based inputs, when the input path is a directory, nested files are not enumerated by default. Instead, only the files inside the base directory are read, while nested files are ignored. Recursive enumeration of nested files can be enabled through the recursive.file.enumeration configuration parameter, like in the following example.

// enable recursive enumeration of nested input files
ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
// create a configuration object
Configuration parameters = new Configuration();
// set the recursive enumeration parameter
parameters.setBoolean("recursive.file.enumeration", true);
// pass the configuration to the data source
DataSet<String> logs = env.readTextFile("file:///path/with.nested/files")
              .withParameters(parameters);

Data sources create the initial data sets, such as from files or from Java collections. The general mechanism of creating data sets is abstracted behind an InputFormat. Flink comes with several built-in formats to create data sets from common file formats. Many of them have shortcut methods on the ExecutionEnvironment.

File-based:

• readTextFile(path) / TextInputFormat - Reads files line wise and returns them as Strings.

• readTextFileWithValue(path) / TextValueInputFormat - Reads files line wise and returns them as StringValues. StringValues are mutable strings.

• readCsvFile(path) / CsvInputFormat - Parses files of comma (or another char) delimited fields. Returns a DataSet of tuples, case class objects, or POJOs. Supports the basic java types and their Value counterparts as field types.

• readFileOfPrimitives(path, delimiter) / PrimitiveInputFormat - Parses files of new-line (or another char sequence) delimited primitive data types such as String or Integer using the given delimiter.

• readSequenceFile(Key, Value, path) / SequenceFileInputFormat - Creates a JobConf and reads file from the specified path with type SequenceFileInputFormat, Key class and Value class and returns them as Tuple2<Key, Value>.

Collection-based:

• fromCollection(Iterable) - Creates a data set from an Iterable. All elements returned by the Iterable must be of the same type.

• fromCollection(Iterator) - Creates a data set from an Iterator. The class specifies the data type of the elements returned by the iterator.

• fromElements(elements: _*) - Creates a data set from the given sequence of objects. All objects must be of the same type.

• fromParallelCollection(SplittableIterator) - Creates a data set from an iterator, in parallel. The class specifies the data type of the elements returned by the iterator.

• generateSequence(from, to) - Generates the sequence of numbers in the given interval, in parallel.

Generic:

• readFile(inputFormat, path) / FileInputFormat - Accepts a file input format.

• createInput(inputFormat) / InputFormat - Accepts a generic input format.

Examples

val env  = ExecutionEnvironment.getExecutionEnvironment
// read text file from local files system
val localLines = env.readTextFile("file:///path/to/my/textfile")
// read text file from an HDFS running at nnHost:nnPort
val hdfsLines = env.readTextFile("hdfs://nnHost:nnPort/path/to/my/textfile")
// read a CSV file with three fields
val csvInput = env.readCsvFile[(Int, String, Double)]("hdfs:///the/CSV/file")
// read a CSV file with five fields, taking only two of them
val csvInput = env.readCsvFile[(String, Double)](
  "hdfs:///the/CSV/file",
  includedFields = Array(0, 3)) // take the first and the fourth field
// CSV input can also be used with Case Classes
case class MyCaseClass(str: String, dbl: Double)
val csvInput = env.readCsvFile[MyCaseClass](
  "hdfs:///the/CSV/file",
  includedFields = Array(0, 3)) // take the first and the fourth field
// read a CSV file with three fields into a POJO (Person) with corresponding fields
val csvInput = env.readCsvFile[Person](
  "hdfs:///the/CSV/file",
  pojoFields = Array("name", "age", "zipcode"))
// create a set from some given elements
val values = env.fromElements("Foo", "bar", "foobar", "fubar")
// generate a number sequence
val numbers = env.generateSequence(1, 10000000)
// read a file from the specified path of type SequenceFileInputFormat
val tuples = env.createInput(HadoopInputs.readSequenceFile(classOf[IntWritable], classOf[Text],
 "hdfs://nnHost:nnPort/path/to/file"))

Configuring CSV Parsing

Flink offers a number of configuration options for CSV parsing:

• lineDelimiter: String specifies the delimiter of individual records. The default line delimiter is the new-line character '\n'.

• fieldDelimiter: String specifies the delimiter that separates fields of a record. The default field delimiter is the comma character ','.

• includeFields: Array[Int] defines which fields to read from the input file (and which to ignore). By default the first n fields (as defined by the number of types in the types() call) are parsed.

• pojoFields: Array[String] specifies the fields of a POJO that are mapped to CSV fields. Parsers for CSV fields are automatically initialized based on the type and order of the POJO fields.

• parseQuotedStrings: Character enables quoted string parsing. Strings are parsed as quoted strings if the first character of the string field is the quote character (leading or tailing whitespaces are not trimmed). Field delimiters within quoted strings are ignored. Quoted string parsing fails if the last character of a quoted string field is not the quote character. If quoted string parsing is enabled and the first character of the field is not the quoting string, the string is parsed as unquoted string. By default, quoted string parsing is disabled.

• ignoreComments: String specifies a comment prefix. All lines that start with the specified comment prefix are not parsed and ignored. By default, no lines are ignored.

• lenient: Boolean enables lenient parsing, i.e., lines that cannot be correctly parsed are ignored. By default, lenient parsing is disabled and invalid lines raise an exception.

• ignoreFirstLine: Boolean configures the InputFormat to ignore the first line of the input file. By default no line is ignored.

Recursive Traversal of the Input Path Directory

For file-based inputs, when the input path is a directory, nested files are not enumerated by default. Instead, only the files inside the base directory are read, while nested files are ignored. Recursive enumeration of nested files can be enabled through the recursive.file.enumeration configuration parameter, like in the following example.

// enable recursive enumeration of nested input files
val env  = ExecutionEnvironment.getExecutionEnvironment
// create a configuration object
val parameters = new Configuration
// set the recursive enumeration parameter
parameters.setBoolean("recursive.file.enumeration", true)
// pass the configuration to the data source
env.readTextFile("file:///path/with.nested/files").withParameters(parameters)

Read Compressed Files

Flink currently supports transparent decompression of input files if these are marked with an appropriate file extension. In particular, this means that no further configuration of the input formats is necessary and any FileInputFormat support the compression, including custom input formats. Please notice that compressed files might not be read in parallel, thus impacting job scalability.

The following table lists the currently supported compression methods.

Data Sinks

Data sinks consume DataSets and are used to store or return them. Data sink operations are described using an OutputFormat. Flink comes with a variety of built-in output formats that are encapsulated behind operations on the DataSet:

• writeAsText() / TextOutputFormat - Writes elements line-wise as Strings. The Strings are obtained by calling the toString() method of each element.
• writeAsFormattedText() / TextOutputFormat - Write elements line-wise as Strings. The Strings are obtained by calling a user-defined format() method for each element.
• writeAsCsv(...) / CsvOutputFormat - Writes tuples as comma-separated value files. Row and field delimiters are configurable. The value for each field comes from the toString() method of the objects.
• print() / printToErr() / print(String msg) / printToErr(String msg) - Prints the toString() value of each element on the standard out / standard error stream. Optionally, a prefix (msg) can be provided which is prepended to the output. This can help to distinguish between different calls to print. If the parallelism is greater than 1, the output will also be prepended with the identifier of the task which produced the output.
• write() / FileOutputFormat - Method and base class for custom file outputs. Supports custom object-to-bytes conversion.
• output()/ OutputFormat - Most generic output method, for data sinks that are not file based (such as storing the result in a database).

A DataSet can be input to multiple operations. Programs can write or print a data set and at the same time run additional transformations on them.

Examples

Standard data sink methods:

// text data
DataSet<String> textData = // [...]
// write DataSet to a file on the local file system
textData.writeAsText("file:///my/result/on/localFS");
// write DataSet to a file on an HDFS with a namenode running at nnHost:nnPort
textData.writeAsText("hdfs://nnHost:nnPort/my/result/on/localFS");
// write DataSet to a file and overwrite the file if it exists
textData.writeAsText("file:///my/result/on/localFS", WriteMode.OVERWRITE);
// tuples as lines with pipe as the separator "a|b|c"
DataSet<Tuple3<String, Integer, Double>> values = // [...]
values.writeAsCsv("file:///path/to/the/result/file", "\n", "|");
// this writes tuples in the text formatting "(a, b, c)", rather than as CSV lines
values.writeAsText("file:///path/to/the/result/file");
// this writes values as strings using a user-defined TextFormatter object
values.writeAsFormattedText("file:///path/to/the/result/file",
    new TextFormatter<Tuple2<Integer, Integer>>() {
        public String format (Tuple2<Integer, Integer> value) {
            return value.f1 + " - " + value.f0;
        }
    });

Using a custom output format:

DataSet<Tuple3<String, Integer, Double>> myResult = [...]
// write Tuple DataSet to a relational database
myResult.output(
    // build and configure OutputFormat
    JdbcOutputFormat.buildJdbcOutputFormat()
                    .setDrivername("org.apache.derby.jdbc.EmbeddedDriver")
                    .setDBUrl("jdbc:derby:memory:persons")
                    .setQuery("insert into persons (name, age, height) values (?,?,?)")
                    .finish()
    );

Locally Sorted Output

The output of a data sink can be locally sorted on specified fields in specified orders using tuple field positions or field expressions. This works for every output format.

The following examples show how to use this feature:

DataSet<Tuple3<Integer, String, Double>> tData = // [...]
DataSet<Tuple2<BookPojo, Double>> pData = // [...]
DataSet<String> sData = // [...]
// sort output on String field in ascending order
tData.sortPartition(1, Order.ASCENDING).print();
// sort output on Double field in descending and Integer field in ascending order
tData.sortPartition(2, Order.DESCENDING).sortPartition(0, Order.ASCENDING).print();
// sort output on the "author" field of nested BookPojo in descending order
pData.sortPartition("f0.author", Order.DESCENDING).writeAsText(...);
// sort output on the full tuple in ascending order
tData.sortPartition("*", Order.ASCENDING).writeAsCsv(...);
// sort atomic type (String) output in descending order
sData.sortPartition("*", Order.DESCENDING).writeAsText(...);

Globally sorted output is not supported yet.

Data sinks consume DataSets and are used to store or return them. Data sink operations are described using an OutputFormat. Flink comes with a variety of built-in output formats that are encapsulated behind operations on the DataSet:

• writeAsText() / TextOutputFormat - Writes elements line-wise as Strings. The Strings are obtained by calling the toString() method of each element.
• writeAsCsv(...) / CsvOutputFormat - Writes tuples as comma-separated value files. Row and field delimiters are configurable. The value for each field comes from the toString() method of the objects.
• print() / printToErr() - Prints the toString() value of each element on the standard out / standard error stream.
• write() / FileOutputFormat - Method and base class for custom file outputs. Supports custom object-to-bytes conversion.
• output()/ OutputFormat - Most generic output method, for data sinks that are not file based (such as storing the result in a database).

A DataSet can be input to multiple operations. Programs can write or print a data set and at the same time run additional transformations on them.

Examples

Standard data sink methods:

// text data
val textData: DataSet[String] = // [...]
// write DataSet to a file on the local file system
textData.writeAsText("file:///my/result/on/localFS")
// write DataSet to a file on an HDFS with a namenode running at nnHost:nnPort
textData.writeAsText("hdfs://nnHost:nnPort/my/result/on/localFS")
// write DataSet to a file and overwrite the file if it exists
textData.writeAsText("file:///my/result/on/localFS", WriteMode.OVERWRITE)
// tuples as lines with pipe as the separator "a|b|c"
val values: DataSet[(String, Int, Double)] = // [...]
values.writeAsCsv("file:///path/to/the/result/file", "\n", "|")
// this writes tuples in the text formatting "(a, b, c)", rather than as CSV lines
values.writeAsText("file:///path/to/the/result/file")
// this writes values as strings using a user-defined formatting
values map { tuple => tuple._1 + " - " + tuple._2 }
  .writeAsText("file:///path/to/the/result/file")

Locally Sorted Output

The output of a data sink can be locally sorted on specified fields in specified orders using tuple field positions or field expressions. This works for every output format.

The following examples show how to use this feature:

val tData: DataSet[(Int, String, Double)] = // [...]
val pData: DataSet[(BookPojo, Double)] = // [...]
val sData: DataSet[String] = // [...]
// sort output on String field in ascending order
tData.sortPartition(1, Order.ASCENDING).print()
// sort output on Double field in descending and Int field in ascending order
tData.sortPartition(2, Order.DESCENDING).sortPartition(0, Order.ASCENDING).print()
// sort output on the "author" field of nested BookPojo in descending order
pData.sortPartition("_1.author", Order.DESCENDING).writeAsText(...)
// sort output on the full tuple in ascending order
tData.sortPartition("_", Order.ASCENDING).writeAsCsv(...)
// sort atomic type (String) output in descending order
sData.sortPartition("_", Order.DESCENDING).writeAsText(...)

Globally sorted output is not supported yet.

Iteration Operators

Iterations implement loops in Flink programs. The iteration operators encapsulate a part of the program and execute it repeatedly, feeding back the result of one iteration (the partial solution) into the next iteration. There are two types of iterations in Flink: BulkIteration and DeltaIteration.

This section provides quick examples on how to use both operators. Check out the Introduction to Iterations page for a more detailed introduction.

Bulk Iterations

To create a BulkIteration call the iterate(int) method of the DataSet the iteration should start at. This will return an IterativeDataSet, which can be transformed with the regular operators. The single argument to the iterate call specifies the maximum number of iterations.

To specify the end of an iteration call the closeWith(DataSet) method on the IterativeDataSet to specify which transformation should be fed back to the next iteration. You can optionally specify a termination criterion with closeWith(DataSet, DataSet), which evaluates the second DataSet and terminates the iteration, if this DataSet is empty. If no termination criterion is specified, the iteration terminates after the given maximum number iterations.

The following example iteratively estimates the number Pi. The goal is to count the number of random points, which fall into the unit circle. In each iteration, a random point is picked. If this point lies inside the unit circle, we increment the count. Pi is then estimated as the resulting count divided by the number of iterations multiplied by 4.

final ExecutionEnvironment env = ExecutionEnvironment.getExecutionEnvironment();
// Create initial IterativeDataSet
IterativeDataSet<Integer> initial = env.fromElements(0).iterate(10000);
DataSet<Integer> iteration = initial.map(new MapFunction<Integer, Integer>() {
    @Override
    public Integer map(Integer i) throws Exception {
        double x = Math.random();
        double y = Math.random();
        return i + ((x * x + y * y < 1) ? 1 : 0);
    }
});
// Iteratively transform the IterativeDataSet
DataSet<Integer> count = initial.closeWith(iteration);
count.map(new MapFunction<Integer, Double>() {
    @Override
    public Double map(Integer count) throws Exception {
        return count / (double) 10000 * 4;
    }
}).print();
env.execute("Iterative Pi Example");

You can also check out the K-Means example, which uses a BulkIteration to cluster a set of unlabeled points.

Delta Iterations

Delta iterations exploit the fact that certain algorithms do not change every data point of the solution in each iteration.

In addition to the partial solution that is fed back (called workset) in every iteration, delta iterations maintain state across iterations (called solution set), which can be updated through deltas. The result of the iterative computation is the state after the last iteration. Please refer to the Introduction to Iterations for an overview of the basic principle of delta iterations.

Defining a DeltaIteration is similar to defining a BulkIteration. For delta iterations, two data sets form the input to each iteration (workset and solution set), and two data sets are produced as the result (new workset, solution set delta) in each iteration.

To create a DeltaIteration call the iterateDelta(DataSet, int, int) (or iterateDelta(DataSet, int, int[]) respectively). This method is called on the initial solution set. The arguments are the initial delta set, the maximum number of iterations and the key positions. The returned DeltaIteration object gives you access to the DataSets representing the workset and solution set via the methods iteration.getWorkset() and iteration.getSolutionSet().

Below is an example for the syntax of a delta iteration

// read the initial data sets
DataSet<Tuple2<Long, Double>> initialSolutionSet = // [...]
DataSet<Tuple2<Long, Double>> initialDeltaSet = // [...]
int maxIterations = 100;
int keyPosition = 0;
DeltaIteration<Tuple2<Long, Double>, Tuple2<Long, Double>> iteration = initialSolutionSet
    .iterateDelta(initialDeltaSet, maxIterations, keyPosition);
DataSet<Tuple2<Long, Double>> candidateUpdates = iteration.getWorkset()
    .groupBy(1)
    .reduceGroup(new ComputeCandidateChanges());
DataSet<Tuple2<Long, Double>> deltas = candidateUpdates
    .join(iteration.getSolutionSet())
    .where(0)
    .equalTo(0)
    .with(new CompareChangesToCurrent());
DataSet<Tuple2<Long, Double>> nextWorkset = deltas
    .filter(new FilterByThreshold());
iteration.closeWith(deltas, nextWorkset)
    .writeAsCsv(outputPath);

Bulk Iterations

To create a BulkIteration call the iterate(int) method of the DataSet the iteration should start at and also specify a step function. The step function gets the input DataSet for the current iteration and must return a new DataSet. The parameter of the iterate call is the maximum number of iterations after which to stop.

There is also the iterateWithTermination(int) function that accepts a step function that returns two DataSets: The result of the iteration step and a termination criterion. The iterations are stopped once the termination criterion DataSet is empty.

The following example iteratively estimates the number Pi. The goal is to count the number of random points, which fall into the unit circle. In each iteration, a random point is picked. If this point lies inside the unit circle, we increment the count. Pi is then estimated as the resulting count divided by the number of iterations multiplied by 4.

val env = ExecutionEnvironment.getExecutionEnvironment()
// Create initial DataSet
val initial = env.fromElements(0)
val count = initial.iterate(10000) { iterationInput: DataSet[Int] =>
  val result = iterationInput.map { i =>
    val x = Math.random()
    val y = Math.random()
    i + (if (x * x + y * y < 1) 1 else 0)
  }
  result
}
val result = count map { c => c / 10000.0 * 4 }
result.print()
env.execute("Iterative Pi Example")

You can also check out the K-Means example, which uses a BulkIteration to cluster a set of unlabeled points.

Delta Iterations

Delta iterations exploit the fact that certain algorithms do not change every data point of the solution in each iteration.

In addition to the partial solution that is fed back (called workset) in every iteration, delta iterations maintain state across iterations (called solution set), which can be updated through deltas. The result of the iterative computation is the state after the last iteration. Please refer to the Introduction to Iterations for an overview of the basic principle of delta iterations.

Defining a DeltaIteration is similar to defining a BulkIteration. For delta iterations, two data sets form the input to each iteration (workset and solution set), and two data sets are produced as the result (new workset, solution set delta) in each iteration.

To create a DeltaIteration call the iterateDelta(initialWorkset, maxIterations, key) on the initial solution set. The step function takes two parameters: (solutionSet, workset), and must return two values: (solutionSetDelta, newWorkset).

Below is an example for the syntax of a delta iteration

// read the initial data sets
val initialSolutionSet: DataSet[(Long, Double)] = // [...]
val initialWorkset: DataSet[(Long, Double)] = // [...]
val maxIterations = 100
val keyPosition = 0
val result = initialSolutionSet.iterateDelta(initialWorkset, maxIterations, Array(keyPosition)) {
  (solution, workset) =>
    val candidateUpdates = workset.groupBy(1).reduceGroup(new ComputeCandidateChanges())
    val deltas = candidateUpdates.join(solution).where(0).equalTo(0)(new CompareChangesToCurrent())
    val nextWorkset = deltas.filter(new FilterByThreshold())
    (deltas, nextWorkset)
}
result.writeAsCsv(outputPath)
env.execute()

Operating on data objects in functions

Flink’s runtime exchanges data with user functions in form of Java objects. Functions receive input objects from the runtime as method parameters and return output objects as result. Because these objects are accessed by user functions and runtime code, it is very important to understand and follow the rules about how the user code may access, i.e., read and modify, these objects.

User functions receive objects from Flink’s runtime either as regular method parameters (like a MapFunction) or through an Iterable parameter (like a GroupReduceFunction). We refer to objects that the runtime passes to a user function as input objects. User functions can emit objects to the Flink runtime either as a method return value (like a MapFunction) or through a Collector (like a FlatMapFunction). We refer to objects which have been emitted by the user function to the runtime as output objects.

Flink’s DataSet API features two modes that differ in how Flink’s runtime creates or reuses input objects. This behavior affects the guarantees and constraints for how user functions may interact with input and output objects. The following sections define these rules and give coding guidelines to write safe user function code.

Object-Reuse Disabled (DEFAULT)

By default, Flink operates in object-reuse disabled mode. This mode ensures that functions always receive new input objects within a function call. The object-reuse disabled mode gives better guarantees and is safer to use. However, it comes with a certain processing overhead and might cause higher Java garbage collection activity. The following table explains how user functions may access input and output objects in object-reuse disabled mode.

Coding guidelines for the object-reuse disabled (default) mode:

• Do not remember and read input objects across method calls.
• Do not read objects after you emitted them.

Object-Reuse Enabled

In object-reuse enabled mode, Flink’s runtime minimizes the number of object instantiations. This can improve the performance and can reduce the Java garbage collection pressure. The object-reuse enabled mode is activated by calling ExecutionConfig.enableObjectReuse(). The following table explains how user functions may access input and output objects in object-reuse enabled mode.

Coding guidelines for object-reuse enabled:

• Do not remember input objects received from an Iterable.
• Do not remember and read input objects across method calls.
• Do not modify or emit input objects, except for input objects of MapFunction, FlatMapFunction, MapPartitionFunction, GroupReduceFunction, GroupCombineFunction, CoGroupFunction, and InputFormat.next(reuse).
• To reduce object instantiations, you can always emit a dedicated output object which is repeatedly modified but never read.

Debugging

Before running a data analysis program on a large data set in a distributed cluster, it is a good idea to make sure that the implemented algorithm works as desired. Hence, implementing data analysis programs is usually an incremental process of checking results, debugging, and improving.

Flink provides a few nice features to significantly ease the development process of data analysis programs by supporting local debugging from within an IDE, injection of test data, and collection of result data. This section give some hints how to ease the development of Flink programs.

Local Execution Environment

A LocalEnvironment starts a Flink system within the same JVM process it was created in. If you start the LocalEnvironment from an IDE, you can set breakpoints in your code and easily debug your program.

A LocalEnvironment is created and used as follows:

final ExecutionEnvironment env = ExecutionEnvironment.createLocalEnvironment();
DataSet<String> lines = env.readTextFile(pathToTextFile);
// build your program
env.execute();

val env = ExecutionEnvironment.createLocalEnvironment()
val lines = env.readTextFile(pathToTextFile)
// build your program
env.execute()

Collection Data Sources and Sinks

Providing input for an analysis program and checking its output is cumbersome when done by creating input files and reading output files. Flink features special data sources and sinks which are backed by Java collections to ease testing. Once a program has been tested, the sources and sinks can be easily replaced by sources and sinks that read from / write to external data stores such as HDFS.

Collection data sources can be used as follows:

final ExecutionEnvironment env = ExecutionEnvironment.createLocalEnvironment();
// Create a DataSet from a list of elements
DataSet<Integer> myInts = env.fromElements(1, 2, 3, 4, 5);
// Create a DataSet from any Java collection
List<Tuple2<String, Integer>> data = ...
DataSet<Tuple2<String, Integer>> myTuples = env.fromCollection(data);
// Create a DataSet from an Iterator
Iterator<Long> longIt = ...
DataSet<Long> myLongs = env.fromCollection(longIt, Long.class);

A collection data sink is specified as follows:

DataSet<Tuple2<String, Integer>> myResult = ...
List<Tuple2<String, Integer>> outData = new ArrayList<Tuple2<String, Integer>>();
myResult.output(new LocalCollectionOutputFormat(outData));

Note: Currently, the collection data sink is restricted to local execution, as a debugging tool.

val env = ExecutionEnvironment.createLocalEnvironment()
// Create a DataSet from a list of elements
val myInts = env.fromElements(1, 2, 3, 4, 5)
// Create a DataSet from any Collection
val data: Seq[(String, Int)] = ...
val myTuples = env.fromCollection(data)
// Create a DataSet from an Iterator
val longIt: Iterator[Long] = ...
val myLongs = env.fromCollection(longIt)

Note: Currently, the collection data source requires that data types and iterators implement Serializable. Furthermore, collection data sources can not be executed in parallel ( parallelism = 1).

Semantic Annotations

Semantic annotations can be used to give Flink hints about the behavior of a function. They tell the system which fields of a function’s input the function reads and evaluates and which fields it unmodified forwards from its input to its output. Semantic annotations are a powerful means to speed up execution, because they allow the system to reason about reusing sort orders or partitions across multiple operations. Using semantic annotations may eventually save the program from unnecessary data shuffling or unnecessary sorts and significantly improve the performance of a program.

Note: The use of semantic annotations is optional. However, it is absolutely crucial to be conservative when providing semantic annotations! Incorrect semantic annotations will cause Flink to make incorrect assumptions about your program and might eventually lead to incorrect results. If the behavior of an operator is not clearly predictable, no annotation should be provided. Please read the documentation carefully.

The following semantic annotations are currently supported.

Forwarded Fields Annotation

Forwarded fields information declares input fields which are unmodified forwarded by a function to the same position or to another position in the output. This information is used by the optimizer to infer whether a data property such as sorting or partitioning is preserved by a function. For functions that operate on groups of input elements such as GroupReduce, GroupCombine, CoGroup, and MapPartition, all fields that are defined as forwarded fields must always be jointly forwarded from the same input element. The forwarded fields of each element that is emitted by a group-wise function may originate from a different element of the function’s input group.

Field forward information is specified using field expressions. Fields that are forwarded to the same position in the output can be specified by their position. The specified position must be valid for the input and output data type and have the same type. For example the String "f2" declares that the third field of a Java input tuple is always equal to the third field in the output tuple.

Fields which are unmodified forwarded to another position in the output are declared by specifying the source field in the input and the target field in the output as field expressions. The String "f0->f2" denotes that the first field of the Java input tuple is unchanged copied to the third field of the Java output tuple. The wildcard expression * can be used to refer to a whole input or output type, i.e., "f0->*" denotes that the output of a function is always equal to the first field of its Java input tuple.

Multiple forwarded fields can be declared in a single String by separating them with semicolons as "f0; f2->f1; f3->f2" or in separate Strings "f0", "f2->f1", "f3->f2". When specifying forwarded fields it is not required that all forwarded fields are declared, but all declarations must be correct.

Forwarded field information can be declared by attaching Java annotations on function class definitions or by passing them as operator arguments after invoking a function on a DataSet as shown below.

Function Class Annotations

• @ForwardedFields for single input functions such as Map and Reduce.
• @ForwardedFieldsFirst for the first input of a functions with two inputs such as Join and CoGroup.
• @ForwardedFieldsSecond for the second input of a functions with two inputs such as Join and CoGroup.

Operator Arguments

• data.map(myMapFnc).withForwardedFields() for single input function such as Map and Reduce.
• data1.join(data2).where().equalTo().with(myJoinFnc).withForwardFieldsFirst() for the first input of a function with two inputs such as Join and CoGroup.
• data1.join(data2).where().equalTo().with(myJoinFnc).withForwardFieldsSecond() for the second input of a function with two inputs such as Join and CoGroup.

Please note that it is not possible to overwrite field forward information which was specified as a class annotation by operator arguments.

Example

The following example shows how to declare forwarded field information using a function class annotation:

@ForwardedFields("f0->f2")
public class MyMap implements
              MapFunction<Tuple2<Integer, Integer>, Tuple3<String, Integer, Integer>> {
  @Override
  public Tuple3<String, Integer, Integer> map(Tuple2<Integer, Integer> val) {
    return new Tuple3<String, Integer, Integer>("foo", val.f1 / 2, val.f0);
  }
}

@ForwardedFields("_1->_3")
class MyMap extends MapFunction[(Int, Int), (String, Int, Int)]{
   def map(value: (Int, Int)): (String, Int, Int) = {
    return ("foo", value._2 / 2, value._1)
  }
}

Non-Forwarded Fields

Non-forwarded fields information declares all fields which are not preserved on the same position in a function’s output. The values of all other fields are considered to be preserved at the same position in the output. Hence, non-forwarded fields information is inverse to forwarded fields information. Non-forwarded field information for group-wise operators such as GroupReduce, GroupCombine, CoGroup, and MapPartition must fulfill the same requirements as for forwarded field information.

IMPORTANT: The specification of non-forwarded fields information is optional. However if used, ALL! non-forwarded fields must be specified, because all other fields are considered to be forwarded in place. It is safe to declare a forwarded field as non-forwarded.

Non-forwarded fields are specified as a list of field expressions. The list can be either given as a single String with field expressions separated by semicolons or as multiple Strings. For example both "f1; f3" and "f1", "f3" declare that the second and fourth field of a Java tuple are not preserved in place and all other fields are preserved in place. Non-forwarded field information can only be specified for functions which have identical input and output types.

Non-forwarded field information is specified as function class annotations using the following annotations:

• @NonForwardedFields for single input functions such as Map and Reduce.
• @NonForwardedFieldsFirst for the first input of a function with two inputs such as Join and CoGroup.
• @NonForwardedFieldsSecond for the second input of a function with two inputs such as Join and CoGroup.

Example

The following example shows how to declare non-forwarded field information:

@NonForwardedFields("f1") // second field is not forwarded
public class MyMap implements
              MapFunction<Tuple2<Integer, Integer>, Tuple2<Integer, Integer>> {
  @Override
  public Tuple2<Integer, Integer> map(Tuple2<Integer, Integer> val) {
    return new Tuple2<Integer, Integer>(val.f0, val.f1 / 2);
  }
}

@NonForwardedFields("_2") // second field is not forwarded
class MyMap extends MapFunction[(Int, Int), (Int, Int)]{
  def map(value: (Int, Int)): (Int, Int) = {
    return (value._1, value._2 / 2)
  }
}

Read Fields

Read fields information declares all fields that are accessed and evaluated by a function, i.e., all fields that are used by the function to compute its result. For example, fields which are evaluated in conditional statements or used for computations must be marked as read when specifying read fields information. Fields which are only unmodified forwarded to the output without evaluating their values or fields which are not accessed at all are not considered to be read.

IMPORTANT: The specification of read fields information is optional. However if used, ALL! read fields must be specified. It is safe to declare a non-read field as read.

Read fields are specified as a list of field expressions. The list can be either given as a single String with field expressions separated by semicolons or as multiple Strings. For example both "f1; f3" and "f1", "f3" declare that the second and fourth field of a Java tuple are read and evaluated by the function.

Read field information is specified as function class annotations using the following annotations:

• @ReadFields for single input functions such as Map and Reduce.
• @ReadFieldsFirst for the first input of a function with two inputs such as Join and CoGroup.
• @ReadFieldsSecond for the second input of a function with two inputs such as Join and CoGroup.

Example

The following example shows how to declare read field information:

@ReadFields("f0; f3") // f0 and f3 are read and evaluated by the function.
public class MyMap implements
              MapFunction<Tuple4<Integer, Integer, Integer, Integer>,
                          Tuple2<Integer, Integer>> {
  @Override
  public Tuple2<Integer, Integer> map(Tuple4<Integer, Integer, Integer, Integer> val) {
    if(val.f0 == 42) {
      return new Tuple2<Integer, Integer>(val.f0, val.f1);
    } else {
      return new Tuple2<Integer, Integer>(val.f3+10, val.f1);
    }
  }
}

@ReadFields("_1; _4") // _1 and _4 are read and evaluated by the function.
class MyMap extends MapFunction[(Int, Int, Int, Int), (Int, Int)]{
   def map(value: (Int, Int, Int, Int)): (Int, Int) = {
    if (value._1 == 42) {
      return (value._1, value._2)
    } else {
      return (value._4 + 10, value._2)
    }
  }
}

Broadcast Variables

Broadcast variables allow you to make a data set available to all parallel instances of an operation, in addition to the regular input of the operation. This is useful for auxiliary data sets, or data-dependent parameterization. The data set will then be accessible at the operator as a Collection.

• Broadcast: broadcast sets are registered by name via withBroadcastSet(DataSet, String), and
• Access: accessible via getRuntimeContext().getBroadcastVariable(String) at the target operator.

// 1. The DataSet to be broadcast
DataSet<Integer> toBroadcast = env.fromElements(1, 2, 3);
DataSet<String> data = env.fromElements("a", "b");
data.map(new RichMapFunction<String, String>() {
    @Override
    public void open(Configuration parameters) throws Exception {
      // 3. Access the broadcast DataSet as a Collection
      Collection<Integer> broadcastSet = getRuntimeContext().getBroadcastVariable("broadcastSetName");
    }
    @Override
    public String map(String value) throws Exception {
        ...
    }
}).withBroadcastSet(toBroadcast, "broadcastSetName"); // 2. Broadcast the DataSet

Make sure that the names (broadcastSetName in the previous example) match when registering and accessing broadcast data sets. For a complete example program, have a look at K-Means Algorithm.

// 1. The DataSet to be broadcast
val toBroadcast = env.fromElements(1, 2, 3)
val data = env.fromElements("a", "b")
data.map(new RichMapFunction[String, String]() {
    var broadcastSet: Traversable[String] = null
    override def open(config: Configuration): Unit = {
      // 3. Access the broadcast DataSet as a Collection
      broadcastSet = getRuntimeContext().getBroadcastVariable[String]("broadcastSetName").asScala
    }
    def map(in: String): String = {
        ...
    }
}).withBroadcastSet(toBroadcast, "broadcastSetName") // 2. Broadcast the DataSet