This documentation is for an out-of-date version of Apache Flink. We recommend you use the latest stable version.

Python Programming Guide Beta

Analysis 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 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.

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

Example Program

The following program is a complete, working example of WordCount. You can copy & paste the code to run it locally.

from flink.plan.Environment import get_environment
from flink.functions.GroupReduceFunction import GroupReduceFunction

class Adder(GroupReduceFunction):
  def reduce(self, iterator, collector):
    count, word =
    count += sum([x[0] for x in iterator])
    collector.collect((count, word))

env = get_environment()
data = env.from_elements("Who's there?",
 "I think I hear them. Stand, ho! Who's there?")

data \
  .flat_map(lambda x, c: [(1, word) for word in x.lower().split()]) \
  .group_by(1) \
  .reduce_group(Adder(), combinable=True) \


Back to top

Program Skeleton

As we already saw in the example, Flink programs look like regular python programs. Each program consists of the same basic parts:

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

We will now give an overview of each of those steps but please refer to the respective sections for more details.

The Environment is the basis for all Flink programs. You can obtain one using these static methods on class Environment:


For specifying data sources the execution environment has several methods to read from files. To just read a text file as a sequence of lines, you can use:

env = get_environment()
text = env.read_text("file:///path/to/file")

This will give you a DataSet on which you can then apply transformations. For more information on data sources and input formats, please refer to Data Sources.

Once you have a DataSet you can apply transformations to create a new DataSet which you can then write to a file, transform again, or combine with other DataSets. You apply transformations by calling methods on DataSet with your own custom transformation function. For example, a map transformation looks like this: x: x*2)

This will create a new DataSet by doubling every value in the original DataSet. For more information and a list of all the transformations, please refer to Transformations.

Once you have a DataSet that needs to be written to disk you can call one of these methods on DataSet:

data.write_text("<file-path>", WriteMode=Constants.NO_OVERWRITE)
write_csv("<file-path>", line_delimiter='\n', field_delimiter=',', write_mode=Constants.NO_OVERWRITE)

The last method is only useful for developing/debugging on a local machine, it will output the contents of the DataSet to standard output. (Note that in a cluster, the result goes to the standard out stream of the cluster nodes and ends up in the .out files of the workers). The first two do as the name suggests. Please refer to Data Sinks for more information on writing to files.

Once you specified the complete program you need to call execute on the Environment. This will submit your program for execution on a cluster. Back to top

Project setup

Apart from setting up Flink, no additional work is required. The python package can be found in the /resource folder of your Flink distribution. The flink package, along with the plan and optional packages are automatically distributed among the cluster via HDFS when running a job.

The Python API was tested on Linux/Windows systems that have Python 2.7 or 3.4 installed.

By default Flink will start python processes by calling “python”. By setting the “python.binary.path” key in the flink-conf.yaml you can modify this behaviour to use a binary of your choice.

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 one of the execute() methods is invoked on the Environment object.

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

Back to top


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.

</tr> </tr>
Transformation Description

Takes one element and produces one element. x: x * 2)

Takes one element and produces zero, one, or more elements.

  lambda x,c: [(1,word) for word in line.lower().split() for line in x])

Transforms a parallel partition in a single function call. The function get the partition as an `Iterator` and can produce an arbitrary number of result values. The number of elements in each partition depends on the parallelism and previous operations.

data.map_partition(lambda x,c: [value * 2 for value in x])

Evaluates a boolean function for each element and retains those for which the function returns true.

data.filter(lambda x: x > 1000)

Combines a group of elements into a single element by repeatedly combining two elements into one. Reduce may be applied on a full data set, or on a grouped data set.

data.reduce(lambda x,y : x + y)

Combines a group of elements into one or more elements. ReduceGroup may be applied on a full data set, or on a grouped data set.

class Adder(GroupReduceFunction):
  def reduce(self, iterator, collector):
    count, word =
    count += sum([x[0] for x in iterator)      
    collector.collect((count, word))


Performs a built-in operation (sum, min, max) on one field of all the Tuples in a data set or in each group of a data set. Aggregation can be applied on a full dataset or on a grouped data set.

# This code finds the sum of all of the values in the first field and the maximum of all of the values in the second field
data.aggregate(Aggregation.Sum, 0).and_agg(Aggregation.Max, 1)

# min(), max(), and sum() syntactic sugar functions are also available
data.sum(0).and_agg(Aggregation.Max, 1)
Join Joins two data sets by creating all pairs of elements that are equal on their keys. Optionally uses a JoinFunction to turn the pair of elements into a single element. See keys on how to define join keys.
# 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.
result = input1.join(input2).where(0).equal_to(1)

The two-dimensional variant of the reduce operation. Groups each input on one or more fields and then joins the groups. The transformation function is called per pair of groups. See keys on how to define coGroup keys.


Builds the Cartesian product (cross product) of two inputs, creating all pairs of elements. Optionally uses a CrossFunction to turn the pair of elements into a single element.

result = data1.cross(data2)

Produces the union of two data sets.


Assigns consecutive indexes to each element. For more information, please refer to the [Zip Elements Guide](zip_elements_guide.html#zip-with-a-dense-index).


Back to top

Specifying Keys

Some transformations (like Join or CoGroup) require that a key is defined on its argument DataSets, and other transformations (Reduce, GroupReduce) allow that the DataSet is grouped on a key before they are applied.

A DataSet is grouped as

reduced = data \
  .group_by(<define key here>) \
  .reduce_group(<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 a data set of Tuples on one or more fields of the Tuple:

reduced = data \
  .group_by(0) \
  .reduce_group(<do something>)

The data set is grouped on the first field of the tuples. The group-reduce function will thus receive groups of tuples with the same value in the first field.

grouped = data \
  .group_by(0,1) \
  .reduce(/*do something*/)

The data set is grouped on the composite key consisting of the first and the second fields, therefore the reduce function will receive groups with the same value for both fields.

A note on nested Tuples: If you have a DataSet with a nested tuple specifying group_by(<index of tuple>) will cause the system to use the full tuple as a key.

Back to top

Certain operations require user-defined functions, whereas all of them accept lambda functions and rich functions as arguments.

data.filter(lambda x: x > 5)
class Filter(FilterFunction):
    def filter(self, value):
        return value > 5


Rich functions allow the use of imported functions, provide access to broadcast-variables, can be parameterized using init(), and are the go-to-option for complex functions. They are also the only way to define an optional combine function for a reduce operation.

Lambda functions allow the easy insertion of one-liners. Note that a lambda function has to return an iterable, if the operation can return multiple values. (All functions receiving a collector argument)

Back to top

Data Types

Flink’s Python API currently only offers native support for primitive python types (int, float, bool, string) and byte arrays.

The type support can be extended by passing a serializer, deserializer and type class to the environment.

class MyObj(object):
    def __init__(self, i):
        self.value = i

class MySerializer(object):
    def serialize(self, value):
        return struct.pack(">i", value.value)

class MyDeserializer(object):
    def _deserialize(self, read):
        i = struct.unpack(">i", read(4))[0]
        return MyObj(i)

env.register_type(MyObj, MySerializer(), MyDeserializer())


You can use the tuples (or lists) for composite types. Python tuples are mapped to the Flink Tuple type, that contain a fix number of fields of various types (up to 25). Every field of a tuple can be a primitive type - including further tuples, resulting in nested tuples.

word_counts = env.from_elements(("hello", 1), ("world",2))

counts = x: x[1])

When working with operators that require a Key for grouping or matching records, Tuples let you simply specify the positions of the fields to be used as key. You can specify more than one position to use composite keys (see Section Data Transformations).

wordCounts \
    .group_by(0) \

Back to top

Data Sources

Data sources create the initial data sets, such as from files or from collections.


  • read_text(path) - Reads files line wise and returns them as Strings.
  • read_csv(path, type) - Parses files of comma (or another char) delimited fields. Returns a DataSet of tuples. Supports the basic java types and their Value counterparts as field types.


  • from_elements(*args) - Creates a data set from a Seq. All elements
  • generate_sequence(from, to) - Generates the sequence of numbers in the given interval, in parallel.


env  = get_environment

\# read text file from local files system
localLiens = env.read_text("file:#/path/to/my/textfile")

\# read text file from a HDFS running at nnHost:nnPort
hdfsLines = env.read_text("hdfs://nnHost:nnPort/path/to/my/textfile")

\# read a CSV file with three fields, schema defined using constants defined in flink.plan.Constants
csvInput = env.read_csv("hdfs:///the/CSV/file", (INT, STRING, DOUBLE))

\# create a set from some given elements
values = env.from_elements("Foo", "bar", "foobar", "fubar")

\# generate a number sequence
numbers = env.generate_sequence(1, 10000000)

Back to top

Data Sinks

Data sinks consume DataSets and are used to store or return them:

  • write_text() - Writes elements line-wise as Strings. The Strings are obtained by calling the str() method of each element.
  • write_csv(...) - Writes tuples as comma-separated value files. Row and field delimiters are configurable. The value for each field comes from the str() method of the objects.
  • output() - Prints the str() value of each element on the standard out.

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.


Standard data sink methods:

 write DataSet to a file on the local file system

 write DataSet to a file on a HDFS with a namenode running at nnHost:nnPort

 write DataSet to a file and overwrite the file if it exists
textData.write_text("file:///my/result/on/localFS", WriteMode.OVERWRITE)

 tuples as lines with pipe as the separator "a|b|c"
values.write_csv("file:///path/to/the/result/file", line_delimiter="\n", field_delimiter="|")

 this writes tuples in the text formatting "(a, b, c)", rather than as CSV lines

Back to top

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 with_broadcast_set(DataSet, String)
  • Access: accessible via self.context.get_broadcast_variable(String) at the target operator
class MapperBcv(MapFunction):
    def map(self, value):
        factor = self.context.get_broadcast_variable("bcv")[0][0]
        return value * factor

# 1. The DataSet to be broadcast
toBroadcast = env.from_elements(1, 2, 3)
data = env.from_elements("a", "b")

# 2. Broadcast the DataSet"bcv", toBroadcast)

Make sure that the names (bcv in the previous example) match when registering and accessing broadcast data sets.

Note: As the content of broadcast variables is kept in-memory on each node, it should not become too large. For simpler things like scalar values you can simply parameterize the rich function.

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 (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 or degree of parallelism (DOP).

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

Execution Environment Level

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 set_parallelism() method. To execute all operators, data sources, and data sinks of the WordCount example program with a parallelism of 3, set the default parallelism of the execution environment as follows:

env = get_environment()

text.flat_map(lambda x,c: x.lower().split()) \
    .group_by(1) \
    .reduce_group(Adder(), combinable=True) \


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

Executing Plans

To run the plan with Flink, go to your Flink distribution, and run the script from the /bin folder. The script containing the plan has to be passed as the first argument, followed by a number of additional python packages, and finally, separated by - additional arguments that will be fed to the script.

./bin/ <Script>[ <pathToPackage1>[ <pathToPackageX]][ - <param1>[ <paramX>]]

Back to top