Data Types #

In Apache Flink’s Python DataStream API, a data type describes the type of a value in the DataStream ecosystem. It can be used to declare input and output types of operations and informs the system how to serailize elements.

Pickle Serialization #

If the type has not been declared, data would be serialized or deserialized using Pickle. For example, the program below specifies no data types.

from pyflink.datastream import StreamExecutionEnvironment


def processing():
    env = StreamExecutionEnvironment.get_execution_environment()
    env.set_parallelism(1)
    env.from_collection(collection=[(1, 'aaa'), (2, 'bbb')]) \
        .map(lambda record: (record[0]+1, record[1].upper())) \
        .print()  # note: print to stdout on the worker machine

    env.execute()


if __name__ == '__main__':
    processing()

However, types need to be specified when:

• Passing Python records to Java operations.
• Improve serialization and deserialization performance.

Passing Python records to Java operations #

Since Java operators or functions can not identify Python data, types need to be provided to help to convert Python types to Java types for processing. For example, types need to be provided if you want to output data using the FileSink which is implemented in Java.

from pyflink.common.serialization import Encoder
from pyflink.common.typeinfo import Types
from pyflink.datastream import StreamExecutionEnvironment
from pyflink.datastream.connectors.file_system import FileSink


def file_sink():
    env = StreamExecutionEnvironment.get_execution_environment()
    env.set_parallelism(1)
    env.from_collection(collection=[(1, 'aaa'), (2, 'bbb')]) \
        .map(lambda record: (record[0]+1, record[1].upper()),
             output_type=Types.ROW([Types.INT(), Types.STRING()])) \
        .add_sink(FileSink
                  .for_row_format('/tmp/output', Encoder.simple_string_encoder())
                  .build())

    env.execute()


if __name__ == '__main__':
    file_sink()

Improve serialization and deserialization performance #

Even though data can be serialized and deserialized through Pickle, performance will be better if types are provided. Explicit types allow PyFlink to use efficient serializers when moving records through the pipeline.

Supported Data Types #

You can use pyflink.common.typeinfo.Types to define types in Python DataStream API. The table below shows the types supported now and how to define them:

The table below shows the array types supported: