Skip to content


Subversion checkout URL

You can clone with
Download ZIP
A Python wrapper for Cascading
Python Java Shell
Pull request Compare This branch is 21 commits ahead, 3 commits behind master.
Fetching latest commit...
Cannot retrieve the latest commit at this time.
Failed to load latest commit information.

NOTE AS OF 4/7/2012: We are in the process of making significant changes to the code base, revamping the documentation, and adding a tutorial. Please check back in a few days for these as well.


PyCascading is a Python wrapper for Cascading. You can control the full data processing workflow from Python.

  • Pipelines are built with Python operators
  • User-defined functions are written in Python
  • Passing arbitrary contexts to user-defined functions
  • Caching of interim results in pipes for faster replay
  • Uses Jython 2.5.2, easy integration with Java and Python libraries


There can't be a big data tutorial without counting words. Here it is:

def split_words(tuple):
    for word in tuple.get(1).split():
        yield [word]

def main():
    input | split_words | GroupBy('word') | Count() | output

Above, the user-defined function that reshapes the stream is annotated with a PyCascading decorator, and the workflow is created by chaining operations into each other.

More examples for the different use cases can be found in the examples folder. See also the the docstrings in the sources for a complete documentation of the arguments.

To try the examples, first build the Java sources as described below in the Building section. Then, change to the 'examples' folder, and issue either


for a simulated Hadoop local run, or

../ -m -s hadoop_server

to deploy automatically on a Hadoop server. hadoop_server is the SSH address of an account where the master jar and script will be scp'd to. Note that the '-m' option has to be used only once in the beginning. The '-m' option copies the master jar to the server, and any subsequent deploys will use this master jar, and only the actual Python script will be copied over the network.


PyCascading may be used in one of two modes: in local Hadoop mode or with remote Hadoop deployment. Please note that you need to specify the locations of the dependencies in the java/ file.

In local mode, the script is executed in Hadoop's local mode. All files reside on the local file system, and creating a bundled deployment jar is not necessary.

To run in this mode, use the script, with the first parameter being the PyCascading script. Additional command line parameters may be used to pass on to the script.

In Hadoop mode, we assume that Hadoop runs on a remote SSH server (or localhost). First, a master jar is built and copied to the server. This jar contains all the PyCascading classes and other dependencies (but not Hadoop) needed to run a job, and may get rather large if there are a few external jars included. For this reason it is copied to the Hadoop deployment server only once, and whenever a new PyCascading script is run by the user, only the Pythn script is copied to the remote server and bundled there for submission to Hadoop. The first few variables in the script specify the Hadoop server and the folders where the deployment files should be placed.

Use the script to deploy a PyCascading script to the remote Hadoop server.


Requirements for building:

Requirements for running:

  • Hadoop installed and set up on the target server (
  • SSH access to the remote server
  • If testing scripts locally, a reasonable JVM callable by "java"

PyCascading consists of Java and Python sources. Python sources need no compiling, but the Java part needs to be built with Ant. For this, change to the 'java' folder, and invoke ant. This should build the sources and create a master jar for job submission.

The locations of the Jython, Cascading, and Hadoop folders on the file system are specified in the java/ file. You need to correctly specify these before compiling the source.

Also, check the script and the locations defined in the beginning of that file on where to put the jar files on the Hadoop server.


Have a bug or feature request? Please create an issue here on GitHub!

Mailing list

Currently we are using the cascading-user mailing list for discussions. Any questions, please ask there.


Gabor Szabo


Copyright 2011 Twitter, Inc.

Licensed under the Apache License, Version 2.0

Something went wrong with that request. Please try again.