Permalink
438dbca Nov 8, 2016
@bzz @agoodm
53 lines (29 sloc) 2.98 KB

Overview

Python interpreter for Apache Zeppelin

Architecture

Current interpreter implementation spawns new system python process through ProcessBuilder and re-directs it's stdin\strout to Zeppelin

Details

  • UnitTests

    To run full suit of tests, including ones that depend on real Python interpreter AND external libraries installed (like Pandas, Pandasql, etc) do

    mvn -Dpython.test.exclude='' test -pl python -am
    
  • Py4j support

    Py4j enables Python programs to dynamically access Java objects in a JVM. It is required in order to use Zeppelin dynamic forms feature.

  • bootstrap process

    Interpreter environment is setup with thex bootstrap.py It defines help() and z convenience functions

Dev prerequisites

  • Python 2 or 3 installed with py4j (0.9.2) and matplotlib (1.31 or later) installed on each

  • Tests only checks the interpreter logic and starts any Python process! Python process is mocked with a class that simply output it input.

  • Code wrote in bootstrap.py and bootstrap_input.py should always be Python 2 and 3 compliant.

  • Use PEP8 convention for python code.

Technical overview

  • When interpreter is starting it launches a python process inside a Java ProcessBuilder. Python is started with -i (interactive mode) and -u (unbuffered stdin, stdout and stderr) options. Thus the interpreter has a "sleeping" python process.

  • Interpreter sends command to python with a Java outputStreamWiter and read from an InputStreamReader. To know when stop reading stdout, interpreter sends print "*!?flush reader!?*"after each command and reads stdout until he receives back the *!?flush reader!?*.

  • When interpreter is starting, it sends some Python code (bootstrap.py and bootstrap_input.py) to initialize default behavior and functions (help(), z.input()...). bootstrap_input.py is sent only if py4j library is detected inside Python process.

  • Py4J python and java libraries is used to load Input zeppelin Java class into the python process (make java code with python code !). Therefore the interpreter can directly create Zeppelin input form inside the Python process (and eventually with some python variable already defined). JVM opens a random open port to be accessible from python process.

  • JavaBuilder can't send SIGINT signal to interrupt paragraph execution. Therefore interpreter directly send a kill SIGINT PID to python process to interrupt execution. Python process catch SIGINT signal with some code defined in bootstrap.py

  • Matplotlib figures are displayed inline with the notebook automatically using a built-in backend for zeppelin in conjunction with a post-execute hook.

  • %python.sql support for Pandas DataFrames is optional and provided using https://github.com/yhat/pandasql if user have one installed