I wrote this launcher script a while back expressly for that purpose. I wanted to be able to interact with the pyspark shell from within the bpython(1) code-completion interpreter and WING IDE, or any IDE for that matter because they have code completion as well as provide a complete development experience. Learning Spark core by just typing 'pyspark' isn't good enough. So I wrote this. This was written in a Cloudera CDH5 environment, but with a little tweaking you can get this to work in whatever your environment is (even manually installed ones).
How to use:
NOTE: You can place all of the following in your .profile (or equivalent).
(1) linux$ export MASTER='yarn-client | local[NN] | spark://host:port'
(2) linux$ export SPARK_HOME=/usr/lib/spark # Your's will vary.
(3) linux$ export JAVA_HOME=/usr/java/latest # Your's will vary.
(4) linux$ export NAMENODE='vps00' # Your's will vary.
(5) linux$ export PYSTART=${PYTHONSTARTUP} # See in-line commends about the reason for the need for this alias to PYTHONSTARTUP.
(6) linux$ export HADOOP_CONF_DIR=/etc/hadoop/conf # Your's will vary. This one may not be necessary to set. Try and see.
(7) linux$ export HADOOP_HOME=/usr/lib/hadoop # Your's will vary. This one may not be necessary to set. Try and see.
(8) bpython -i /path/to/script/below # The moment of truth. Note that this is 'bpython' (not just plain 'python', which would not give the code completion you desire).
>>> sc
<pyspark.context.SparkContext object at 0x2798110>
>>>
Now for use with an IDE, you simply determine how to specify the equivalent of a PYTHONSTARTUP script for that IDE, and set that to '/path/to/script/below'. For example, as I described in the in-line comments below, for WING IDE you simply set the key/value pair 'PYTHONSTARTUP=/path/to/script/below' inside the project's properties section.
See in-line comments for more information.
#! /usr/bin/env python
# -*- coding: utf-8 -*-
#
# ===========================================================================
# Author: Noel Milton Vega (PRISMALYTICS, LLC.)
# ===========================================================================
# Start-up script for 'python(1)', 'bpython(1)', and Python IDE iterpreters
# when you want a 'client-mode' SPARK Shell (i.e. interactive SPARK shell)
# environment either LOCALLY, on a SPARK Standalone Cluster, or on SPARK
# YARN cluster. The code-sense/intelligence of bpython(1) and IDEs, in
# particular will aid in learning the SPARK core API.
#
# This script basically (1) first sets up an environment to launch a SPARK
# Shell, then (2) launches the SPARK Shell using the 'shell.py' python script
# provided in the distribution's SPARK_HOME; and finally (3) imports our
# favorite Python modules (for convenience; e.g. numpy, scipy; etc.).
#
# IMPORTANT:
# DON'T RUN THIS SCRIPT DIRECTLY. It is meant to be read in by interpreters
# (similar, in that respect, to a PYTHONSTARTUP script).
#
# Thus, there are two ways to use this file:
# # We can't refer to PYTHONSTARTUP inside this file b/c that causes a recursion loop
# # when calling this from within IDEs. So in step (0) we alias PYTHONSTARTUP to
# # PYSTARTUP at the O/S level, and use that alias here (since no conflict with that).
# (0): user$ export PYSTARTUP=${PYTHONSTARTUP} # We can't use PYTHONSTARTUP in this file
# (1): user$ export MASTER='yarn-client | local[NN] | spark://host:port'
# user$ bpython|python -i /path/to/this/file
#
# (2): From within your favorite IDE, specify it as your python startup
# script. For example, from within a WINGIDE project, set the following
# variables within a WING Project: 'Project -> Project Properties':
# 'PYTHONSTARTUP=/path/to/this/very/file'
# 'MASTER=yarn-client | local[NN] | spark://host:port'
# ===========================================================================
import sys, os, glob, subprocess, random
namenode = os.getenv('NAMENODE')
SPARK_HOME = os.getenv('SPARK_HOME')
# ===========================================================================
# =================================================================================
# This functions emulates the action of "source" or '.' that exists in bash(1),
# and can be used to set PYTHON environment variables (in Pythons globals dict).
# =================================================================================
def source(script, update=True):
proc = subprocess.Popen(". %s; env -0" % script, stdout=subprocess.PIPE, shell=True)
output = proc.communicate()[0]
env = dict((line.split("=", 1) for line in output.split('x00') if line))
if update: os.environ.update(env)
return env
# ================================================================================
# ================================================================================
# Here, we get the name of our current SPARK Assembly JAR file name (locally). We
# use that to create a HDFS URL that points to it's location in HDFS when using
# YARN (i.e. when 'export MASTER=yarn-client'; we ignore it otherwise).
# ================================================================================
# Remember to always upload/update your distribution's current SPARK Assembly JAR
# to HDFS like this:
# $ hdfs dfs -mkdir -p /user/spark/share/lib" # Only necessary to do once!
# $ hdfs dfs -rm "/user/spark/share/lib/spark-assembly-*.jar" # Remove old version.
# $ hdfs dfs -put ${SPARK_HOME}/assembly/lib/spark-assembly-[0-9]*.jar /user/spark/share/lib/
# ================================================================================
SPARK_JAR_LOCATION = glob.glob(SPARK_HOME + '/lib/' + 'spark-assembly-[0-9]*.jar')[0].split("/")[-1]
SPARK_JAR_LOCATION = 'hdfs://' + namenode + ':8020/user/spark/share/lib/' + SPARK_JAR_LOCATION
# ================================================================================
# ================================================================================
# Update Pythons globals environment variable dict with necessary environment
# variables that the SPARK Shell will be looking for. Some we set explicitly via
# an in-line dictionary, as shown below. And the rest are set by 'source'ing the
# global SPARK environment file (although we could have included those explicitly
# here too, if we preferred not to touch that system-wide file -- and leave it as FCS).
# ================================================================================
spark_jar_opt = None
MASTER = os.getenv('MASTER') if os.getenv('MASTER') else 'local[8]'
if MASTER.startswith('yarn-'): spark_jar_opt = ' -Dspark.yarn.jar=' + SPARK_JAR_LOCATION
elif MASTER.startswith('spark://'): pass
else: HADOOP_HOME = ''
# ================================================================================
# ================================================================================
# Build '--driver-java-options' options for spark-shell, pyspark, or spark-submit.
# Many of these are set in '/etc/spark/conf/spark-defaults.conf' (and thus
# commented out here, but left here for reference completeness).
# ================================================================================
# Default UI port is 4040. The next statement allows us to run multiple SPARK shells.
DRIVER_JAVA_OPTIONS = '-Dspark.ui.port=' + str(random.randint(1025, 65535))
DRIVER_JAVA_OPTIONS += spark_jar_opt if spark_jar_opt else ''
# ================================================================================
# ================================================================================
# Build PYSPARK_SUBMIT_ARGS (i.e. the sames ones shown in 'pyspark --help'), and
# apply them to the O/S environment.
# ================================================================================
DRIVER_JAVA_OPTIONS = "'" + DRIVER_JAVA_OPTIONS + "'"
PYSPARK_SUBMIT_ARGS = ' --master ' + MASTER # Remember to set MASTER on UNIX CLI or in the IDE!
PYSPARK_SUBMIT_ARGS += ' --driver-java-options ' + DRIVER_JAVA_OPTIONS # Built above.
# ================================================================================
os.environ.update(source('/etc/spark/conf/spark-env.sh', update = False))
os.environ.update({ 'PYSPARK_SUBMIT_ARGS' : PYSPARK_SUBMIT_ARGS })
# ================================================================================
# ================================================================================
# Next, adjust 'sys.path' so SPARK Shell has the python modules it needs.
# ================================================================================
SPARK_PYTHON_DIR = SPARK_HOME + '/python'
PY4J = glob.glob(SPARK_PYTHON_DIR + '/lib/' + 'py4j-*-src.zip')[0].split("/")[-1]
sys.path = [SPARK_PYTHON_DIR, SPARK_PYTHON_DIR + '/lib/' + PY4J] + sys.path
# ================================================================================
# ================================================================================
# With our environment set, we start the SPARK Shell; and then to that, we add
# our favorite Python imports (e.g. numpy, scipy; etc).
# ================================================================================
print('PYSPARK_SUBMIT_ARGS:' + PYSPARK_SUBMIT_ARGS) # For visual debug.
execfile(SPARK_HOME + '/python/pyspark/shell.py', globals()) # Start the SPARK Shell.
execfile(os.getenv('PYSTARTUP')) # Next, load our favorite Python modules.
# ================================================================================
Enjoy and good luck! =:)
