Uploaded image for project: 'Spark'
  1. Spark
  2. SPARK-16367

Wheelhouse Support for PySpark




      Is it recommended, in order to deploying Scala packages written in Scala, to build big fat jar files. This allows to have all dependencies on one package so the only "cost" is copy time to deploy this file on every Spark Node.

      On the other hand, Python deployment is more difficult once you want to use external packages, and you don't really want to mess with the IT to deploy the packages on the virtualenv of each nodes.

      This ticket proposes to allow users the ability to deploy their job as "Wheels" packages. The Python community is strongly advocating to promote this way of packaging and distributing Python application as a "standard way of deploying Python App". In other word, this is the "Pythonic Way of Deployment".

      Previous approaches
      I based the current proposal over the two following bugs related to this point:

      First part of my proposal was to merge, in order to support wheels install and virtualenv creation

      Virtualenv, wheel support and "Uber Fat Wheelhouse" for PySpark
      In Python, the packaging standard is now the "wheels" file format, which goes further that good old ".egg" files. With a wheel file (".whl"), the package is already prepared for a given architecture. You can have several wheels for a given package version, each specific to an architecture, or environment.

      For example, look at https://pypi.python.org/pypi/numpy all the different version of Wheel available.

      The pip tools knows how to select the right wheel file matching the current system, and how to install this package in a light speed (without compilation). Said otherwise, package that requires compilation of a C module, for instance "numpy", does not compile anything when installing from wheel file.

      pypi.pypthon.org already provided wheels for major python version. It the wheel is not available, pip will compile it from source anyway. Mirroring of Pypi is possible through projects such as http://doc.devpi.net/latest/ (untested) or the Pypi mirror support on Artifactory (tested personnally).

      pip also provides the ability to generate easily all wheels of all packages used for a given project which is inside a "virtualenv". This is called "wheelhouse". You can even don't mess with this compilation and retrieve it directly from pypi.python.org.

      Use Case 1: no internet connectivity
      Here my first proposal for a deployment workflow, in the case where the Spark cluster does not have any internet connectivity or access to a Pypi mirror. In this case the simplest way to deploy a project with several dependencies is to build and then send to complete "wheelhouse":

      • you are writing a PySpark script that increase in term of size and dependencies. Deploying on Spark for example requires to build numpy or Theano and other dependencies
      • to use "Big Fat Wheelhouse" support of Pyspark, you need to turn his script into a standard Python package:
        • write a requirements.txt. I recommend to specify all package version. You can use pip-tools to maintain the requirements.txt
          astroid==1.4.6 # via pylint 
          click==6.6 # via pip-tools 
          colorama==0.3.7 # via pylint 
          enum34==1.1.6 # via hypothesis 
          findspark==1.0.0 # via spark-testing-base 
          first==2.0.1 # via pip-tools 
          hypothesis==3.4.0 # via spark-testing-base 
          lazy-object-proxy==1.2.2 # via astroid 
          linecache2==1.0.0 # via traceback2 
          pep8==1.7.0 # via autopep8 
          py==1.4.31 # via pytest 
          pytest==2.9.2 # via spark-testing-base 
          six==1.10.0 # via astroid, pip-tools, pylint, unittest2 
          traceback2==1.4.0 # via unittest2 
          unittest2==1.1.0 # via spark-testing-base 
          wrapt==1.10.8 # via astroid 
        • write a setup.py with some entry points or package. Use PBR it makes the jobs of maitaining a setup.py files really easy
        • create a virtualenv if not already in one:
          virtualenv env 
        • Work on your environment, define the requirement you need in requirements.txt, do all the pip install you need.
      • create the wheelhouse for your current project
        pip install wheelhouse 
        pip wheel . --wheel-dir wheelhouse 

        This can take some times, but at the end you have all the .whl required for your current system in a directory wheelhouse.

      • zip it into a wheelhouse.zip.

      Note that you can have your own package (for instance 'my_package') be generated into a wheel and so installed by pip automatically.

      Now comes the time to submit the project:

      bin/spark-submit --master master --deploy-mode client --files /path/to/virtualenv/requirements.txt,/path/to/virtualenv/wheelhouse.zip --conf "spark.pyspark.virtualenv.enabled=true" ~/path/to/launcher_script.py 

      You can see that:

      • no extra argument is add in the command line. All configuration goes through --conf argument (this has been directly taken from SPARK-13587). According to the history on spark source code, I guess the goal is to simplify the maintainance of the various command line interface, by avoiding too many specific argument.
      • The wheelhouse deployment is triggered by the --conf "spark.pyspark.virtualenv.enabled=true" }} argument. The {{requirements.txt and wheelhouse.zip are copied through --files. The names of both files can be changed through --conf arguments. I guess with a proper documentation this might not be a problem
      • you still need to define the path to requirement.txt and wheelhouse.zip (they will be automatically copied to each node). This is important since this will allow pip install, running of each node, to pick only the wheels he needs. For example, if you have a package compiled on 32 bits and 64 bits, you will have 2 wheels, and on each node, pip will only select the right one
      • I have choosen to keep the script at the end of the command line, but for me it is just a launcher script, it can only be 4 lines:
        /#!/usr/bin/env python	
        from mypackage import run 
      • on each node, a new virtualenv is created at each deployment. This has a cost, but not so much, since the pip install will only install wheel, no compilation nor internet connection will be required. The command line for installing the wheel on each node will be like:
        pip install --no-index --find-links=/path/to/node/wheelhouse -r requirements.txt 


      • quick installation, since there is no compilation
      • no Internet connectivity support, no need mess with the corporate proxy or require a local mirroring of pypi.
      • package versionning isolation (two spark job can depends on two different version of a given library)


      • creating a virtualenv at each execution takes time, not that much but still it can take some seconds
      • and disk space
      • slighly more complex to setup than sending a simple python script, but this feature is not lost
      • support of heterogenous Spark nodes (ex: 32 bits, 64 bits) is possible but one has to send all wheels flavours and ensure pip is able to install in every environment. The complexity of this task is on the hands of the developer and no more on the IT persons! (TMHO, this is an advantage)

      Use Case 2: the Spark cluster has access to Pypi or a mirror of Pypi

      This is the more elegant situation. The Spark cluster (each node) can install the dependencies of your project independently from the wheels provided by Pypi. Your internal dependencies and your job project can also comes in independent wheel files as well. In this case the workflow is much simpler:

      • Turn your project into a Python module
      • write requirements.txt and setup.py like in Use Case 1
      • create the wheel with pip wheels. But now we will not send ALL the dependencies. Only the one that are not on Pypi (current job project, other internal dependencies, etc).
      • no need to create a wheelhouse. You can still copy the wheels either with --py-files (will be automatically installed) or inside a wheelhouse named wheelhouse.zip

      Deployment becomes:

      Now comes the time to submit the project:

      bin/spark-submit --master master --deploy-mode client --files /path/to/project/requirements.txt --py-files /path/to/project/internal_dependency_1.whl,/path/to/project/internal_dependency_2.whl,/path/to/project/current_project.whl --conf "spark.pyspark.virtualenv.enabled=true" --conf "spark.pyspark.virtualenv.index_url=http://pypi.mycompany.com/" ~/path/to/launcher_script.py 

      or with a wheelhouse that only contains internal dependencies and current project wheels:

      bin/spark-submit --master master --deploy-mode client --files /path/to/project/requirements.txt,/path/to/project/wheelhouse.zip --conf "spark.pyspark.virtualenv.enabled=true" --conf "spark.pyspark.virtualenv.index_url=http://pypi.mycompany.com/" ~/path/to/launcher_script.py 

      or if you want to use the official Pypi or have configured pip.conf to hit the internal pypi mirror (see doc bellow):

      bin/spark-submit --master master --deploy-mode client --files /path/to/project/requirements.txt,/path/to/project/wheelhouse.zip --conf "spark.pyspark.virtualenv.enabled=true" ~/path/to/launcher_script.py 

      On each node, the deployment will be done with a command such as:

      pip install --index-url http://pypi.mycompany.com --find-links=/path/to/node/wheelhouse -r requirements.txt 


      • --conf "spark.pyspark.virtualenv.index_url=http://pypi.mycompany.com/" allows to specify a Pypi mirror, for example a mirror internal to your company network. If not provided, the default Pypi mirror (pypi.python.org) will be requested
      • to send a wheelhouse, use --files. To send individual wheels, use --py-files. With the latter, all wheels will be installed. For multiple architecture cluster, prepare all needed wheels for all architecture and use a wheelhouse archive, this allows pip to choose the right version of the wheel automatically.

      code submission
      I already started working on this point, starting by merging the 2 mergerequests #5408 and #13599
      I'll upload a patch asap for review.
      I see two major interogations:

      • I don't know that much YARN or MESOS, so I might require some help for the final integration
      • documentation should really be carefully crafted so users are not lost in all these concepts

      I really think having this "wheelhouse" support for spark will really helps using, maintaining, and evolving Python scripts on Spark. Python has a rich set of mature libraries Spark should do anythink to help developers easily access and use them in their everyday job.

      Important notes about some complex package such as numpy

      Numpy is the kind of package that take several minutes to deploy and we want to avoid having all nodes install it each time. Pypi provides several precompiled wheel but it may occurs that the wheel are not right for your platform or the platform fo your cluster.

      Wheels are not cached for pip version < 7.0. From pip v7.0 and +, wheel are automatically cached when built (if needed), so the first installation might take some time, but after the installation will be straight forward.

      On most of my machines, numpy is installed without any compilation thanks to wheels


      pip does not use system ssl certificate. If you use a local pypi mirror behind https with internal certificate, you'll have to setup pypi correctly with the following content in ~/.pip/pip.conf:

      cert = /path/to/your/internal/certificates.pem 

      First creation might take some times, but pip will automatically cache the wheel for your system in /.cache/pip/wheels. You can of course recreate the wheel with pip wheel or find the wheel in /.cache/pip/wheels. You can use pip -v install numpy to see where it has placed the wheel in cache.

      If you use Artifactory, you can upload your wheels at a local, central cache that can be shared accross all your slave. See this documentation to see how this works. This way, you can insert wheels in this local cache and it will be seens as if it has been uploaded to the official repository (local cache + remote cache can be "merged" into a virtual repository with artifactory)

      Set use of internal pypi mirror
      Ask your IT to update the ~/.pip/pip.conf of the node to point by default to the internal mirror:

      ; Low timeout 
      timeout = 20 
      index-url = https://&lt;user&gt;:&lt;pass&gt;@pypi.mycompany.org/ 

      Now, no more need to specify the --conf "spark.pyspark.virtualenv.index_url=http://pypi.mycompany.com/" in your Spark submit command line

      Note: this will not work when installing package with python setup.py install syntax. In this case you need to update ~/.pypirc and use the -r argument. This syntax is not used in spark-submit

      Approach vulgarized at the following blog post


        Issue Links



              Unassigned Unassigned
              gaetan@xeberon.net gsemet
              16 Vote for this issue
              29 Start watching this issue



                Time Tracking

                  Original Estimate - 168h
                  Remaining Estimate - 168h
                  Time Spent - Not Specified
                  Not Specified