Fridolín Pokorný - Thoth - how to recommend the best possible libraries for your application

Having libraries in your Python project properly locked to a specific version is a well known best practice. Dependency management tools in the Python ecosystem lock dependencies to the latest version available, but what if the latest version available is not the best fit for your application? Open source project Thoth is an advanced Python dependency resolver which recommends libraries for your project based on observations that are gathered for Python libraries for specific runtime environments. How these recommendations look like? How are different observations like performance characteristics of machine learning libraries for a particular hardware gathered?

https://us.pycon.org/2019/schedule/presentation/185/

Transcript

1. Thoth
   How to recommend the best possible packages for
   your application
   Fridolin Pokorny
   2019-May-4
   

   View Slide

2. Thoth Station
   ● Fridolín “fridex” Pokorný
   ● Senior Software Engineer at Red Hat
   ● Distributed systems, AI/ML and (of course) Python fan
   ● Projects:
   ○ Reverse engineer RetDec (AVG)
   ○ Linux kernel TLS/DTLS module AF_KTLS
   ○ Selinon - distributed task flows scheduler on top of Celery
   ○ Project Thoth
   $ whoami
   https://fridex.github.io
   

   View Slide

3. Thoth Station
   ● Project Thoth - https://github.com/thoth-station
   ● Red Hat - Office of the CTO
   ○ Emerging technologies
   ○ AI team - https://github.com/aicoe
   ● Initially 2 engineers, now growing
   ○ Christoph Görn
   ○ Francesco Murdaca
   ○ Fridolín Pokorný
   ○ Harshad Reddy Nalla
   ○ Marek Cermak
   ○ Subin Modeel
   $ whoarewe # project Thoth
   https://thoth-station.ninja/
   

   View Slide

4. Thoth Station
   What is Thoth?
   Why Thoth?
   

   View Slide

5. Thoth Station
   Why Thoth?
   ● PyPI - Python Package Index
   ○ https://pypi.org/
   ○ 178,016 projects
   ○ 1,303,926 releases (approx. 7 releases per project)
   

   View Slide

6. Thoth Station
   Why Thoth?
   import tensorflow as tf
   from flask import Flask
   application = Flask()
   

   View Slide

7. Thoth Station
   Why Thoth?
   import tensorflow as tf
   from flask import Flask
   application = Flask()
   

   View Slide

8. Thoth Station
   $ pip3 install --user tensorflow
   $ pip3 install --user flask
   $ python3 ./app.py
   Error: tensorflow 1.10.1 has requirement
   numpy<=1.14.5,>=1.13.3, but you'll have numpy 1.15.1 which
   is incompatible.
   $
   

   View Slide

9. Thoth Station
   Why Thoth?
   import tensorflow as tf
   from flask import Flask
   application = Flask()
   59 releases 28 releases
   

   View Slide

10. Thoth Station
    Why Thoth?
    import tensorflow as tf
    from flask import Flask
    application = Flask()
    59 releases 28 releases
    All combinations how to install libraries directly used:
    59 * 28 = 1,652
    

    View Slide

11. Thoth Station
    Transitive dependencies
    ● Flask
    ○ click, itsdangerous, jinja2, markupsafe, werkzeug
    Estimatimated number of combinations: 54,395,000
    

    View Slide

12. Thoth Station
    Transitive dependencies
    ● TensorFlow
    ○ absl-py, astor, backports-weakref, bleach, enum34, gast, google-pasta, grpcio,
    h5py, html5lib, keras, keras-applications, keras-preprocessing, markdown, mock,
    numpy, pbr, protobuf, pyyaml, scipy, setuptools, six, tensorboard,
    tensorflow-estimator, tensorflow-tensorboard, termcolor, tf-estimator-nightly,
    werkzeug, wheel
    Estimated number of combinations: 139,740,802,927,165,440,000
    approx. 1.39*1020
    

    View Slide

13. Thoth Station
    Why Thoth?
    import tensorflow as tf
    from flask import Flask
    application = Flask()
    1.39*1020
    combinations
    54,395,000
    combinations
    All combinations how to install application stack of libraries
    directly and indirectly used (estimation):
    1.39*1020 * 54,395,000 = 7.6*1027
    

    View Slide

14. Thoth Station
    Why Thoth?
    import pandas as pd
    import tensorflow as tf
    from flask import Flask
    application = Flask()
    Operating System
    Fedora 30
    Fedora 29
    ...
    CentOS 7.6
    CentOS 7.5
    …
    Python interpreter
    

    View Slide

15. Thoth Station
    Why Thoth?
    import pandas as pd
    import tensorflow as tf
    from flask import Flask
    application = Flask()
    Operating System
    Python interpreter
    glibc cuda
    

    View Slide

16. Thoth Station
    Hardware
    Why Thoth?
    import pandas as pd
    import tensorflow as tf
    from flask import Flask
    application = Flask()
    Operating System
    Python interpreter
    glibc cuda
    GPU CPU
    

    View Slide

17. Thoth Station
    Why Thoth?
    Python application
    

    View Slide

18. Thoth Station
    Hardware
    Why Thoth?
    Python application
    Operating System
    Python interpreter
    Native dependecies
    Kernel modules
    Direct Python dependencies
    Transitive Python dependencies
    

    View Slide

19. Thoth Station
    Why Thoth?
    ● Create knowledge base
    ○ What packages in which versions should I use?
    ■ Application builds correctly
    ■ Application runs correctly
    ■ Application behaves and performs well
    ● Create an advanced Python resolver which uses knowledge base to
    resolve software stacks
    Latest versions are not always greatest choices.
    

    View Slide

20. Thoth Station
    Building Thoth’s knowledge base
    

    View Slide

21. Thoth Station
    Gathering data for Thoth’s knowledge base
    ● Resolving software stacks
    ○ own resolution algorithm
    ● Analyses of container images
    ○ JupyterHub images
    ○ Thoth’s container images
    ● Amun, Dependency Monkey
    ○ running CI and perf related tests
    ○ performance related analyses
    ● ...
    

    View Slide

22. Thoth Station
    Optimized TensorFlow builds by Thoth team
    ● Automated tests of libraries
    ● Tests targeting performance
    ● Optimized TensorFlow builds
    https://tensorflow.pypi.thoth-station.ninja/
    

    View Slide

23. Thoth Station
    Recommendations
    

    View Slide

24. Thoth Station
    How good is my software stack?
    simplelib
    anotherlib
    

    View Slide

25. Thoth Station
    

    View Slide

26. Thoth Station
    v1 v2
    simplelib
    v1 v2
    anotherlib
    v1 v2
    dependency1
    v1
    dependency2
    v2
    

    View Slide

27. Thoth Station
    v1 v2
    simplelib
    v1 v2
    anotherlib
    v1 v2
    dependency1
    v1
    dependency2
    v2
    pip/Pipenv (always latest):
    simplelib ==v2
    anotherlib ==v2
    dependency2 ==v2
    

    View Slide

28. Thoth Station
    v1 v2
    simplelib
    v1 v2
    anotherlib
    v1 v2
    dependency1
    v1
    dependency2
    v2
    

    View Slide

29. Thoth Station
    v1 v2
    simplelib
    v1 v2
    anotherlib
    v1 v2
    dependency1
    v1
    dependency2
    v2
    Causes errors based on
    Thoth’s knowledge base.
    

    View Slide

30. Thoth Station
    v1 v2
    simplelib
    v1 v2
    anotherlib
    v1 v2
    dependency1
    v1
    dependency2
    

    View Slide

31. Thoth Station
    v1 v2
    simplelib
    v1 v2
    anotherlib
    v1 v2
    dependency1
    v1
    dependency2
    Simplelib in version v1
    performs better together
    with dependency1 in
    version v1 based on
    Thoth’s knowledge base.
    

    View Slide

32. Thoth Station
    v1
    v2
    simplelib
    v1 v2
    anotherlib
    v1
    v2
    dependency1
    v1
    dependency2
    

    View Slide

33. Thoth Station
    v1
    v2
    simplelib
    v1 v2
    anotherlib
    v1
    v2
    dependency1
    v1
    dependency2
    Thoth (always greatest):
    simplelib ==v1
    anotherlib ==v2
    dependency1 ==v1
    dependency2 ==v1
    

    View Slide

34. Thoth Station
    Stack generation pipeline
    Remove
    pre-releases
    Construct dependency
    graph
    Remove install
    errors
    Remove run
    errors
    Adjust based on
    performance
    Adjust based on
    security
    Sort based on
    semver
    Resolved stacks
    steram
    Performance
    based scoring
    Security based
    scoring
    Lockfile generation
    Final score gating
    Runtime
    environment
    Requirements
    Analysis of
    application
    Lock file
    Justification
    

    View Slide

35. Thoth Station
    Extending information about Python packages
    

    View Slide

36. Thoth Station
    ● License
    ● Classifiers
    ○ Programming Language :: Python :: 3.6
    ○ Programming Language :: Python :: Implementation :: CPython
    ● Package purpose
    ○ machine learning library
    ○ plugin
    ○ …
    ● Is the given package affecting performance?
    Python package metadata
    

    View Slide

37. Thoth Station
    ● A vector space model
    ● Each vector in vector space corresponds to a project
    ● Each item in vector represents a feature
    ● Allows feature based queries and similar projects search
    F = {python, machine-learning, web, django-framework, webassembly, sql, spark, gpu-support, Java}
    F
    tf
    = {1, 1, 0, 0, 0, 0, 0, 1, 0}
    F - feature vector
    F
    tf
    - feature vector for project TensorFlow
    project2vec
    

    View Slide

38. Thoth Station
    Image address
    project2vec
    

    View Slide

39. Thoth Station
    Information about Thoth
    ● Website:
    ○ https://thoth-station.ninja/
    ● Twitter
    ○ https://twitter.com/thothstation
    ■ Follow for updates on public availability
    ● GitHub
    ○ https://github.com/thoth-station
    

    View Slide

40. Thoth Station
    

    View Slide

41. Thoth Station
    ● Community sitting at https://github.com/thoth-station/
    ● Bot Kebechet
    ○ https://thoth-station.ninja/kebechet/
    ● Twitter
    ○ https://twitter.com/thothstation
    $ pip3 install thamos
    $ cd ~/repositories/my-repo/
    $ thamos config
    $ thamos advise
    It’s all on you
    

    View Slide

42. THANK YOU
    plus.google.com/+RedHat
    linkedin.com/company/red-hat
    youtube.com/user/RedHatVideos
    facebook.com/redhatinc
    twitter.com/RedHat
    

    View Slide