/bazel-mvn-demo

Comparison of building a Java/C++ application with Maven vs Bazel

Primary LanguageC++Apache License 2.0Apache-2.0

TensorFlow/Java Build demo

about

This project provides a comparison between build systems/techniques. It's a simple Java application that wraps TensorFlow's Java JNI wrapper, which calls a native library. There are also a number of not-terribly-useful generated Java libraries, designed to increase the complexity of the build. The codebase is a disconnected fork of TensorFlow1 1.9, with the demo app added (in folder demo-app), and a few tweaks to TF to facilitate the demo.

This application has been successfully built on Ubuntu 16.04. It should work on other OSes, though native library install may be different.

Setup

  • Required software
    • Combined prereqs: sudo apt-get install pkg-config zip g++ zlib1g-dev unzip python python-dev python-pip openjdk-8-jdk
    • PIP bits: pip install -U --user pip six numpy wheel mock
    • Maven: sudo apt-get install maven
    • Bazel 0.15
      • wget https://github.com/bazelbuild/bazel/releases/download/0.15.2/bazel-0.15.2-installer-linux-x86_64.sh
      • chmod +x bazel-0.15[...]
      • ./bazel-0.15[...].sh --user
  • Setup local machine (at root of repo)
    • Run ./configure to prepare TensorFlow Bazel workspace
      • Note: this might fail b/c the repo is pre-configured with an Ubuntu config. TBH I'm not sure what happens if you try to configure over it.
    • Build the TF native library and Java wrapper:
      • bazel build //tensorflow/java:tensorflow
      • bazel build //tensorflow/java:libtensorflow_jni
      • bazel build //tensorflow/java:pom [??? needed]
    • Configure the local JDK to load the native library:
     echo "export LD_LIBRARY_PATH=\$LD_LIBRARY_PATH:${PWD}/bazel-bin/tensorflow/java" >> ~/.bashrc
 source ~/.bashrc
    • set command prompt to make more room: PS1="\[\033[42m\]\[\033[30m\]${HOSTNAME/bazel-mvn-/}:\[\033[40m\]\[\033[34m\]…/\W\[\033[00m\]\$ "
  • Setup GCP (at root of repo)
    • Authenticate: gcloud init
    • Create a service account and store it as creds/rbe-cred.json
  • Build with Maven and verify that it runs
    • cd demo-app && mvn package
    • ./demo.sh
  • Build with Bazel and verify that it runs
    • bazel test :all
    • bazel run tfjavademo [NOTE: this won't actually work b/c Bazel looks in a different place for the image files. TODO: fix]
    • bazel test :all --config=rbe

Demo script part A

  1. Show app:
  • cd demo-app
  • ./demo.sh
  1. Java change:
  • vim src/main/java/com/davidstanke/tfjavademo/Greeter.java
  1. Test:
  • mvn test / bazel test :all
  • [fail]
  1. Fix:
  • vim src/test/java/com/davidstanke/tfjavademo/GreeterTest.java
  • mvn test / bazel test :all

Demo script part B

  1. C++ change:
  • cd ../tensorflow/core/
  • vim lib/strings/numbers.h
  1. Test:
  • bazel test :all / bazel test :all --config=rbe
  • [fail]
  • [show BBRUI]
  1. Fix:
  • vim lib/stirngs/numbers_test.cc
  • bazel test :all --config=rbe