Skip to content

Latest commit

 

History

History
161 lines (122 loc) · 6.06 KB

README.md

File metadata and controls

161 lines (122 loc) · 6.06 KB

How to develop and build Reverb with the Docker containers

Overview

This document covers a couple scenarios:

While there is overlap in the above scenarios, treating them separately seems the most clear at the moment. Before you get started, setup a local variable pointing to your local Reverb GitHub repo.

$ export REVERB_DIR=/path/to/reverb/github/repo

Create a stable Reverb release

There are two steps for building the Reverb package.

  • Build the Docker container with the version of TensorFlow to build Reverb against.
  • Execute the build and declare any specific TensorFlow dependency for the pip install. The dependency is only enforced if the user uses pip install reverb[tensorflow].

Execute from the root of the git repository. The end result will end up in $REVERB_DIR/dist.

⚠️ Bazel is caching the first version of files or a config of the first version of Python used. If building reverb consecutively for different versions of Python --clear_bazel_cache true is needed and included in the commands below. Running the docker once for each python version also works. This issue may resolve itself in the future.

##################################
# Creates the Docker container.
##################################
# Builds the container with Python 3.8, 3.9, 3.10, and 3.11. Set the
# `build-arg tensorflow_pip` to the version of TensorFlow to build against.
# If building against an RC, use == rather than ~= for `tensorflow_pip`.
$ docker build --pull --no-cache \
  --tag tensorflow:reverb_release \
  --build-arg tensorflow_pip=tensorflow~=2.13.0 \
  --build-arg python_version="python3.8 python3.9 python3.10 python3.11" \
  - < "$REVERB_DIR/docker/release.dockerfile"

#################################################
# Builds Reverb against TensorFlow stable or rc.
#################################################

# Builds Reverb against most recent stable release of TensorFlow and
# requires `tensorflow~=2.8.0` if using `pip install reverb[tensorflow]`.
# Packages for Python 3.8, 3.9, 3.10, and 3.11 are created.
$ docker run --rm --mount "type=bind,src=$REVERB_DIR,dst=/tmp/reverb" \
  tensorflow:reverb_release bash oss_build.sh --clean true \
  --clear_bazel_cache true --tf_dep_override "tensorflow~=2.13.0" \
  --release --python "3.8 3.9 3.10 3.11"

# Builds Reverb against an RC of TensorFlow. `>=` and `~=` are not effective
# because pip does not recognize 2.4.0rc0 as greater than 2.3.0. RC builds need
# to have a strict dependency on the RC of TensorFlow used.
$ docker run --rm --mount "type=bind,src=$REVERB_DIR,dst=/tmp/reverb" \
  tensorflow:reverb_release bash oss_build.sh --clean true \
  --clear_bazel_cache true --tf_dep_override "tensorflow==2.13.0rc0" \
  --release --python "3.8 3.9 3.10 3.11"

# Builds a debug version of Reverb. The debug version is not labeled as debug
# as that can result in a user installing both the debug and regular packages
# making it unclear which is installed as they both have the same package
# namespace. The command below puts the .whl files in ./dist/debug/**.
# Debug builds are ~90M compared to normal builds that are closer to 7M.
$ docker run --rm --mount "type=bind,src=$REVERB_DIR,dst=/tmp/reverb" \
  tensorflow:reverb_release bash oss_build.sh --clean true --debug_build true \
  --clear_bazel_cache true --output_dir /tmp/reverb/dist/debug/ \
  --tf_dep_override "tensorflow~=2.13.0" --release --python "3.8 3.9 3.10 3.11"

Develop Reverb inside a Docker container

  1. Build the Docker container. By default the container is setup for python 3.9. Use the python_version arg to configure the container with 3.8 or 3.9.
$ docker build --tag tensorflow:reverb - < "$REVERB_DIR/docker/release.dockerfile"

# Alternatively you can build the container with Python 3.8 support.
$ docker build --tag tensorflow:reverb \
    --build-arg python_version=python3.8 \
    - < "$REVERB_DIR/docker/release.dockerfile"
  1. Run and enter the Docker container.
$ docker run --rm -it \
  --mount "type=bind,src=$REVERB_DIR,dst=/tmp/reverb" \
  --mount "type=bind,src=$HOME/.gitconfig,dst=/etc/gitconfig,ro" \
  --name reverb tensorflow:reverb bash
  1. Compile Reverb.
$ python3.8 configure.py
$ bazel build //reverb/pip_package:build_pip_package
  1. Build the .whl file and output it to /tmp/reverb_build/dist/.
$ ./bazel-bin/reverb/pip_package/build_pip_package \
  --dst /tmp/reverb_build/dist/

$ ./bazel-bin/reverb/pip_package/build_pip_package \
  --release --dst /tmp/reverb_build/dist/
  1. Install the .whl file.
# If multiple versions were built, pass the exact wheel to install rather than
# *.whl.
$ $PYTHON_BIN_PATH -mpip install --upgrade /tmp/reverb_build/dist/*.whl

Builds Tips and Hints

protoc / protobuf version mismatch

There is a check in the Reverb build process that checks if the protoc library in Tensorflow matches what Reverb is using. It throws this error if there is a mismatch: "Please update the PROTOC_VERSION in your WORKSPACE". Here for search reasons :-).

Tensorflow sets its version of Protobuf in this WORKSPACE file. Reverb sets its version of protoc in this WORKSPACE file. The twist is the protobuf people release protoc as 21.9 and the protobuf library as 3.21.0 as seen here.

Until Feb-2023, Tensorflow was using a version of Protobuf that was ~2 years old. In Feb 2023, TF jumped to the latest protobuf 3.21.9.

Note: While I am pretty sure Tensorflow was on 21.9 as of Feb-2023, The checker in Reverb saw it as 21.0. 21.0 worked so I (tobyboyd) did not look into it any farther.