Skip to content

Latest commit

 

History

History
91 lines (67 loc) · 6.21 KB

README.md

File metadata and controls

91 lines (67 loc) · 6.21 KB
Raw

Bokeh docker images

This directory contains different versions of Docker images for use with Bokeh. There are many ways to deploy Bokeh on Docker, each serving a different use-case and also dependent on personal preference. The containers provided here are experimental and supported on a best-effort basis. Feel free to submit a pull requested for new versions! Good starting points are:

Alpine vs Debian (and other Linux distributions)

As explained by Itamar the major difference between Alpine and Debian (or any other Linux distribution), is that Alpine uses a different C library, musl, instead of the more common glibc. As a result, binary wheels won’t work on Alpine Linux, so many packages that 'just work' on other Linux distributions will need to be compiled from scratch. This means longer build times the first time you install a package.

Furthermore, while in theory musl and glibc are mostly compatible, in practice the differences can cause problems. Particularly, when you are using Bokeh in combination with Numpy, Pandas and/or SciPy, you will need to do extra work to compile the C binaries from these libraries from yourself when using Alpine. This is do-able, with benefits on the users side (small image footprint) and the downsides mostly on the developers side (having to compile python packages from scratch).

These differences are shown in the table below, which compares the uncompressed image footprint for Alpine vs Debian slim, using pip vs. conda.

Alpine Debian slim
pip jfloff/alpine-python:3.7-slim: 82 MB python:3.7-slim 143 MB
conda continuumio/miniconda3:alpine from GitHub 163 MB continuumio/miniconda3:debian-slim from GitHub 245 MB

While there is a difference, given that a modest Bokeh app (like demo.bokeh.org) quickly adds up to an image size of around 1 GB (uncompressed), the difference for production systems is around 10-15%. Besides the difference in image footprint between Alpine and Debian, the conda-based images are also larger as they include more libraries. Note that these conda images from ContinuumIO are already optimised as described by Jim Crist.

Building images

To build a docker image, choose to build with Alpine or Debian. All the commands below should be run from the root of bokeh's repository. Note the Alpine images build on frolvlad's Alpine images, as continuumio/miniconda3:alpine are not published on Docker Hub yet.

using Alpine from source

The docker-tools/alpine/Dockerfile-from-source pulls bokeh from github and builds bokeh and bokehjs. Note that this will probably fail for old versions! The BOKEH_VERSION build arg is interpreted as the commit to fetch. To build:

docker build --file docker-tools/alpine/Dockerfile-from-source --build-arg BOKEH_VERSION=master --tag bokeh:latest .

You can test this image by running one of the examples:

docker run -p 5006:5006 -it bokeh:latest bokeh serve /bokeh_examples/app/sliders.py

Note that some examples require additional Python modules and hence will result in an error.

using Alpine with conda

Choose a bokeh version to install from the available versions on anaconda. Run conda search bokeh for the available versions. Then, build the image using, for example, version 1.4.0:

docker build -t bokeh:dev-py3 --build-arg PYTHON=3.6 -f docker-tools/Dockerfile-from-travis .

Also here, you can test the image by docker run -it

docker run -p 5006:5006 -it bokeh:1.4.0 bokeh serve /bokeh_examples/app/sliders.py

For older versions the tornado version might need to be pinned. If so, specify TORNADO_VERSION

sudo docker build --file docker-tools/debian/Dockerfile-from-conda --build-arg BOKEH_VERSION=0.12.7 --build-arg TORNADO_VERSION=4.5 --tag bokeh:0.12.7 docker-tools/alpine

using debian-slim with conda

In this example we use docker-tools/debian/environment.yml to specify the conda environment for the Docker container.

docker build --file docker-tools/debian/Dockerfile --tag demo-bokeh-conda docker-tools/debian

To run the container:

docker run --rm -p 8080:8080 -it demo-bokeh-conda

using debian-slim with pip

Here we use docker-tools/debian/requirements.txt to specify pip packages for the Docker container.

docker build -f docker-tools/debian/Dockerfile-pip --tag demo-bokeh-pip docker-tools/debian

To run:

docker run -it -p 8080:8080 --rm demo-bokeh-pip

development image

TO DO

Deploying on Google App Engine

One way to deploy Bokeh apps on Google App Engine (GAE) is to use GAE Python Flexible Environment. Make sure to read at least the the instructions on custom runtimes. At the time of writing (January 2020), Google App Engine only supports websockets on the Flex environment. Once you have setup the project and gcloud SDK, you can deploying to GAE with

cd docker-tools/debian
gcloud app deploy

Note this is just a bare-bones deployment, using the default Dockerfile with no SSL or any security measures. Using a load-balancer is recommended as described in the GAE documentation.

TO DO: add more details on how to configure GAE more robustly.