I'm currently building a docker image that can be used to deploy a deep learning application. The image is fairly large with a size roughly 6GB. Since the deployment time is affected by the size of docker container, I wonder if there are some of best practices to reduce the image size of ml-related applications.

CodePudding user response：

First, keep the data (if any) apart from the image (in volumes for example).Also, use .dockerignore to ignore files you don't want in your image.

Now some techniques:

A first technique is to use multistage builds. For example, an image just to install dependencies and another image that starts from the first one and run the app.

A second technique is to minimize the number of image layers. Each RUN , COPY and FROM command creates a different layer. Try to combine commands in a single one using linux operators (like &&).

A third technique is to take profit of the caching in docker image builds. Run every command you can before copying the actual content into the image. For exemple, for a python app, you might install dependencies before copying the contents of the app inside the image.