We have a monolithic build process for a ROS-based application and want to use Docker to package it up so we can test and deploy it more simply.
My current thought is to create an image that has all of the dependencies, then use ENTRYPOINT and appropriate —mount commands to perform our final cmake and build the local source into a container.
However, there is one big design issue for which I have yet to find a solution or best practice. First, I have the OpenCV+CUDA image and an acceptable ROS picture, therefore I need to merge them into a new container, but the FROM command does not appear to be sufficient.
For example, OpenCV has been installed with make install, and CUDA has gone through whatever apt-get processing is required, so I can’t just COPY the files —from one to the other. Currently, when I try to build, CMake complains about a missing CUDA TOOLKIT ROOT DIR, and if I hack that, there will very certainly be another similar issue a few more lines into the makefile.
How can individuals avoid “multiple image inheritance” (for want of a better term) like in this example? Should I just run apt-get cuda8 in my final Dockerfile instead of depending on FROM docker-OpenCV-Cuda to give an environment indirectly? That, however, poses the issue of ensuring that my CUDA versions are in sync between images.