Module merging changes

      3 min read        ·

We are very close to release CEKit version 3.0. This is a great time to discuss one of the biggest changes we made in the codebase in preparation for 3.0: the module merging process rewrite.

Old style merging

Since early days of CEKit all modules were processed and merged internally. What does it mean? Let’s imagine we have an image descriptor that requires a few modules to install. These modules can define anything what is available in the module descriptor.

CEKit iterated over these modules and tried to build something that we called effective image descriptor. If there were for example labels to add, we had an array of labels which was extended with values found (if any) in next modules. Same applied to environment variables, packages and everything else actually.

Such internal image descriptor was then processed by (a fairly simple) template which produced the Dockerfile used to build the image.

What’s wrong with this approach?

We found that above approach can be tricky to implement correctly.

One of problematic areas is package installation. All packages, from all modules were combined together in a list. Such list of packages was installed in the image in one go, as one of the first steps in the template. What if package installation makes only sense when some requirements are fulfilled? For example when a repository definition is required to be present, which is added by some module?

Environment variables can be also problematic. If we define the same environment variable in multiple modules, only the last module’s value would be added to the container image. In most cases this is fine, but what if we rely on this environment variable at the time of module installation? We expect something that we set in this module, but it was overridden by some other module which leads to an error.

As you can probably already see — there are potentially other cases where such approach can provide unexpected results.

The change

We found that it’s best to look at modules as units of work. Everything that is defined in the module should be applied exactly where it belongs to, as a whole, without the attempt to smartly handle it.

With the new approach we iterate over modules in the defined installation order in the template itself. Modules are applied (processed) serially.

This gives full control over how images are composed to image developer. If there are some prerequisites for a module, these can be satisfied by changes in a previous module on the list to install.

Tradeoffs of new approach

Of course there are always tradeoffs. We gained more control over the content, but the image build process can take more time. This is the result of multiple executions of the same instruction. For example if we define labels in all 10 modules we are going to install, we will run the LABEL instruction 10 times. Same applies to package installation, which highlights another tradeoff: intermediate image can be bigger.

This shouldn’t be a problem in general, because we now squash by default every image that is built with the Docker engine. Other engines may squash images automatically for you. For example Buildah and OSBS do this.

Note
You can disable squashing for Docker engine by specifying th --no-squash flag.

Summary

In many cases you will not notice this change since this is actually an implementation detail and the resulting images should be exactly the same. But in case if you build sophisticated images, it may be important to understand how it works internally to fully leverage CEKit.

Note
You may want to read about module processing in our documentation too!

comments powered by Disqus