keskiviikko 15. heinäkuuta 2015

Strange requirements of low level packages

As some of you might know, I'm the main developer of the Meson build system. It is implemented in Python 3 but has strict requirements on dependencies. Every now and then this confuses people so here's an explanation why we do things slightly unconventionally.

One of the greatest things about modern development environments is the number of third party packages and how easy they are to use. This is why Meson has spent a lot of effort in creating a simple way to get dependencies automatically on multiple platforms. This system is called Wrap, but this text is not about that. Instead it's about the lack of dependencies in Meson itself. The policy is that in order to run Meson and its test suite, you are only allowed to depend on Python standard library. No external dependencies are allowed. For non-core stuff (coverage, static analysis) external dependencies are ok.

This seems a bit strange and has caused us to have to reinvent some wheels. There are valid reasons for this, though, which have to do with the fact that packages at the lowest levels of the software stack have requirements that regular applications don't. The most important of these is that a build system needs to work in very limited distros and environments.

Let's look at Mer project for an example. It aims to provide an extremely slimmed down Linux distribution for use in mobile applications. They support only a small number of packages by design. There are only 8 Python packages in total. Six are extension packages and those are not useful for a build system. If we wanted to get Meson there, what would it actually mean?

If Meson depends on any other package, it can not be accepted into Mer. Since source bundling is prohibited and the system does not allow net access during package creation, it just can't be done. Either you disable the part that requires the extension or you don't get your package accepted. You might get an exception but you need to apply for it and it takes time an effort for many people. If only the unit test system depends on external packages you can disable it but then you can't tell if your project is working correctly. This is not acceptable either.

There are a bunch of distros like this. Then there are the BSDs that might not have your dependencies in their repos. Rather than force them all to deal with this dependency issue we instead write our code so that it has no external dependencies. This means more work for us but less for everyone else. Luckily Python 3's standard library is very good and has almost everything we need.

Another use case that many people don't think about is that a lot of development is done on machines that neither have Internet access nor the capability to install additional software. Several corporations have requirements such as these for real or imagined security reasons. Even some OSS development is done like this. Requesting permission to use a new dependency can take six weeks.

I'm not saying that this is right or wrong or even sane. What I'm saying is that this is reality and that you ignore reality at your own peril.

In an environment like this getting Meson approved for use takes only one round of bureaucracy. Every dependency requires an additional round (if you are lucky). To reduce friction and ease uptake, we must again cut down on our dependencies. The easier Meson is to deploy, the more people are willing to switch to it.

There are other advantages, too. It is easier to get contributions from Windows and OSX developers since they can hack on the code without any additional steps. You'd be surprised how tough it can be to download Python dependencies, even with setuptools. Deployment is easier, too: "Download Python 3 and Meson, run" as opposed to "download Python 3 and Meson, use setuptool to move it places, then install these dependencies and then set up these paths and ...".

Because of these reasons we don't use external dependencies. This is the price you pay for being at the very bottom of the software stack.

Ei kommentteja:

Lähetä kommentti