What do we mean when we say "work"?
In this post we are going to use the word "work" in a very specific way. A dependency application is said to work if and only if we can take two separate code projects where one uses the other and use them together without needing to write special case code. That is, we should be able to snap the two projects together like Lego. If this can be done to arbitrary projects with a success rate of more than 95%, then the approach can be said to work.
It should be especially noted that "I tried this with two trivial helloworld projects and it worked for me" does not fulfill the requirements of working. Sadly this line of reasoning is used all too often in online dependency discussions, but it is not a response that holds any weight. Any approach that has not been tested with at least tens (preferably hundreds) of packages does not have enough real world usage experience to be taken seriously.
The phases of a project
Every project has three distinctive phases on its way from source code to a final executable.
- Original source in the source directory
- Compiled build artifacts in the build directory
- Installed build artifacts on the system
The dependency classes
Each of these phases has a corresponding way to use dependencies. The first and last ones are simple so let's examine those first.
The first one is the simplest. In a source-only world you just copy the dependency's source inside your own project, rewrite the build definition files and use it as if it was an integral part of your own code base. The monorepos used by Google, Facebook et al are done in this fashion. The main downsides are that importing and updating dependencies is a lot of work.
The third approach is the traditional Linux distro approach. Each project is built in isolation and installed either on the system or in a custom prefix. The dependencies provide a pkg-config file explaining which defines both the dependency and how it should be used. This approach is easy to use and scales really well, the main downside being that you usually end up with multiple versions of some dependency libraries on the same file system, which means that they will eventually get mixed up and crash in spectacular but confusing ways.
The second approach
A common thing people want to do is to mix two different languages and build systems in the same build directory. That is, to build multiple different programming languages with their own build systems intermixed so that one uses the built artifacts of the other directly from the build dir.
This turns out to be much, much, much more difficult than the other two. But why is that?
Approach #3 works because each project is clearly separated and the installed formats are simple, unambiguous and well established. This is not the case for build directories. Most people don't know this, but binaries in build directories are not the same as the installed ones Every build system conjures its own special magic and does things slightly differently. The unwritten contract has been that the build directory is each build system's internal implementation detail. They can do with it whatever they want, just as long as after install they provide the output in the standard form.
Mixing the contents of two build systems' build directories is not something that "just happens". Making one "just call" the other does not work simply because of the N^2 problem. For example, currently you'd probably want to support C and C++ with Autotools, CMake and Meson, D with Dub, Rust with Cargo, Swift with SwiftPM, Java with Maven (?) and C# with MSBuild. That is already up to 8*7 = 56 integrations to write and maintain.
The traditional way out is to define a data interchange protocol to declare build-dir dependencies. This has to be at least as rich in semantics as pkg-config, because that is what it is: a pkg-config for build dirs. In addition to that you need to formalise all the other things about setup and layout that pkg-config gets for free by convention and in addition you need to make every build system adhere to that. This seems like a tall order and no-one's really working on it as far as I know.
What can we do?
If build directories can't be mixed and system installation does not work due to the potential of library mixups, is there anything that we can do? It turns out not only that we can, but that there is already a potential solution (or least an approach for one): Flatpak.
The basic idea behind Flatpak is that it defines a standalone file system for each application that looks like a traditional Linux system's root file system. Dependencies are built and installed there as if one was installing them to the system prefix. What makes this special is that the filesystem separation is enforced by the kernel. Within each application's file system only one version of any library is visible. It is impossible to accidentally use the wrong version. This is what traditional techniques such as rpath and LD_LIBRARY_PATH have always tried to achieve, but have never been able to do reliably. With kernel functionality this becomes possible, even easy.
What sets Flatpak apart from existing app container technologies such as iOS and Android apps, UWP and so on is its practicality. Other techs are all about defining new, incompatible worlds that are extremely limited and invasive (for example spawning new processes is often prohibited). Flatpak is not. It is about making the app environment look as much as possible like the enclosing system. In fact it goes to great lengths to make this work transparently and it succeeds admirably. There is not a single developer on earth who would tolerate doing their own development inside a, say, iOS app. It is just too limited. By contrast developing inside Flatpak is not only possible and convenient, but something people already do today.
The possible solution, then, is to shift the dependency consumption from option 2 to option 3 as much as possible. It has only one real new requirement: each programming language must have a build system agnostic way of providing prebuilt libraries. Preferably this should be pkg-config but any similar neutral format will do. (For those exclaiming "we can't do that, we don't have a stable ABI", do not worry. Within the Flatpak world there is only one toolchain, system changes cause a full rebuild.)
With this the problem is now solved. All one needs to do is to write a Flatpak builder manifest that builds and installs the dependencies in the correct order. In this way we can mix and match languages and build systems in arbitrary combinations and things will just work. We know it will, because the basic approach is basically how Debian, Fedora and all other distros are already put together.