Fork me on GitHub

The Future Futhark Package Manager

Posted on July 20, 2018 by Troels Henriksen

There was a time when it was considered crucial for a programming language to come bundled with a large standard library. Python is perhaps the most famous example of this idea of “batteries included”. Unfortunately, this approach ties development of the libraries to the release schedule of the language implementation (and in particular makes it impossible to upgrade one without the other), which often leads to these “standard” libraries falling behind compared to independently developed libraries. Another issue is that the standard libraries typically must be general enough to cover everyones needs, which easily makes them convoluted. For more, see the post Where modules go to die and its corresponding discussion on Hacker News.

Instead of having “batteries included”, many recent languages instead contain minimal standard libraries, and instead provide a standard package manager that makes it easy to use third party packages. Rust is perhaps the most prominent example of this approach. Since reusable Futhark code is beginning to crop up, and I’d rather not stuff all of it into futlib itself, it is time to design and implement a simple package manager for Futhark.

There are lots of things we would like the package manager to do. Unfortunately, because we live in an imperfect world, there are also things we cannot expect it to do.

  1. It must be simple to implement. There are lots of improvements to make to Futhark, and we cannot afford to spend too long on introducing and fixing bugs in a package manager. The best way to avoid complexity is to avoid features, so we will keep the package manager as bare-bones as possible.

  2. It must be simple to use. Futhark lives in the desert, and our users are not interested in learning a new complicated package manager. It is better to restrict the flexibility of the package manager than to introduce complexity in the user experience.

    This also extends to package authors - ideally, issuing a release should be as simple as tagging a revision in a source control system.

  3. No user configuration required. Whatever format a package lists its dependencies in must contain enough information to also allow the package manager to fetch them. If you download a Futhark program, all you should do is run futhark-pkg get (or whatever the command ends up being), and then you can compile it. This is in contrast to a system like Smackage, which requires package URLs to be manually added to a user-global configuration file.

  4. The compiler should not be modified. In particular, the compiler does not support include paths, so the package manager must make the fetched packages available as files in the file system, such that they can be directly imported by the program. Concretely, I think the package manager will simply put everything it fetches in a lib/ sub-directory, such that a program will have to say import "lib/foo/bar" to import file bar.fut from the foo package. The package manager will not be a build system. You will still need to manually invoke the compilers after fetching the dependencies. This creates operational simplicity.

  5. No need for a central server. We do not want more infrastructure maintenance burdens. However, the design should permit us to later add such a server as a package registry, as this seems likely to be useful for documentation and discoverability.

  6. The package manager must fetch only pre-packaged tarballs. Interfacing directly with version control systems is too complex. Fortunately, code hosting sites like GitHub make it trivial to generate tarballs corresponding to revision tags.

  7. It need not be possible for a program to simultaneously depend on multiple versions of the same package. Such support might be difficult to implement, and is anyway only needed for large programs. Futhark programs are supposed to be small, so this is complexity that is not worth it.

  8. “Vendoring” (copying the dependencies directly into your own source repository) must still be possible for those who prefer that. Ideally, just by committing the lib/ directory alluded to in point 4.

  9. Initially focus on supporting a GitHub-centred workflow. This

    is only relevant if it becomes necessary to make a choice about what code hosting service to support initially (if a generic solution is not possible). Ultimately it is not appropriate to depend exclusively on a centralised proprietary service, but we need to start somewhere.

I have never implemented a package manager before, although I have used a few. To figure out how to turn the above wish list into a program, I have been doing some reading. In particular, I have found this series of blog posts by Russ Cox very interesting. Russ is also working on adding a new package manager, called vgo, to the Go language. While I am not a big fan of Go as such, I have a lot of respect for the thoughtfulness that tends to go into the language and its tools. For example, Minimum Version Selection seems like an elegant solution that both manages to avoid complex version resolution, and provides reproducible builds without the need for a lock file - and all by preferring to use the oldest compatible packages rather than the newest.

Let’s look at how vgo addresses the requirements for Futhark. Simplicity of use is certainly the case, as vgo automatically reads import statements in the source code and downloads necessary dependencies before building. It also does not require any user configuration, but does use the $GOPATH directory as sort of a cache. vgo does not require a central server, but inherits a nice scheme from go get whereby any HTML page can act as a sort of package proxy by serving appropriate META tags. Previous package managers for Go interacted directly with the zoo of different version systems, but Russ Cox explains how this is far too much complexity, so vgo fetches only tarballs. In fact, his explanation is why I added this requirement for the Futhark package manager. Multiple versions of the same package are explicitly supported by vgo via semantic import versioning. This makes sense for Go, which must scale to gigantic programs, but less so for Futhark. Finally, vendoring is unaffected, and vgo seems to also initially support GitHub import paths, and uses the GitHub web API to fetch tags and tarballs instead of invoking git directly. Looks like a pretty close match for our needs!

The only missing piece is the question about simplicity of implementation, so I guess the only option is to start writing some code and see what happens.