The true root of the PVP debate
September 24, 2015
I recently wrote a new Stack feature and blogged about it. The feature is about adding support for PVP bounds in cabal files. I did my very best to avoid stirring up anything controversial. But as is usual when the PVP comes up, a disagreement broke out on Reddit about version bounds. This essentially comes down to two different ways to view an upper bound on a package:
- We've tested, and know for certain that the new version of a dependency is incompatible with our package
- I can't guarantee that any new versions of a dependency will be compatible with our package
If you look through the history of PVP debates, you'll see that this argument comes up over and over again. I'm going to make a bold statement: if a core feature to how we do package management is so easily confused, there's a problem with how we're using the feature. I made an offhand comment about this on Twitter:
Instead of cabal file version ranges, we should have a set of "built with" versions per package. Let tooling generate and interpret the data— Michael Snoyman (@snoyberg) September 24, 2015
Based on the positive feedback to that tweet, I'm moving ahead with making a more official proposal.
How PVP bounds work today
Here's the theory of how you're supposed to write PVP bounds in a file:
- Test your package against a range of dependencies. For example, let's say we tested with text-1.1.2 and text-22.214.171.124
- Modify the build-depends in your .cabal file to say
text >= 1.1.2 && < 1.3, based on the fact that it's known to work with at least version 1.1.2, and unknown to work on anything than major version 1.2.
- Next time you make a release, go through this whole process again.
PVP detractors will respond with a few points:
- You don't know that your package won't work with text-1.1.1
- You don't know that your package won't work with text-1.3
- You don't know for certain that your package will work with text-1.1.3, text-126.96.36.199, or text-1.2.1 (yes, it should based on PVP rules, but mistakes can happen.
- Repeating this testing/updating process manually each time you make a code change is tedious and error-prone.
Collect the real information
If you notice, what we did in the above was extract the cabal file's metadata (version bounds) from what we actually know (known versions that the package works with). I'm going to propose a change: let's capture that real information, instead of the proxy data. The data could go into the cabal file itself, a separate metadata file, or a completely different database. In fact, the data doesn't even need to be provided by the package author. Stackage Nightly, for instance, would be a wonderful source of this information.
A dependency solver - perhaps even cabal-install's - would then be able to extract exactly the same version bound information we have today from this data. We could consider it an intersection between the .cabal file version bounds and the external data. Or, we could ignore .cabal file bounds and jump straight to the database. Or we could even get more adventurous, e.g. preferring known-good build plans (based on build plan history).
In theory, this functionality - if done as a separate database from the .cabal files themselves - means that on-Hackage revisions would be much less important, possibly even a feature that no one needs in the future.
And here's the best part: this doesn't require authors to do anything. We can automate the entire process. There could even be build servers sitting and churning constantly trying to find combinations that build together. We've seen already how difficult it is to get authors to adopt a policy. The best policy is one that can be automated and machine run.
Problems I've thought of so far:
- Some packages (notably Edward Kmett's) have a versioning scheme which expresses more information than the PVP itself, and therefore the generated version bounds from this scheme may be too strict. But that won't necessarily be a problem, since a build server will be able to just test new versions as they come out.
- This very blog post may start a flame war again, which I sincerely hope doesn't happen.
In order for this to really happen, we need:
- Broad support for the idea
- Changes to the cabal-install dependency solver (or an alternate replacement)
- Central infrasturcture for tracking the build successes
- Tooling support for generating the build success information
And to be blunt: this is not a problem that actually affects me right now, or most people I work with (curation is in a sense a simplified version of this). If the response is essentially "don't want it," let's just drop it. But if people relying on version bounds today think that this may be a good path forward, let's pursue it.