The concept I'll be describing here is strongly related to GPS Haskell, something Mark, Duncan, and I started working on at ICFP. I'll expand on the relation to that project in the questions section below.
There's a very simple, easily understood problem that I'm sure many of us writing software in Haskell have faced: we want to have a fixed API to write code against, but we also want to get bug fixes against our dependencies. (And dare I say it, possibly even some new features.) Currently, there's no easy way to do that. Curated systems like Stackage provide us with fixed API's to code against, but even to get the benefit of just one tiny new bug fix, we currently need to move over over to a brand new snapshot, providing a brand new API, which crucially, compared to the old API may include arbitrary breaking changes. (The same applies to Linux distributions like Debian and Nix, and to the Haskell Platform as well.)
This is a well understood problem in a different context: Linux distributions themselves. What we have today in Stackage is akin to Debian unstable: a rolling release system where you update your entire environment to a new state of being. Each state of being is internally consistent, but may not be compatible with what you've done locally. In the case of Debian, your config files might break, for instance. In the case of Haskell, your program may no longer compile.
By contrast, we have systems like Ubuntu Long Term Support (LTS). In an LTS release, bug fixes are backported to a stable version for an extended period of time. This allows users to have stability without stagnation. Over the next month, a few of us in the community will be working towards the goal of an experimental "LTS Haskell" kind of project, and hope to have it ready to start testing by January. This blog post is intended to describe how this will work, and encourage people to both provide feedback to improve the system, and to get involved in the project.
On January 1, 2015, we're going to take the most recent Stackage unstable snapshot and promote it to be "LTS Haskell 1.0". It will have its own URL on stackage.org, and will be tracked in a Github repo. On a regular basis (tentatively: once a week), we'll take all of the packages in this snapshot and bump them to the newest version from Hackage with the same major version number (see "example bump" below). We'll then run the full Stackage build/test procedure on this new set of package versions. Once this passes, we'll release this as "LTS Haskell 1.1". That's the whole process.
The significance of this is that every release in the 1.X series will have a backwards-compatible API with previous releases, in the same sense that we're used to with minor version bumps. That means that, barring issues of name collisions, your code will continue to compile with new releases. However, you will also get new features rolled out in minor version bumps of your dependencies and, more importantly, bug fixes that have been released.
After a certain period of time (tentatively: three months, see questions below), we'll again take the newest unstable Stackage snapshot and call that LTS Haskell 2.0. There will be an overlap period where both LTS Haskell 1 and 2 are supported (tentatively: one month), and then LTS Haskell 1 will be retired in favor of LTS Haskell 2. This will give users a chance to upgrade to a new supported release. Note that even after being retired, the old snapshots will still be available for use, the only question is whether bugfixes will still be backported.
To clarify the bump procedure, consider the following fictitious set of packages in LTS Haskell 1.0:
After 1.0 is released, the following releases are made to Hackage:
In our bumping procedure, we would replace foo-2.4.1 with foo-220.127.116.11, since it has the same major version number. Similarly, bar-3.2.2 would be bumped to 3.2.3. However, baz-5.1.9 would not be bumped to baz-5.2.0, since that introduces a breaking API change. (baz's author, however, would be able to make a baz-18.104.22.168 or baz-5.1.10 release, and those would be included in the next bump.)
There are two primary design goals underlying this simple process.
We want the smallest change possible for users, and the smallest
amount of work to be created for library authors. To use LTS Haskell, you would
just modify your
remote-repo, like you do today to use Stackage. (And
hopefully in the future, even that will be simplified, once changes are adopted
by the Haskell Platform and Hackage.) Library authors already release their code to
Hackage with bugfixes. Instead of making them go through a process to get their
changes adopted, we will automatically include them.
We want to make the process as automatic as possible. The process listed above allows a new LTS Haskell candidate to be produced with zero human intervention (though some massaging may be necessary for funny situations on Hackage, see questions section below). Making the process automatic makes it that much easier to provide regular releases.
Note that these design goals are built around what's made Stackage such a successful project: minimal author dependencies, simple user story, and automation. I believe we can recreate that success, with even greater benefit now.
There is one change that library authors can make that would improve the experience: support the current LTS Haskell major version of your packages, and provide bug fixes for them. That means that, if you're the maintainer of foo, LTS Haskell has foo-1.2.1, you've release foo-1.3.0, and a bug is discovered in both the 1.2 and 1.3 versions, please fix the bug with both a foo-22.214.171.124 and foo-126.96.36.199 release. This not only helps LTS Haskell users, but library users in general looking to avoid unnecessary API changes.
This sounds a lot like GPS Haskell. What's the difference? It should sound very similar; the goal of this project is to be a testing ground for what GPS Haskell will become. GPS Haskell involves multiple moving parts: Stackage, the Haskell Platform, and Hackage. It's difficult to coordinate work among all those parts and keep stability. Stackage is well set up to support experiments like this by having the multiple snapshot concept built in. The goal is to try out this idea, shake out the flaws, and hopefully when we've stabilized it, the Haskell Platform and Hackage will be ready to adopt it.
Ubuntu LTS doesn't allow new features to be added. Why are you allowing new features in addition to bugfixes? I'll admit that I originally argued against adding new features, while Mark was in favor of it. Ultimately, I changed my mind for two reasons: I saw people asking for this exact thing to be present in Stackage, and allowing backporting of new features eases the maintenance burden of library authors considerably, which is an incredible selling point. If there's demand in the future for a bugfix-only version of this, I'd be very much open to exploring the idea. But I think it's pragmatic to start initial investigation with this.
Is LTS Haskell a separate project from Stackage? I'd describe it more as an extension of Stackage, with the goal to ultimately expand to multiple projects: Stackage, the Haskell Platform, and Hackage. Said another way: on a code level, this is clearly an extension of the Stackage tooling. But ideologically, it's trying to adopt the best features of all three of those projects in a way that all of them will be able to take advantage.
Why such short support windows? The strawmen of three months between releases and a one month grace period are ridiculously short support windows. The reason I propose them is because- like I mentioned in the design goals- we want the smallest delta from how people work today. Right now, there is no concept of stable versions, and we're trying to introduce it. Starting off with a more standard support window of something like two years would be a radical shift in how library users and authors operate today. Three months is very short, but it's long enough for us to test the process. As time goes on, we should have serious community discussions on how long a support window we should have. (I, for one, am fully in favor of extending it well beyond three months.)
What kind of funny Hackage situations do you mean above? I mentioned above
that manual intervention may sometimes be necessary. Consider the following
situation: foo-1.1 depends on bar-1.1, and both are included in LTS Haskell
1.0. bar-1.2 is then released, which by the rules stated above, will not be
included in LTS Haskell 1.1. foo-1.1.1 is also released, which should be
included. However, suppose that foo-1.1.1 has a lower bound
bar >= 1.2. Even
though foo itself isn't changing its API, it's demanding an API change for
another package. In this case, we'd have to disallow foo-1.1.1 from being
included in LTS Haskell 1.1. I'm not sure if we'll be able to automate this
kind of detection.
As a side note, I've long considered this a shortcoming of the Package Versioning Policy's stance on when version bumps are required, and have debated proposing a change. I'm still debating that proposal, but wouldn't object if someone else wants to make that proposal instead.
How does this affect Linux distributions? It doesn't necessarily affect them at all. However, LTS Haskell could be a very interesting set of packages for Linux distros to track, for all the same reasons given above regarding backported bugfixes. In this sense, you can think of LTS Haskell as having multiple delivery mechanisms. We're experimenting with one delivery mechanism via stackage.org now; we can have future delivery mechanisms via Debian, Fedora, Nix, and even with direct support in Hackage/cabal-install.
What can I do to help? The areas that jump to mind are:
Do you have any more details? I originally wrote a two-part blog post with much more detail, but got feedback that the content was a bit too dense, so I rewrote the content in the format here. There's still lots of information present in those blog posts that may be of interest to some, so I've posted them as a Github Gist in case they're useful to anyone. (Note: I'm not aware of contradictions between that Gist and this post. If there are contradictions, this post takes precedence.)
What does this blog post have to do with the recent Stackage survey? Nothing directly, yet. The survey is intended to help gather more information about how people are using Stackage, and to help us make more informed decisions with future Stackage and LTS Haskell work. This blog post was written before I posted the survey, and has not incorporated the survey results in any way. Stay tuned for more information about those results in a separate blog post.