26 Mar 2015
TL;DR: we propose to factor Hackage into a separate, very simple service serving a stash of Haskell packages, with the rest of Hackage built on top of that, in the name of availability, reliability and extensibility.
One of the main strengths of Haskell is just how much it encourages composable code. As programmers, we are goaded along a good path for composability by the strictures of purity, which forces us to be honest about the side effects we might use, but first and foremost because first class functions and lazy evaluation afford us the freedom to decompose solutions into orthogonal components and recompose them elsewhere. In the words of John Hughes, Haskell provides the necessary glue to build composable programs, ultimately enabling robust code reuse. Perhaps we ought to build our shared community infrastructure along the same principles: freedom to build awesome new services by assembling together existing ones, made possible by the discipline to write these basic building blocks as stateless, scalable, essentially pure services. Let's think about how, taking packages hosting as an example, with a view towards solving three concrete problems:
- Availability of package metadata and source code (currently these are no longer available when hackage.haskell.org goes down).
cabal updatedownload times.
- The difficulty of third party services and other community services to interoperate with hackage.haskell.org and extend it in any direction the community deems fit.
Today Haskell packages are sets of files with a distinguished
*.cabal file containing the package metadata. We host these files on
a central package repository called Hackage, a community supported
service. Hackage is a large service that has by and large served the
community well, and has done so since 2007. The repository has grown
tremendously, by now hosting no less than 5,600 packages. It
implements many features, some of which include package management. In
particular, Hackage allows registered users to:
Upload a new package: either from the browser or via
Download an index of all packages available: this index includes the full content of all
*.cabalfiles for all packages and all versions.
Query the package database via a web interface: from listing all packages available by category, to searching packages by name. Hackage maintains additional metadata for each package not stored in the package itself, such as download counts, package availability in various popular Linux distributions. Perhaps in the future this metadata will also include social features such as number of "stars", à la Github.
Some of the above constitute the defining features of a central package repository. Of course, Hackage is much more than just that today - it is a portal for exploring what packages are out there through a full blown web interface, running nightly builds on all packages to make sure they compile and putting together build reports, generating package API documentation and providing access to the resulting HTML files, maintaining RSS feeds for new package uploads, generating activity graphs, integration with Hoogle and Hayoo, etc.
In the rest of this blog post, we'll explore why it's important to tweeze out the package repository from the rest, and build the Hackage portal on top of that. That is to say, talk separately about Hackage-the-repository and Hackage-the-portal.
A central hub is the cement of the community
A tell-tale sign of a thriving development community is that a number of services pop up independently to address the needs of niche segments of the community or indeed the community as a whole. Over time, these community resources together as a set of resources form an ecosystem, or perhaps even a market, in much the same way that the set of all Haskell packages form an ecosystem. There is no central authority deciding which package ought to be the unique consecrated package for e.g. manipulating filesystem paths: on Hackage today there are at least 5, each exploring different parts of the design space.
However, we do need common infrastructure in place, because we do need consensus about what package names refer to what code and where to find it. People often refer to Hackage as the "wild west" of Haskell, due to its very permissive policies about what content makes it on Hackage. But that's not to say that it's an entirely chaotic free-for-all: package names are unique, only designated maintainers can upload new versions of some given package and version numbers are bound to a specific set of source files and content for all time.
The core value of Hackage-the-repository then, is to establish
consensus about who maintains what package, what versions are
available and the metadata associated with each version. If Alice has
created a package called
foolib, then Bob can't claim
any of his own packages, he must instead choose another name. There is
therefore agreement across the community about what
Agreement makes life much easier for users, tools and developers
talking about these packages.
What doesn't need consensus is anything outside of package metadata and authorization: we may want multiple portals to Haskell code, or indeed have some portals dedicated to particular views (a particular subset of the full package set) of the central repository. For example, stackage.org today is one such portal, dedicated to LTS Haskell and Stackage Nightly, two popular views of consistent package sets maintained by FP Complete. We fully anticipate that others will over time contribute other views — general-purpose or niche (e.g. specialized for a particular web framework) — or indeed alternative portals — ranging from small, fast and stable to experimental and replete with social features aplenty. Think powerful new search functionality, querying reverse dependencies, pre-built package sets for Windows, OS X and Linux, package reviews, package voting ... you name it!
Finally, by carving out the central package repository into its own simple and reliable service, we limit the impact of bugs on both availability and reliably, and thus preserve on one of our most valuable assets: the code that we together as a community have written. Complex solutions invariably affect reliability. Keeping the core infrastructure small and making it easy to build on top is how we manage that.
The next section details one way to carve the central package repository, to illustrate my point. Alternative designs are possible of course - I merely wish to seek agreement that a modular architecture with at its core a set of very small and simple services as our community commons would be beneficial to the community.
Pairing down the central hub to its bare essence
Before we proceed, let's first introduce a little bit of terminology:
- A persistent data structure is an data
structure that is never destructively updated: when modified, all
previous versions of the data structure are still available.
containerspackage is persistent in this sense, as are lists and most other data structures in Haskell.
- A service is stateless if its response is a function of the state of other services and the content of the request. Stateless services are trivially scaled horizontally - limited only by the scalability of the services they depend on.
- A persistent service is a service that maintains its only state as a persistent data structure. Most resources served by a persistent service are immutable. Persistent services share many of the same properties as stateless services: keeping complexity down and scaling them is easy because concurrent access and modification of a persistent data structure requires little to no coordination (think locks, critical sections, etc).
A central hub for all open source Haskell packages might look something like this:
A persistent read-only directory of metadata for all versions of all packages (i.e. the content of the
.cabalfile). Only the upload service may modify this directory, and even then, only in a persistent way.
A persistent read-only directory of the packages themselves, that is to say the content of the archives produced by
An upload service, for uploading new revisions of the metadata directory. This service maintains no state of its own, therefore multiple upload services can be spawned if necessary.
An authentication service, granting access tokens to users for adding a new package or modifying the metadata for their own existing packages via the upload service(s).
The metadata and package directories together form a central
repository of all open source Haskell packages. Just as is the case
with Hackage today, anyone is allowed to upload any package they like
via the upload service. We might call these directories collectively
The Haskell Stash. End-user command-line tools, such as
cabal-install, need only interact with the Stash to get the latest
list of packages and versions. If the upload or authentication
services go down, existing packages can still be downloaded without
Availability is a crucial property for such a core piece of infrastructure: users from around the world rely on it today to locate the dependencies necessary for building and deploying Haskell code. The strategy for maintaining high availability can be worked out independently for each service. A tried and tested approach is to avoid reinventing as much of the wheel as we can, reusing existing protocols and infrastructure where possible. I envision the following implementation:
Serve the metadata directory as a simple Git repository. A Git repository is persistent (objects are immutable and live forever), easy to add new content to, easy to backup, easy to mirror and easy to mine for insights on how packages changes over time. Advanced features such as package candidates fall out nearly for free. Rather than serving in its entirety a whole new static tarball of all package metadata (totalling close to 9MB of compressed data) as we do today, we can leverage the existing Git wire protocol to transfer new versions to end users much more efficiently. In short, a faster
The point here is very much not to use Git as a favoured version control system (VCS), fine as it may be for that purpose, at the expense of any other such tool. Git is at its core an efficient persistent object store first and foremost, with a VCS layered on top. The idea is to not reinvent our own object store. It features a simple disk format that has remained incredibly stable over the years. Hosting all our metadata as a simple Git repository means we can leverage any number of existing Git hosting providers to serve our community content with high uptime guarantees.
Serve package source archives (produced by
cabal sdist) via S3, a de facto standard API for file storage, supported by a large array of cloud providers. These archives can be large, but unlike package metadata, their content is fixed for all time. Uploading a new version of a package means uploading a new source archive with a different name. Serving our package content via a standard API means we can have that content hosted on a reliable cloud platform. In short, better uptime and higher chance that
cabal installwill not randomly fail.
The Haskell Stash is a repository in which to store our community's shared code assets in as simple, highly available and composable a manner as possible. Reduced to its bare essence, easily consumable by all manner of downstream services, most notably, Hackage itself, packdeps.haskellers.org, hdiff.luite.com, stackage.org, etc. It is by enabling people to extend core infrastructure in arbitrary directions that we can hope to build a thriving community that meets not just the needs of those that happened to seed it, but that furthermore embraces new uses, new needs, new people.
Provided community interest in this approach, the next steps would be:
- implement the Haskell Stash;
- implement support for the Haskell Stash in Hackage Server;
- in the interim, if needed, mirror Hackage content in the Haskell Stash.
In the next post in this series, we'll explore ways to apply the same principles of composability to our command-line tooling, in the interest of making our tools more hackable, more powerful and ship with fewer bugs.