Foundation

A new hope. Foundation is a new library that tries to define a new modern Haskell framework. It is also trying to be more than a library: A common place for the community to improve things and define new things

It started as a thought experiment:

What would a modern Haskell base looks like if I could start from scratch ?

What would I need to complete my Haskell projects without falling into traditional pitfalls like inefficient String, all-in-one Num, un-productive packing and unpacking, etc.

One of the constraints, that was set early on, was not depending on any external packages, instead depending only on what GHC provides (itself, base, and libraries like ghc-prim). While it may sound surprising, especially considering the usually high quality and precision of libraries available on hackage, there are many reasons for not depending on anything; I’ll motivate the reason later in the article, so hang on.

A very interesting article from Stephen Diehl on production, that details well some of the pitfalls of Haskell, or the workaround for less than ideal situation/choice, outline pretty well some of the reasons for this effort.

Starting with the basic

One of the few basic things that you’ll find in any modern haskell project, is ByteArray, Array and packed strings. Usually in the form of the bytestring, vector and text packages.

We decided to start here. One of the common problem of those types is their lack of inter-operability. There’s usually a way to convert one into another, but it’s either exposed in an Unsafe or Internal module, or has a scary name like unsafeFromForeignPtr.

Then, if you’re unlucky you will see some issues with unpinned and pinned (probably in production settings to maximise fun); the common ByteString using the pinned memory, ShortByteString and Text using unpinned memory, and Vector, well, it’s complicated (there’s 4 different kind of Vectors).

Note: pinned and unpinned represent whether the memory is allowed to move by the GC. Unpinned usually is better as it allows the memory system to reduce fragmentation, but pinned memory is crucial for dealing with Input/Output with the real world, large data (and some other uses).

Unboxed Array

Our corner stone is the unboxed array. The unboxed array is a native Haskell ByteArray (represented by the ByteArray# primitive type), and it is allowed to be unpinned or pinned (at allocation time). To also support further interesting stuff, we supplement it with another constructor to make it able to support natively a chunk of memory referenced by a pointer.

In simplified terms it looks like:

data UArray ty = UArrayBA (Offset ty) (Size ty) PinnedStatus ByteArray#
               | UArrayAddr (Offset ty) (Size ty) (Ptr ty)

With this capability, we have the equivalent of ByteString, ShortByteString, Unboxed Vector and (Some) Storable Vector, implemented in one user friendly type. This is a really big win for users, as suddenly all those types play better together; they are all the same thing working the same way.

Instead of differentiating ByteString and Vector, now ByteString disappears completely in favor of just being a UArray Word8. This has been tried before with the current ecosystem with vector-bytestring.

String

String is a big pain point. Base represents it as a list of Char [Char], which as you can imagine is not efficient for most purpose. Text from the popular text package implements a packed version of this, using UTF-16 and unpinned native ByteArray#.

While text is almost a standard in haskell, it’s very likely you’ll need to pack and unpack this representation to interact with base functions, or switch representation often to interact with some libraries.

Note on Unicode: UTF-8 is an encoding format where unicode codepoints are encoded in sequence of 1 to 4 bytes (4 different cases). UTF-16 represent unicode sequences with either 2 or 4 bytes (2 different cases).

Foundation’s String are packed UTF-8 data backed by an unboxed vector of bytes. This means we can offer a lightweight type based on UArray Word8:

newtype String = String (UArray Word8)

So by doing this, we inherit directly all the advantages of our vector types; namely we have a String type that is unpinned or pinned (depending on needs), and supports native pointers. It is extremely lightweight to convert between the two: provided UTF8 binary data, we only validate the data, without re-allocating anything.

There’s no perfect representation of unicode; each representation has it own advantages and disadvantages, and it really depends on what types of data you’re actually processing. One of the easy rules of thumb is that the more your representation has cases, the slower it will be to process the highest unicode sequences.

By extension, it means that choosing a unique representation leads to compromise. In early benchmarks against text we are consistently outperforming Text when the data is predominantly ASCII (i.e. 1-byte encoding). In other type of data, it really depends; sometimes we’re faster still, sometimes slower, and sometimes par.

Caveat emptor: benchmarks are far from reliable, and only been run on 2 machines with similar characteristic so far.

Other Types

We also support already:

  • Boxed Array. This is an array to any other Haskell types. Think of it as array of pointers to another Haskell value
  • Bitmap. 1 bit packed unboxed array

In the short term, we expect to add:

  • tree like structure.
  • hash based structure.

Unified Collection API

Many types of collections support the same kind of operations.

For example, commonly you have very similar functions defined with different types:

    take :: Int -> UArray a -> UArray a
    take :: Int -> [a]      -> [a]
    take :: Int -> String   -> String

    head :: UArray a -> a
    head :: String   -> Char
    head :: [a]      -> a

So we tried to avoid monomorphic versions of common Haskell functions and instead provide type family infused versions of those functions. In foundation we have:

    take :: Int -> collection -> collection

    head :: collection -> Element collection

Note: Element collection is a type family. this allow from a type “collection” to define another type. for example, the Element of a [a] is a, and the Element of String is Char.

Note2: head is not exactly defined this way in foundation: This was the simplest example that show Type families in action and the overloading. foundation’s head is not partial and defined: head :: NonEmpty collection -> Element collection

The consequence is that the same head or take (or other generic functions) works the same way for many different collection types, even when they are monomorphic (e.g. String).

For another good example of this approach being taken, have a look at the mono-traversable package

For other operations that are specific to a data structure, and hard to generalize, we still expose dedicated operations.

The question of dependencies

If you’re not convinced by how we provide a better foundation to the standard Haskell types, then it raises the question: why not depend on those high quality libraries doing the exact same thing ?

Consistency. I think it’s easier to set a common direction, and have a consistent approach when working in a central place together, than having N maintainers working on M packages independently.

Common place. An example speaks better than words sometime: I have this X thing, that depends on the A package, and the B package. Should I add it to A, to B, or create a new C package ?

Atomic development. We don’t have to jump through hoops to improve our types or functions against other part of foundation. Having more things defined in a same place, means we can be more aggressive about improving things faster, while retaining an overall package that make sense.

Versions, and Releases. Far easier to depends on a small set of library than depends on hundreds of different versions. Particularly in an industrial settings, I will be much more confident tracking 1 package, watch 1 issue tracker and deal with a set of known people, than having to deal with N packages, N issues trackers (possibly in different places), and N maintainers.

Some final notes

A fast iterative release schedule. Planning to release early, release often. and with a predictable release schedule.

We’re still in the early stage. While we’re at an exciting place, don’t expect a finish product right now.

You don’t need to be an expert to help. anyone can help us shape foundation.

Join us. If you want to get involved: all Foundation works take place in the open, on the haskell-foundation organisation with code, proposals, issues and voting, questions.


posted by Vincent Hanquez on September 9, 2016.

tags foundation, haskell, text, vector, bytestring.

in haskell.