Crit-bit trees for Haskell

This is the first purely functional implementation of crit-bit trees that I'm aware of.

A crit-bit tree is a key/value container that allows efficient lookups and ordered traversal for data that can be represented as a string of bits.

This package exists in part with education in mind:

• The core data structures are simple.

• The core algorithms are easy to grasp.

• I have intentionally structured the source to be easy to follow and extend.

• I've deliberately left the package incomplete. Ever thought to yourself, "I'd write a bit of Haskell if only I had a project to work on"? Well, here's your chance! I will set aside time to review your code and answer what questions I can.

Education aside, crit-bit trees offer some interesting features compared to other key/value container types in Haskell.

• For many operations, they are much faster than Data.Map from the containers package. For instance, lookup is about 3x faster.

• Compared to Data.HashMap, you get about the same lookup performance, but also some features that a hash-based structure can't provide: prefix-based search, efficient neighbour lookup, ordered storage.

Of course crit-bit trees have some downsides, too. For example, building a tree from randomly ordered inputs is somewhat slow, and of course the set of usable key types is small (only types that can be interpreted as bitstrings "for free").

Compared to the most easily findable crit-bit tree code you'll come across that's written in C, the core of this library has a lot less accidental complexity, and so may be easier to understand. It also handles arbitrary binary data that will cause the C library to go wrong.

How to contribute

I've purposely published this package in an incomplete state, and I'd like your help to round it out. In return, you get to learn a little Haskell, have your code reviewed by someone who wants to see you succeed, and contribute to a rather nifty library.

Do you need any prior experience with Haskell to get started? No! All you need is curiosity and the ability to learn from context. Oh, and a github account.

My aim with this library is drop-in API compatibility with the widely used Haskell containers library, which has two happy consequences:

• There are lots of functions to write!

• In almost every case, you'll find a pre-existing function in containers that (from a user's perspective) does exactly what its counterparts in this library ought to do.

Getting started

If you want to contribute or play around, please use the most modern version of the Haskell Platform.

Once you have the Platform installed, there are just a few more steps.

Set up your local database of known open source Haskell packages.

cabal update

Install the latest version of the cabal command, without which you won't be able to build or run benchmarks. You'll also want a sandbox environment. I like cabal-dev, and there are plenty of others.

cabal install cabal-install
cabal install cabal-dev

Both the new cabal command and cabal-dev will install to $HOME/.cabal/bin, so put that directory at the front of your shell's search path before you continue.

Get the critbit source.

git clone git://github.com/bos/critbit

Set up a sandbox.

The first time through, you need to download and install a ton of dependencies, so hang in there.

cd critbit
cabal-dev install \
    --enable-tests \
    --enable-benchmarks \
    --only-dependencies \
    -j

The cabal-dev command is just a sandboxing wrapper around the cabal command. The -j flag above tells cabal to use all of your CPUs, so even the initial build shouldn't take more than a few minutes.

cabal-dev configure \
    --enable-tests \
    --enable-benchmarks
cabal-dev build

Running the test suite

Once you've built the code, you can run the entire test suite in a few seconds.

dist/build/tests/tests +RTS -N

(The +RTS -N above tells GHC's runtime system to use all available cores.)

If you want to explore, the tests program accepts a --help option. Try it out.

Running benchmarks

It is just as easy to benchmark stuff as to test it.

First, you need a dictionary. If your system doesn't have a file named /usr/share/dict/words, you can download a dictionary here.

If you've downloaded a dictionary, tell the benchmark suite where to find it by setting the WORDS environment variable.

export WORDS=/my/path/to/linuxwords

You can then run benchmarks and generate a report. For instance, this runs every benchmark that begins with bytestring/lookup.

dist/build/benchmarks/benchmarks -o lookup.html \
    bytestring/lookup

Open the lookup.html file in your browser. Here's an example of what to expect.

As with tests, run the benchmarks program with --help if you want to do some exploring.

What your code should look like

Okay, so you've bought into this idea, and would like to try writing a patch. How to begin?

I've generally tried to write commits with a view to being readable, so there are examples you can follow.

For instance, here's the commit where I added the keys function. This commit follows a simple pattern:

• Uncomment the export of the function.

• Write the function definition. In this case, the documentation is taken almost verbatim from the corresponding function in the Data.Map module.

• Write a test and make sure it gets run.

• Add an entry to the benchmark suite so it's easy to see how this compares to other key/value map types.

Naturally, you'll follow the prevailing coding and formatting style. If you forget, I'll be sad and offer you only a terse "fix your formatting" review, and then you'll be sad too.

Setting expectations

I have no idea whether this experiment will attract zero contributors or a hundred. If the former, that's too bad, and I'll flesh the library out at my own pace. If the latter, I'll do my best to keep up, and we'll be more systematic if necessary (it would be a shame to see several redundant pull requests implementing the same functions, is what I'm thinking).

But the main point of this is: have fun!