laborantin-hs

Laborantin is a Haskell framework for running controlled experiments. It is already quite stable and only few things should change in the near future. Comments and pull requests are warmly welcome.

Features

• Write your experiments in Haskell and benefit from compile-time bug squashing, plenty of libraries.
• Embedded specific language to express experiments in a no-brainer, systematic way.
• Compile to binary with:
  • commands to support experimental to lifecycle (run/analyze/find/delete experiments)
  • free support for bash completion

Install

The easiest way to install Laborantin is to use the package published on hackage.

  cabal update
  cabal install laborantin-hs

Alternatively you can clone this repository with git to get the latest development version.

  git clone https://github.com/lucasdicioccio/laborantin-hs	
  cd laborantin-hs
  cabal update
  cabal sandbox init # only if you want a sandboxed install
  cabal install # you can also use cabal configure && cabal build to just build the repo

Two-minutes tutorial

When using Laborantin the typical workflow is as follows:

1. Write one or multiple scenarios using the DSL, e.g. my-experiment.hs.

2. Compile the application with ghc --make -O2 my-experiment.hs. If you built Laboratin using a sandbox, you need to specify the local package database, e.g.

    ghc -no-user-package-db \
        -package-db /path/to/laboratin/.cabal-sandbox/x86_64-linux-ghc-7.6.3-packages.conf.d \
        --make -O2 my-experiment.hs
   

3. Run experiments with

    ./my-experiment run -m "@sc.param 'some-param' in [42, 'toto'] and @sc.param 'other-param' == 1234"
   

Example, annotated, code is as follows. Inline comments start with "--". Please note that the actual implementation of the executePingCommand is left as an exercise.

{-# LANGUAGE OverloadedStrings #-}

module Main where
-- import the world
import Data.Text (Text)
import Control.Monad.IO.Class (liftIO)
import Laborantin.DSL
import Laborantin.Types
import Laborantin.CLI
import Laborantin.Implementation

-- declare one scenario
ping :: ScenarioDescription EnvIO
ping = scenario "ping" $ do -- start defining a scenario called "ping"
  -- enters the description for this scenario
  describe "ping to a remote server"
  -- declares a first parameter, called "destination".
  -- this parameter has some description for documentation purposes
  -- we should explore two values (strings) by default for this parameter
  parameter "destination" $ do
    describe "a destination server (host or ip)"
    values [str "example.com", str "probecraft.net"]
  -- declares a second parameter, called "packet-size".
  -- we should also explore two values (rational numbers) by default for this
  -- parameter
  parameter "packet-size" $ do
    describe "packet size in bytes"
    values [num 50, num 1500] 

  -- now implement the "run hook", which is the actual code to run
  run $ do
    (StringParam srv) <- param "destination" -- lookup "destination" parameter, it should be a string
    (NumberParam ps) <- param "packet-size" -- lookup "packet-size" parameter, it should be a rational number
    liftIO (executePingCommand srv ps) >>= writeResult "raw-result" -- executes the ping action defined below, and dumps the result into a file called "raw-result"
    where executePingCommand :: Text -> Rational -> IO (Text)
          executePingCommand host packetSize = ...

-- list your scenarios in the defaultMain to get a command-line app
main :: IO ()
main = defaultMain [ping]

At first, it looks like Laborantin requires a lot of boilerplate if your experiment is a one-liner. Nothing is free and this is the small cost you have to pay. However you get a handy command-line tool for this cost.

For instance, you get some documentation command: ./my-experiment describe will output:

# Scenario: ping
    ping to a remote server
    4 parameter combinations by default
## Parameters:
### destination
(destination)
    a destination server (host or ip)
    2 values:
    - StringParam "example.com"
    - StringParam "probecraft.net"

### packet-size
(packet-size)
    packet size in bytes
    2 values:
    - NumberParam (50 % 1)
    - NumberParam (1500 % 1)

You can run experiments with: ./my-experiment run (stripped output).

backend> execution finished

backend> "preparing ping"
backend> "resolving dependencies"
backend> scenario: "ping"
         rundir: results/ping/81e44c78-4fd8-4ab9-8f97-5494dac646a2
         json-params: {"packet-size":{"val":1500.0,"type":"num"},"destination":{"val":"probecraft.net","type":"string"}}

backend> execution finished

Then find where experiments results are located with: ./my-experiment find.

results/ping/2ead949a-ed36-4523-9a9c-7c7e2c22a1b2 ping (Success) {"packet-size":{"val":1500.0,"type":"num"},"destination":{"val":"example.com","type":"string"}}
results/ping/81e44c78-4fd8-4ab9-8f97-5494dac646a2 ping (Success) {"packet-size":{"val":1500.0,"type":"num"},"destination":{"val":"probecraft.net","type":"string"}}
results/ping/866b63e8-d407-442c-bc33-f1fb4e96c2a8 ping (Success) {"packet-size":{"val":50.0,"type":"num"},"destination":{"val":"example.com","type":"string"}}
results/ping/a34826bc-5160-4e12-95cc-12fb5c02fc7b ping (Success) {"packet-size":{"val":50.0,"type":"num"},"destination":{"val":"probecraft.net","type":"string"}}

Discussion

Designing scientific experiments is hard. Scientific experiments often must test an hypothesis such as does parameter X influences the outcome of process A ? Experimenters must write code to run the process A, a task that may be daunting when it comes to setting up machines remotely or calling half-a-dozen of shell scripts to edit configurations such as firewall settings for computer networks. As a result of spending time to prepare and debug the code for process A, little time is left for writing the code around parameter X.

In general, running scientific experiments requires a number of actions such as:

• writing code for the experiments themselves
• preparing the system for conducting experiments
• actually conducting the experiments
• organizing results for the experiments
• documenting the experiments
• analyzing the results of experiments

Analyzing results itself is an experiment because a sound analyses also requires steps such as:

• writing code for the analyses themselves
• actually conducting the analyses
• organizing results for the analyses
• [...]

After each analysis step, the scientist may want to run extra experiments if some questions are not fully answered, or if new questions arise. These extra experiments again ask for the same care in preparing/conducting/analyzing experiments. Laborantin is a framework to help you along this iterative process.

Laborantin is moving away from Ruby and adopted Haskell for two distinct reasons. First, Haskell is a functional programming language and it is easier to reuse chunk of codes in declarative DSLs (Domain Specific Languages) with functional programming languages than scripting languages (although Ruby is great for DSLs too). Second, Haskell has a very powerful type system and it allows to catch a whole set of bugs at compile time. While I may consider the first point on DSLs open for debate, this second point is the nail in the coffin: real-world experiments generally involve a time-consuming phase where we act on the physical world (e.g., sending hundreds of network packets spaced in time). You do not want to lose minutes because a typo crashed your experiment: it is infuriating and stressful. You typically cannot write tests for this type of "effects on the real-world-only" code. Nor it is possible to mock and write unit tests for the whole world when you are under pressure for getting results for your research. Thus, Haskell's opinionated choices to segregate effectful code from pure code and Haskell's obnoxious type system are a time saver in code for running experiments. One drawback of using Haskell is that Laborantin-Hs needs a compilation phase now (i.e., Laborantin-Hs is more a library than a command-line utility). Somehow, I think that the pros far outweigh the cons. Plus, it seems possible to write a labor-like script for Laborantin-Hs that will compile the project or call runhaskell underneath.

Laborantin-Hs brings the following to the experimenter:

• a clean DSL to express scenarios, parameters, as well as raw data an product analysis
• an execution engine that runs scenarios, exhausting the parameter space for you
• a default backend to store executed scenarios data and metadata in the filesystem
• auto-documentation features to later generate simple HTML reports
• the full power of Haskell type-system to catch runtime errors at compile time

With releasing Laborantin for free and under an open-source license, our goal is to make sure that your precious expert time is used in productive efforts. This is not a purely altruistic goal because I want to benefit from your good science asap. Another goal of Laborantin is to empower scientists who want to open their code and datasets more easily than possible nowadays.

Historical anecdote

A pattern I started with, and that I have seen often while observing my peers is to encode parameters in filename such as process_A_parameterX1.dat to store a data result and process_A_parameterX2.dat. For instance ping_grenouille.com_1500.dat. Such a scheme is okay for a first shot but quickly becomes opaque past a few parameters, or when you have special characters in string parameters. Plus, with evolving version of process A into finer and finer refinements a whole genealogy of experiments unrolls and the number of required result files explodes. A similar effect happens as the number of result files needed grows or as the number of parameters increases.

After failing to manage sound experiments with mere bash scripts encoding parameters in filenames, an obvious next step is to use a build system such as make to run experiments. Makefiles help, but they only work for so long. Make is a build system for managing dependencies in a build process. I think you can use make to explore a set of parameters, but it looks totally unnatural to write rules for encoding and decoding parameters in filenames. In the end, your Makefile is barely decipherable and you need the Rosetta stone to remember what %< means because your web search engine will simply ignore the glyph. Thus, documenting the experiments while evolving the set of experiments quickly becomes painful. You can try to work around by combining make with shell scripts and environment variables but the coupling between different files becomes so tight that your experiment project is impossible to maintain and it will fail in absurd and totally obscure manner. Similarly, I let you imagine how annoying it is, when you spent a full day writing and debugging your Makefile and it turns out you need to spend another day to make sure make clean does not wipe all your results because you had an experiment crash.

The only way around managing results for an exploding parameter space is to use a sort of database, and to let the computer manage the database. If you ever start encoding parameters values or experiment names in filenames, you are doing it wrong. It is exactly the same as that using a spoon to drive a screw: it works but it is a lot of effort. Laborantin is out, free, and open-source: use it, fork it, or clone it.

Roadmap

For version 0.1.6.x

For version 0.1.7.x

• [dsl] propagate the whole parameter space to every result-dependency when aggregating results (this way, we can easily do things like labor run -s some-analysis -p some-param-for-a-raw-data-collection:int:42
• [cli] command to plot the dep graph between results/scenarios

For version 0.1.8.x

• [dsl] labor-script binary for easy-integration of other scripts
  • will need to expand parameter spaces by extending undeclared parameters