1. How do Haskell developers use monads?
A preliminary assessment
Ismael Figueroa

2. This works aims to bridge this gap in empirical knowledge
About this work
Programming with monads is perhaps one of the defining characteristics of the Haskell language
However, despite the prevalence of monads and their use in programming Haskell software…
we are not aware of any empirical studies on how Haskell developers use monads in their software
This works aims to bridge this gap in empirical knowledge
According to our results, the State monad is by far the most used one

3. Monads in a nutshell

4. Monads come from Category Theory (Moggi, ‘91)
Notions of Computation: a value of type A is different from a computation of type T A
Monads provide a mechanism for reasoning about
non-pure programs in the lambda calculus

5. Monads in Functional Programming (Wadler, ‘92)
class Monad m where
return :: a -> m a
(>>=) m a -> (a -> m b) -> m b
Wadler proposes the use of monads for practical functional programming with notions of computation.
In Haskell this is reflected by the Monad typeclass

6. mtl: monad transformers library
Haskell provides a standardized programming interface for monadic programming: the mtl library.
Based on monad transformers (Liang+ ’96), a mechanism for modular composition of monads
Each notion of computation is implemented in several modules, data types and typeclasses

7. mtl Identity Error List Continuations State Reader Writer RWS
pure computations
computations that may fail
may yield multiple non-deterministic values
can be suspended, passed around, and resumed
mtl
mutable access to a collection of values
read-only access to a collection of values
write-only access to a collection of values
combines State, Reader, and Writer
State
Reader
Writer
RWS

8. Monads as a “Design Pattern”
dequeue :: StateT Identity [Int]
dequeue = do
queue <- get
put (tail queue)
return (head queue)

9. Monads as a “Design Pattern”
dequeue :: StateT Identity [Int]
dequeue = do
queue <- get
put (tail queue)
return (head queue)
StateT transformer
Identity
“Monad Stack”
do is a sequence of the monad >>= function

10. Monads as a “Design Pattern”
ErrorT transformer
dequeue :: ErrorT String (StateT Identity [Int])
dequeue = do
queue <- get
put (tail queue)
return (head queue)
StateT transformer
Identity
Code is uniform with respect to the monad stack in use, avoids code repetition and boilerplate

11. Monads as a “Design Pattern”
dequeue :: ErrorT String (StateT Identity [Int])
dequeue = do
queue <- get
if null queue
then throwError "Empty queue“
else do put (tail queue)
return (head queue)
ErrorT transformer
StateT transformer
Identity
Code is uniform with respect to the monad stack in use, avoids code repetition and boilerplate

12. Research Questions
RQ1: How many packages directly depend on the mtl library? What is their distribution with respect to package metadata?
RQ2: What is the usage distribution of mtl’s monads in packages that directly import mtl?
RQ3: What is the situation of alternative implementations to the mtl, and their usage on existing packages?

13. Methodology

14. Methodology / Processing Pipeline
Part 1: Simpler, answers just RQ1
Part 2: Complex, answers RQ2 and RQ3

15. Methodology / Package Index
hackage.haskell.org is the de-facto repository for open-source Haskell software
A tar.gz file with the hierarchical structure of Hackage and the package metadata

16. Methodology / Package Index
Each subfolder is a version
Each package has a folder
Each versión is described by a .cabal file

17. Methodology / .cabal files
key/value metadata

18. Methodology / Initial Package Data
Name
Version
Stability
Categories
Dependencies
Modules provided
Main modules, for executables
Other info for further processing…
This data is enough to answer RQ1
It only requires processing the .cabal files, which is simple with the Cabal API

19. Methodology / Package Archives Processing
Full source code of the package
This is quite a complex step!
Custom analysis program, using haskell-src-exts
Set of modules imported by the package modules

20. Methodology /Packages with monad usage
Initial Package Data, plus:
Set of imported modules
Indicator flags, 0 or 1, for each module that is available in the mtl
With this we can answer RQ2 and RQ3!
But it is complex and costly, we need to download and analyse the source code of each package version

21. Results

22. Results / RQ1
How many packages directly depend on the mtl library? What is their distribution with respect to package metadata?
mtl-packages by stability , threshold at 3%
Stability distribution of relevant categories
mtl-packages by category, threshold at 5%
On the packages analysed, we found 2803 mtl-packages
Some packages had several categories, after considering this multiplicity we got 3670 unique package/category entries

23. Results / RQ2
What is the usage distribution of mtl’s monads in packages that directly import mtl?
600+ packages import mtl but use no monad???
Most packages use between 0 and 3 monads, and only outliers use more than 6…
The State monad is the most used one, followed by the Reader monad.
The Trans typeclass, used to create monad Transformers is also widely used!!

24. Results / RQ3
What is the situation of alternative implementations to the mtl, and their usage on existing packages?
We found 104 packages that do not import mtl, but use one module that matches the name of an mtl module
78 packages use an alternative implementation of mtl
The remaining 26 appear to implement their own customized versión of the monads they use

25. ONLY THE “LATEST” VERSION
Limitations
ONLY THE “LATEST” VERSION
11171 packages
~3GB of data
436 packages with parsing errors
provided modules

26. Summary / Conclusions packages analysed: 11171
mtl-packages: 2803 (3670 single category) alt-mtl packages: 78 (104 with custom impl)
parsed modules:
pkgs with parsing errors: 438
Around 25% of sampled packages import the mtl
The State monad is the most used one
The monad Transformers module is also widely used
Around 1% of sampled packages use alternatives to mtl
We would like to answer qualitative questions:
What are developers actually doing with the monad transformer class?
How and why are developers creating their own monads?
How does the usage of monads evolve over time, versions?