Skip to content

Latest commit

 

History

History
1817 lines (1349 loc) · 88.7 KB

sotu.md

File metadata and controls

1817 lines (1349 loc) · 88.7 KB

State of the Haskell ecosystem

In this post I will describe the current state of the Haskell ecosystem to the best of my knowledge and its suitability for various programming domains and tasks. The purpose of this post is to discuss both the good and the bad by advertising where Haskell shines while highlighting where I believe there is room for improvement.

This post is grouped into two sections: the first section covers Haskell's suitability for particular programming application domains (i.e. servers, games, or data science) and the second section covers Haskell's suitability for common general-purpose programming needs (such as testing, IDEs, or concurrency).

The topics are roughly sorted from greatest strengths to greatest weaknesses. Each programming area will also be summarized by a single rating of either:

  • Best in class: the best experience in any language
  • Mature: suitable for most programmers
  • Immature: only acceptable for early-adopters
  • Bad: pretty unusable

The more positive the rating the more I will support the rating with success stories in the wild. The more negative the rating the more I will offer constructive advice for how to improve things.

Disclaimer #1: I obviously don't know everything about the Haskell ecosystem, so whenever I am unsure I will make a ballpark guess and clearly state my uncertainty in order to solicit opinions from others who have more experience. I keep tabs on the Haskell ecosystem pretty well, but even this post is stretching my knowledge. If you believe any of my ratings are incorrect, I am more than happy to accept corrections (both upwards and downwards)

Disclaimer #2: There are some "Educational resource" sections below which are remarkably devoid of books, since I am not as familiar with textbook-related resources. If you have suggestions for textbooks to add, please let me know.

Disclaimer #3: I am very obviously a Haskell fanboy if you haven't guessed from the name of my blog and I am also an author of several libraries mentioned below, so I'm highly biased. I've made a sincere effort to honestly appraise the language, but please challenge my ratings if you believe that my bias is blinding me! I've also clearly marked Haskell sales pitches as "Propaganda" in my external link sections. :)

Disclaimer #4: I've contributed the majority of these recommendations and I also play an editorial role. This means that although some contributions have been crowd-sourced I reserve the right to decline a pull request or edit/delete content if I feel that a resource is abandoned or if I feel there are better alternatives already listed. I try to be as fair as possible and if you disagree with any decision of mine or you feel that my recommendation does not reflect the consensus of the Haskell community you can challenge my decision by opening an issue and I will either defend my decision or change my mind.

Table of Contents

Legend:

🏆 = Best in class

🥈 = Mature

🌱 = Immature

⛔ = Bad

Application Domains

Compilers

Rating: Best in class

Haskell is an amazing language for writing your own compiler. If you are writing a compiler in another language you should genuinely consider switching.

Haskell originated in academia, and most languages of academic origin (such as the ML family of languages) excel at compiler-related tasks for obvious reasons. As a result the language has a rich ecosystem of libraries dedicated to compiler-related tasks, such as parsing, pretty-printing, unification, bound variables, syntax tree manipulations, and optimization.

Anybody who has ever written a compiler knows how difficult they are to implement because by necessity they manipulate very weakly typed data structures (trees and maps of strings and integers). Consequently, there is a huge margin for error in everything a compiler does, from type-checking to optimization, to code generation. Haskell knocks this out of the park, though, with a really powerful type system with many extensions that can eliminate large classes of errors at compile time.

I also believe that there are many excellent educational resources for compiler writers, both papers and books. I'm not the best person to summarize all the educational resources available, but the ones that I have read have been very high quality.

Finally, there are a large number of parsers and pretty-printers for other languages which you can use to write compilers to or from these languages.

Notable libraries:

Some compilers written in Haskell:

  • Elm
  • Purescript
  • Idris
  • Agda
  • Pugs (the first Perl 6 implementation)
  • ghc (self-hosting)
  • frege (very similar to Haskell, also self-hosting)
  • hython (a Python3 interpreter written in Haskell)
  • Lasca (a small Scala-like language with global type inference and optional dynamic mode on LLVM backend)
  • verve - Functional language with object-oriented support
  • sixten - Haskell/Idris-style language with a focus on precise and efficient memory layout
  • carp - An efficient, statically typed Lisp with ownership tracking.
  • unison - A purely functional distributed programming language with algebraic effects.
  • oden (no longer in active development)

Educational resources:

Propaganda:


Server-side web programming

Rating: Mature

Haskell's second biggest strength is the back-end, both for web applications and services. The main features that the language brings to the table are:

  • Server stability
  • Performance
  • Ease of concurrent programming
  • Excellent support for web standards

The strong type system and polished runtime greatly improve server stability and simplify maintenance. This is the greatest differentiator of Haskell from other backend languages, because it significantly reduces the total-cost-of-ownership. You should expect that you can maintain Haskell-based services with significantly fewer programmers than other languages, even when compared to other statically typed languages.

However, the greatest weakness of server stability is space leaks. The most common solution that I know of is to use ekg (a process monitor) to examine a server's memory stability before deploying to production. The second most common solution is to learn to detect and prevent space leaks with experience, which is not as hard as people think.

Haskell's performance is excellent and currently comparable to Java. Both languages give roughly the same performance in beginner or expert hands, although for different reasons.

Where Haskell shines in usability is the runtime support for the following three features:

  • software transactional memory (which differentiates Haskell from Go)
  • lightweight threads that use non-blocking I/O (which differentiates Haskell from the JVM)
  • garbage collection (which differentiates Haskell from Rust)

If you have never tried out Haskell's software transactional memory you should really, really, really give it a try, since it eliminates a large number of concurrency logic bugs. STM is far and away the most underestimated feature of the Haskell runtime.

Notable libraries:

  • warp / wai - the low-level server and API that all server libraries share, with the exception of snap
  • scotty - A beginner-friendly server framework analogous to Ruby's Sinatra
  • spock - Lighter than the "enterprise" frameworks, but more featureful than scotty (type-safe routing, sessions, conn pooling, csrf protection, authentication, etc)
  • yesod / yesod-* / snap / snap-* / happstack-server / happstack-* - "Enterprise" server frameworks with all the bells and whistles
  • ihp - batteries-included web framework with a friendly and helpful community. The best choice when getting started with haskell.
  • servant / servant-* - Library for type-safe REST servers and clients that might blow your mind
  • graphql-api - Implement a GraphQL API
  • websockets - Standalone websockets client and server
  • authenticate / authenticate-* - Shared authentication libraries
  • ekg / ekg-* - Haskell service monitoring
  • stm - Software-transactional memory
  • lucid - Haskell DSL for building HTML
  • mustache / karver - Templating libraries
  • aeson - Parsing and generation of JSON

Some web sites,services, and projects powered by Haskell:

Propaganda:

Educational resources:

Notable hosting platforms:


Scripting / Command-line applications

Rating: Mature

Haskell's biggest advantage as a scripting language is that Haskell is the most widely adopted language that supports global type inference. Many languages support local type inference (such as Rust, Go, Java, C#), which means that function argument types and interfaces must be declared but everything else can be inferred. In Haskell, you can omit everything: all types and interfaces are completely inferred by the compiler (with some caveats, but they are minor).

Global type inference gives Haskell the feel of a scripting language while still providing static assurances of safety. Script type safety matters in particular for enterprise environments where glue scripts running with elevated privileges are one of the weakest points in these software architectures.

The second benefit of Haskell's type safety is ease of script maintenance. Many scripts grow out of control as they accrete arcane requirements and once they begin to exceed 1000 LOC they become difficult to maintain in a dynamically typed language. People rarely budget sufficient time to create a sufficiently extensive test suite that exercises every code path for each and every one of their scripts. Having a strong type system is like getting a large number of auto-generated tests for free that exercise all script code paths. Moreover, the type system is more resilient to refactoring than a test suite.

However, the main reason I mark Haskell as mature because the language is also usable even for simple one-off disposable scripts. These Haskell scripts are comparable in size and simplicity to their equivalent Bash or Python scripts. This lets you easily start small and finish big.

Haskell has one advantage over many dynamic scripting languages, which is that Haskell can be compiled into a native and statically linked binary for distribution to others.

Haskell's scripting libraries are feature complete and provide all the niceties that you would expect from scripting in Python or Ruby, including features such as:

  • rich suite of Unix-like utilities
  • advanced sub-process management
  • POSIX support
  • light-weight idioms for exception safety and automatic resource disposal

Notable libraries:

Some command-line tools written in Haskell:

Educational resources:


Data science

Rating: Immature

Haskell data science can take advantage of other data science ecosystems via the HaskellR and Sparkle projects. HaskellR is a Haskell-to-R bridge with Jupyter notebook integration, which lets you take advantage of the broad R ecosystem while benefiting from the speed and type safety of Haskell. Sparkle is a Haskell-to-Spark bridge which lets you interface with the Spark subset of the Java/Scala data science ecosystem. However, to get a Mature rating Haskell data science needs to be able to stand alone without depending on other programming language ecosystems.

If you restrict yourself to just the Haskell ecosystem then choices are more limited. I'll primarily compare Haskell to Python since that's the data science ecosystem that I'm more familiar with. Specifically, I'll compare to the scipy suite of libraries:

The Haskell analog of NumPy is the hmatrix library, which provides Haskell bindings to BLAS, LAPACK. hmatrix's main limitation is that the API is a bit clunky, but all the tools are there.

Haskell's charting story is okay. Probably my main criticism of most charting APIs is that their APIs tend to be large, the types are a bit complex, and they have a very large number of dependencies.

Fortunately, Haskell does integrate into IPython so you can use Haskell within an IPython shell or an online notebook. For example, there is an online "IHaskell" notebook that you can use right now located here:

If you want to learn more about how to setup your own IHaskell notebook, visit this project:

The closest thing to Python's pandas is the frames library. I haven't used it that much personally so I won't comment on it much other than to link to some tutorials in the Educational Resources section.

I'm not aware of a Haskell analog to SciPy (the library) or sympy. If you know of an equivalent Haskell library then let me know.

One Haskell library that deserves honorable mention here is the diagrams library which lets you produce complex data visualizations very easily if you want something a little bit fancier than a chart. Check out the diagrams project if you have time:

Areas for improvement:

  • Smooth user experience and integration across all of these libraries
  • Simple types and APIs. The data science programmers I know dislike overly complex or verbose APIs
  • Beautiful data visualizations with very little investment

Notable libraries:


Numerical programming

Rating: Immature

Haskell's numerical programming story is not ready, but steadily improving.

My main experience in this area was from a few years ago doing numerical programming for bioinformatics that involved a lot of vector and matrix manipulation and my rating is largely colored by that experience.

The biggest issues that the ecosystem faces are:

  • Really clunky matrix library APIs
  • Fickle rewrite-rule-based optimizations

When the optimizations work they are amazing and produce code competitive with C. However, small changes to your code can cause the optimizations to suddenly not trigger and then performance drops off a cliff.

There is one Haskell library that avoids this problem entirely which I believe holds a lot of promise: accelerate generates LLVM and CUDA code at runtime and does not rely on Haskell's optimizer for code generation, which side-steps the problem. accelerate has a large set of supported algorithms that you can find by just checking the library's reverse dependencies:

However, I don't have enough experience with accelerate or enough familiarity with numerical programming success stories in Haskell to vouch for this just yet. If somebody has more experience than me in this regard and can provide evidence that the ecosystem is mature then I might consider revising my rating upward.

Notable libraries:

Propaganda:

Educational Resources:


Front-end web programming

Rating: Immature

This boils down to Haskell's ability to compile to JavaScript. ghcjs is the front-runner, but for a while setting up ghcjs was non-trivial. Now that the stack build tool supports ghcjs you can very easily set up a new ghcjs project by following these instructions:

One of the distinctive features of ghcjs compared to other competing Haskell-to-JavaScript compilers is that a huge number of Haskell libraries work out of the box with ghcjs because it supports most of ghc's primitive operations.

I would also like to mention that there are two Haskell-like languages that you should also try out for front-end programming: elm and purescript. These are both used in production today and have equally active maintainers and communities of their own. purescript in particular is extremely similar to Haskell.

Areas for improvement:

  • There needs to be a clear story for smooth integration with existing JavaScript projects
  • There need to be many more educational resources targeted at non-experts explaining how to translate existing front-end programming idioms to Haskell
  • There need to be several well-maintained and polished Haskell libraries for front-end programming
  • The whole ghcjs ecosystem needs much more documentation. There's not even a basic tutorial on how to actually use ghcjs

Notable Haskell-to-JavaScript compilers:

Notable libraries:


Distributed programming

Rating: Immature

This is sort of a broad area since I'm using this topic to refer to both distributed computation (for analytics) and distributed service architectures. For distributed service architectures Haskell is catching up to its peers with service toolkit libraries, but for distributed computation Haskell still lags behind.

There has been a lot of work in replicating Erlang-like functionality in Haskell through the Cloud Haskell project, not just in creating the low-level primitives for code distribution / networking / transport, but also in assembling a Haskell analog of Erlang's OTP. Work on the higher-level libraries seems to have stopped, but the low-level libraries are still good for distributing computation.

Areas for improvement:

  • We need more analytics libraries. Haskell has no analog of scalding or spark. The most we have is just a Haskell wrapper around hadoop
  • We need a polished consensus library (i.e. a high quality Raft implementation in Haskell)

Notable libraries:


Standalone GUI applications

Rating: Immature

All Haskell GUI libraries are wrappers around toolkits written in other languages (such as GTK+ or Qt). The last time I checked the gtk bindings were the most comprehensive, best maintained, and had the best documentation.

The reason for the "Immature" rating is that there still isn't a Haskell binding to a widget toolkit that doesn't have some sort of setup issues with the toolkit.

However, the Haskell bindings to GTK+ have a strongly imperative feel to them. The way you do everything is communicating between callbacks by mutating IORefs. Also, you can't take extensive advantage of Haskell's awesome threading features because the GTK+ runtime is picky about what needs to happen on certain threads. I haven't really seen a Haskell library that takes this imperative GTK+ interface and wraps it in a more idiomatic Haskell API.

My impression is that most Haskell programmers interested in applications programming have collectively decided to concentrate their efforts on improving Haskell web applications instead of standalone GUI applications. Honestly, that's probably the right decision in the long run.

Another post that goes into more detail about this topic is this post written by Keera Studios:

Areas for improvement:

  • A GUI toolkit binding that is maintained, comprehensive, and easy to use
  • Polished GUI interface builders

Notable libraries:

  • gi-gtk and various other bindings such as GStreamer audio/video - GTK+ (and more generally, GObject) bindings done right (autogenerated using GObject Introspection, hence gi)
  • wx - wxWidgets bindings
  • X11 - X11 bindings
  • threepenny-gui - Framework for local apps that use the web browser as the interface
  • hsqml - A Haskell binding for Qt Quick, a cross-platform framework for creating graphical user interfaces.
  • fltkhs - A Haskell binding to FLTK. Easy install/use, cross-platform, self-contained executables.
  • FregeFX - Frege bindings to Java FX (Frege is essentially the Haskell for the JVM)
  • typed-spreadsheet - Library for building composable interactive forms
  • brick - Terminal UI based on vty package

Some example applications:

Educational resources:


Machine learning

Rating: Immature

There are two approaches to using machine learning in Haskell:

  • Use a Haskell binding to an implementation in another language
  • Use a machine learning library implemented in Haskell

You will most likely want to check out Haskell bindings to the tensorflow library if you are interested in the first approach:

You will also want to check out Haskell bindings to the ArrayFire library if you are interested in the first approach:

Also, Tweag.io has released Sparkle, a Haskell integration with Spark. This enables the use of MLib from Haskell. MLib is widely used in the industry for machine learning. Sparkle itself is fairly new.

If you are interested in Haskell implementations of machine learning libraries then the most promising ones are the HLearn project:

... and the grenade project:

Notable libraries:

  • hasktorch - Haskell bindings to libtorch which is the C++ API for PyTorch
  • HLearn-* - Advanced implementations of a subset of machine learning algorithms
  • grenade - Machine learning library implemented in Haskell with a BLAS/LAPACK backend and a high-level type-based API
  • tensorflow - Haskell bindings to Google's tensorflow project
  • arrayfire - Haskell bindings to ArrayFire
  • ad - Automatic differentiation, used as a substrate for many Haskell machine learning projects

Game programming

Rating: Immature

Haskell is a garbage collected language, so Haskell is more appropriate for the scripting / logic layer of a game but not suitable manipulating a large object graph or for implementing a high-performance game engine due to the risk of introducing perceptible pauses due to GC pauses. Also, for simple games you can realistically use Haskell for the entire stack.

Examples of games that could be fully implemented in Haskell:

  • Casual games
  • Turn-based strategy games
  • Adventure games
  • Platform / side-scrolling games
  • First-person shooter

Examples of games that are difficult to implement at all in Haskell:

  • Real-time strategy games
  • MMORPGs

Haskell has SDL, OpenGL, and Vulkan bindings, which are actually quite good, but that's about it. You're on your own from that point onward. There is not a rich ecosystem of higher-level libraries built on top of those bindings. There is some work in this area, but I'm not aware of anything production quality or easy to use.

The primary reason for the immature rating is the difficulty of integrating Haskell with existing game platforms, which often are biased towards a particular language or toolchain. The only game platform where Haskell has no issues is native binaries for desktop games. For the web, you must compile to JavaScript, which is doable. For mobile games on Android you have to cross compile and interface the Haskell logic with Android through JNI + Haskell's foreign function interface. For console games, you have no hope.

Areas for improvement:

  • Improve the garbage collector and benchmark performance with large heap sizes
  • Provide higher-level game engines
  • Improve distribution of Haskell games on proprietary game platforms

Notable libraries:

  • gloss - Simple graphics and game programming for beginners
  • Code World - Similar to gloss, but you can try it in your browser
  • vulkan - Low-level Vulkan bindings
  • gl - Comprehensive OpenGL bindings
  • SDL / SDL-* / sdl2 - Bindings to the SDL library
  • SFML - Bindings to the SFML library
  • quine - Github project with cool 3D demos
  • GPipe - Type-safe OpenGL API that also lets you embed shader code directly within Haskell. See the GPipe wiki to learn more

Educational resources:


ARM processor support

Rating: Immature

On hobbyist boards like the Raspberry Pi its possible to compile Haskell code with GHC. There are limitations; some libraries have problems on the arm platform, and GHCi only works on newer compilers. Cross compiling doesn't work with template Haskell. Stack and other large projects can take more than 1g of memory to compile.

However, if the Haskell code builds, it runs with respectable performance on these machines.

Arch (Banana Pi)

update 2016-02-25:

  • installed today from pacman, current versions are GHC 7.10.3 and cabal-install 1.22.6.0
  • a compatible version of llvm also installed automatically.
  • GHCi passes hello world test; cabal/GHC compiled a modest project normally.

Raspian (Raspberry Pi, pi2, others)

  • current version: GHC 7.4, cabal-install 1.14
  • GHCi doesn't work.

Debian Jesse (Raspberry Pi 3)

  • works with: ghc-7.10.3 and stack-1.1.2
  • Requires llvm version 3.5.2 or higher. Do not use the llvm-3.5 provided by default in the Jessie package distribution

Arch (Raspberry Pi 2)

  • current version 7.8.2, but llvm is 3.6, which is too new.
  • downgrade packages for llvm not officially available.
  • with llvm downgrade to 3.4, GHC and GHCi work, but problems compiling yesod, scotty.
  • compiler crashes, segfaults, etc.

Computer Vision

Rating: Immature

The largest real world Haskell usage of computer vision is LumiGuide, which powers municipal bicycle detection and guidance systems in Amsterdam. They maintain OpenCV bindings in their haskell-opencv library.

There are some interesting projects which try to tackle computer vision in a purely functional manner. cv-combinators, easyVision, and Zef are some examples.

There are Haskell bindings for OpenCV available via HOpenCV which has bindings for versions up to OpenCV 2.0. A fork maintained by Anthony Cowley has bindings available for versions up to OpenCV 2.4, but it pretty much stops there. Currently, OpenCV 3.0 has been released, and there are no Haskell bindings covering it.

Notable libraries:

Propaganda:


Mobile apps

Rating: Immature

This greatly lags behind using languages that are natively supported by the mobile platform (i.e. Java for Android or Objective-C / Swift for iOS).

However, one route is to compile Haskell to a supported language. For example, you can compile Haskell to Java using Eta to port Haskell games to Android.

Educational resources:


Systems / embedded programming

Rating: Bad / Immature (See description)

Since systems programming is an abused word, I will clarify that I mean programs where speed, memory layout, and latency really matter.

Haskell fares really poorly in this area because:

  • The language is garbage collected, so there are no latency guarantees
  • Executable sizes are large
  • Memory usage is difficult to constrain (thanks to space leaks)
  • Haskell has a large and unavoidable runtime, which means you cannot easily embed Haskell within larger programs
  • You can't easily predict what machine code that Haskell code will compile to

Typically people approach this problem from the opposite direction: they write the low-level parts in C or Rust and then write Haskell bindings to the low-level code.

It's worth noting that there is an alternative approach which is Haskell DSLs that are strongly typed that generate low-level code at runtime. This is the approach championed by the company Galois.

Notable libraries:

  • atom / ivory - DSL for generating embedded programs
  • copilot - Stream DSL that generates C code
  • improve - High-assurance DSL for embedded code that generates C and Ada

Educational resources:


Common Programming Needs

Maintenance

Rating: Best in class

Haskell is unbelievably awesome for maintaining large projects. There's nothing that I can say that will fully convey how nice it is to modify existing Haskell code. You can only appreciate this through experience.

When I say that Haskell is easy to maintain, I mean that you can easily approach a large Haskell code base written by somebody else and make sweeping architectural changes to the project without breaking the code.

You'll often hear people say: "if it compiles, it works". I think that is a bit of an exaggeration, but a more accurate statement is: "if you refactor and it compiles, it works". This lets you move fast without breaking things.

Most statically typed languages are easy to maintain, but Haskell is on its own level for the following reasons:

  • Strong types
  • Purity
  • Global type inference
  • Type classes
  • Laziness

The latter three features are what differentiate Haskell from other statically typed languages.

If you've ever maintained code in other languages you know that usually your test suite breaks the moment you make large changes to your code base and you have to spend a significant amount of effort keeping your test suite up to date with your changes. However, Haskell has a very powerful type system that lets you transform tests into invariants that are enforced by the types so that you can statically eliminate entire classes of errors at compile time. These types are much more flexible than tests when modifying code and types require much less upkeep as you make large changes.

The Haskell community and ecosystem use the type system heavily to "test" their applications, more so than other programming language communities. That's not to say that Haskell programmers don't write tests (they do), but rather they prefer types over tests when they have the option.

Global type inference means that you don't have to update types and interfaces as you change the code. Whenever I do a large refactor the first thing I do is delete all type signatures and let the compiler infer the types and interfaces for me as I go. When I'm done refactoring I just insert back the type signatures that the compiler infers as machine-checked documentation.

Type classes also assist refactoring because the compiler automatically infers type class constraints (analogous to interfaces in other languages) so that you don't need to explicitly annotate interfaces. This is a huge time saver.

Laziness deserves special mention because many outsiders do not appreciate how laziness simplifies maintenance. Many languages require tight coupling between producers and consumers of data structures in order to avoid wasteful evaluation, but laziness avoids this problem by only evaluating data structures on demand. This means that if your refactoring process changes the order in which data structures are consumed or even stops referencing them altogether you don't need to reorder or delete those data structures. They will just sit around patiently waiting until they are actually needed, if ever, before they are evaluated.


Single-machine Concurrency

Rating: Best in class

I give Haskell a "Best in class" rating because Haskell's concurrency runtime performs as well or better than mainstream languages and is significantly easier to use due to the runtime support for software-transactional memory.

The best explanation of Haskell's threading module is the documentation in Control.Concurrent:

Concurrency is "lightweight", which means that both thread creation and context switching overheads are extremely low. Scheduling of Haskell threads is done internally in the Haskell runtime system, and doesn't make use of any operating system-supplied thread packages.

In Haskell, all I/O is non-blocking by default, so for example a web server will just spawn one lightweight thread per connection and each thread can be written in an ordinary synchronous style instead of nested callbacks like in Node.js.

The best way to explain the performance of Haskell's threaded runtime is to give hard numbers:

  • The Haskell thread scheduler can easily handle millions of threads
  • Each thread requires 1 kb of memory, so the hard limitation to thread count is memory (1 GB per million threads).
  • Haskell channel overhead for the standard library (using TQueue) is on the order of one microsecond per message and degrades linearly with increasing contention
  • Haskell channel overhead using the unagi-chan library is on the order of 100 nanoseconds (even under contention)
  • Haskell's MVar (a low-level concurrency communication primitive) requires 10-20 ns to add or remove values (roughly on par with acquiring or releasing a lock in other languages)

Haskell also provides software-transactional memory, which allows programmers build composable and atomic memory transactions. You can compose transactions together in multiple ways to build larger transactions:

  • You can sequence two transactions to build a larger atomic transaction
  • You can combine two transactions using alternation, falling back on the second transaction if the first one fails
  • Transactions can retry, rolling back their state and sleeping until one of their dependencies changes in order to avoid wasteful polling

A few other languages provide software-transactional memory, but Haskell's implementation has two main advantages over other implementations:

  • The type system enforces that transactions only permit reversible memory modifications. This guarantees at compile time that all transactions can be safely rolled back.
  • Haskell's STM runtime takes advantage of enforced purity to improve the efficiency of transactions, retries, and alternation.

Haskell is also the only language that supports both software transactional memory and non-blocking I/O.

Notable libraries:

  • stm - Software transactional memory
  • unagi-chan - High performance channels
  • async - Futures library
  • streamly - A streaming library offering high performance concurrency

Educational resources:

Propaganda:


Types / Type-driven development

Rating: Best in class

Haskell definitely does not have the most advanced type system (not even close if you count research languages) but out of all languages that are actually used in production Haskell is probably at the top. Idris is probably the closest thing to a type system more powerful than Haskell that has a realistic chance of use in production in the foreseeable future.

The killer features of Haskell's type system are:

  • Type classes
  • Global type and type class inference
  • Light-weight type syntax

Haskell's type system really does not get in your way at all. You (almost) never need to annotate the type of anything. As a result, the language feels light-weight to use like a dynamic language, but you get all the assurances of a static language.

Many people are familiar with languages that support "local" type inference (like Rust, Java, C#), where you have to explicitly type function arguments but then the compiler can infer the types of local variables. Haskell, on the other hand, provides "global" type inference, meaning that the types and interfaces of all function arguments are inferred, too. Type signatures are optional (with some minor caveats) and are primarily for the benefit of the programmer.

Here is an example of writing a function without any types or interfaces at all and asking the compiler to infer them for you:

>>> let addAndShow x y = show (x + y)
>>> :type addAndShow
addAndShow :: (Num a, Show a) => a -> a -> String

This really benefits projects where you need to prototype quickly but refactor painlessly when you realize you are on the wrong track. You can leave out all type signatures while prototyping but the types are still there even if you don't see them. Then when you dramatically change course those strong and silent types step in and keep large refactors painless.

Some Haskell programmers use a "type-driven development" programming style, analogous to "test-driven development":

  • they specify desired behavior as a type signature which initially fails to type-check (analogous to adding a test which starts out "red")
  • they create a quick and dirty solution that satisfies the type-checker (analogous to turning the test "green")
  • they improve on their initial solution while still satisfying the type-checker (analogous to a "red/green refactor")

"Type-driven development" supplements "test-driven development" and has different tradeoffs:

  • The biggest disadvantage of types is that they don't test as many things as full-blown tests, because Haskell is not (yet) dependently typed
  • The biggest advantage of types is that they can prove the complete absence of programming errors that you can encode in the type system, whereas tests do not typically exercise every possible code path
  • Type-checking is much faster than running tests
  • Type error messages are informative: they explain what went wrong
  • Type-checking never hangs and never gives flaky results

Haskell also provides the "Typed Holes" extension, which lets you add an underscore (i.e. "_") anywhere in the code whenever you don't know what expression belongs there. The compiler will then tell you the expected type of the hole and suggest terms in scope with related types that you can use to fill the hole.

There is also a newly added "Liquid Haskell" extension under development which you can use to program with "refinement types". These types enrich Haskell's type system with the ability to decorate type signatures with logical predicates and arithmetic, and increases the number of invariants that you can encode at the type level.

Educational resources:

Propaganda:


Parsing / Pretty-printing

Rating: Best in class

Haskell parsing is sooooooooooo slick. Recursive descent parser combinators are far-and-away the most popular parsing paradigm within the Haskell ecosystem, so much so that people use them even in place of regular expressions. I strongly recommend reading the "Monadic Parsing in Haskell" functional pearl linked below if you want to get a feel for why parser combinators are so dominant in the Haskell landscape.

If you're not sure what library to pick, I generally recommend the parsec library as a default well-rounded choice because it strikes a decent balance between ease-of-use, performance, good error messages, and small dependencies (since it ships with GHC). There is also the megaparsec library, which is modern and improved version of parsec.

attoparsec deserves special mention as an extremely fast backtracking parsing library. The speed and simplicity of this library will blow you away. The main deficiency of attoparsec is the poor error messages.

The pretty-printing front is also excellent. Academic researchers just really love writing pretty-printing libraries in Haskell for some reason.

Parsing libraries:

  • parsec - Best overall "value"
  • megaparsec - Modern, actively maintained fork of parsec
  • attoparsec - Extremely fast backtracking parser
  • Earley - Earley parsing embedded within the Haskell language. Parses all context-free grammars, even ambiguous ones, with no need to left factor. Returns all valid parses.
  • trifecta - Best error messages (clang-style)
  • parsers - Interface compatible with attoparsec, parsec and trifecta which lets you easily switch between them. People commonly use this library to begin with trifecta or parsec (for better error messages) then switch to attoparsec when done for performance
  • alex / happy - Like lexx / yacc but with Haskell integration

Pretty-printing libraries:

Educational resources:

Propaganda:


Domain-specific languages (DSLs)

Rating: Mature

Haskell rocks at DSL-building. While not as flexible as a Lisp language I would venture that Haskell is the most flexible of the non-Lisp languages. You can overload a large amount of built-in syntax for your custom DSL.

The most popular example of overloaded syntax is do notation, which you can overload to work with any type that implements the Monad interface. This syntactic sugar for Monads in turn led to a huge overabundance of Monad tutorials.

However, there are lesser known but equally important things that you can overload, such as:

  • numeric and string literals
  • if/then/else expressions
  • list comprehensions
  • numeric operators

Educational resources:


Testing

Rating: Mature

There are a few places where Haskell is the clear leader among all languages:

  • property-based testing
  • mocking / dependency injection

Haskell's QuickCheck is the gold standard which all other property-based testing libraries are measured against. The reason QuickCheck works so smoothly in Haskell is due to Haskell's type class system and purity. The type class system simplifies automatic generation of random data from the input type of the property test. Purity means that any failing test result can be automatically minimized by rerunning the check on smaller and smaller inputs until QuickCheck identifies the corner case that triggers the failure.

Mocking is another area where Haskell shines because you can overload almost all built-in syntax, including:

  • do notation
  • if statements
  • numeric literals
  • string literals

Haskell programmers overload this syntax (particularly do notation) to write code that looks like it is doing real work:

example = do str <- readLine
             putLine str

... and the code will actually evaluate to a pure syntax tree that you can use to mock in external inputs and outputs:

example = ReadLine (\str -> PutStrLn str (Pure ()))

Haskell also supports most testing functionality that you expect from other languages, including:

  • standard package interfaces for testing
  • unit testing libraries
  • test result summaries and visualization

Notable libraries:

  • QuickCheck - property-based testing
  • doctest - tests embedded directly within documentation
  • free - Haskell's abstract version of "dependency injection"
  • hspec - Testing library analogous to Ruby's RSpec
  • HUnit - Testing library analogous to Java's JUnit
  • tasty - Combination unit / regression / property testing library
  • hedgehog - property-based testing with integrated shrinking
  • HTF - Preprocessor based unit testing with various output formats

Educational resources:


Data structures and algorithms

Rating: Mature

Haskell primarily uses persistent data structures, meaning that when you "update" a persistent data structure you just create a new data structure and you can keep the old one around (thus the name: persistent). Haskell data structures are immutable, so you don't actually create a deep copy of the data structure when updating; any new structure will reuse as much of the original data structure as possible.

The Notable libraries sections contains links to Haskell collections libraries that are heavily tuned. You should realistically expect these libraries to compete with tuned Java code. However, you should not expect Haskell to match expertly tuned C++ code.

The selection of algorithms is not as broad as in Java or C++ but it is still pretty good and diverse enough to cover the majority of use cases.

Notable libraries:


Benchmarking

Rating: Mature

This boils down exclusively to the criterion library, which was done so well that nobody bothered to write a competing library. Notable criterion features include:

  • Detailed statistical analysis of timing data
  • Beautiful graph output: (Example)
  • High-resolution analysis (accurate down to nanoseconds)
  • Customizable HTML/CSV/JSON output
  • Garbage collection insensitivity

Notable libraries:

  • criterion
  • gauge offers a similar feature set as criterion but has much fewer dependencies
  • tasty-bench even lighter than guage with support for comparing benchmarks

Educational resources:


Unicode

Rating: Mature

Haskell's Unicode support is excellent. Just use the text and text-icu libraries, which provide a high-performance, space-efficient, and easy-to-use API for Unicode-aware text operations.

Note that there is one big catch: the default String type in Haskell is inefficient. You should always use Text whenever possible.

Notable libraries:


Stream programming

Rating: Mature

Haskell's streaming ecosystem is mature. Probably the biggest issue is that there are too many good choices (and a lot of ecosystem fragmentation as a result), but each of the streaming libraries listed below has a sufficiently rich ecosystem including common streaming tasks like:

  • Network transmissions
  • Compression
  • External process pipes
  • High-performance streaming aggregation
  • Concurrent streams
  • Incremental parsing

Notable libraries:

Educational resources:


Serialization / Deserialization

Rating: Mature

Haskell's serialization libraries are reasonably efficient and very easy to use. You can easily automatically derive serializers/deserializers for user-defined data types and it's very easy to encode/decode values.

Haskell's serialization does not suffer from any of the gotchas that object-oriented languages deal with (particularly Java/Scala). Haskell data types don't have associated methods or state to deal with so serialization/deserialization is straightforward and obvious. That's also why you can automatically derive correct serializers/deserializers.

Serialization performance is pretty good. You should expect to serialize data at a rate between 100 Mb/s to 1 Gb/s with careful tuning. Serialization performance still has about 3x-5x room for improvement by multiple independent estimates. See the "Faster binary serialization" link below for details of the ongoing work to improve the serialization speed of existing libraries.

Notable libraries:

Educational resources:


Support for file formats

Rating: Mature

Haskell supports all the common domain-independent serialization formats (i.e. XML/JSON/YAML/CSV). For more exotic formats Haskell won't be as good as, say, Python (which is notorious for supporting a huge number of file formats) but it's so easy to write your own quick and dirty parser in Haskell that this is not much of an issue.

Notable libraries:

  • aeson - JSON encoding/decoding
  • cassava / sv- CSV encoding/decoding
  • yaml - YAML encoding/decoding
  • HsYAML - pure Haskell YAML 1.2 parser
  • xml - XML encoding/decoding
  • tomland - TOML encoding/decoding

Package management

Rating: Mature

This rating is based entirely on the recent release of the stack package tool by FPComplete which greatly simplifies package installation and dependency management. This tool was created in response to a broad survey of existing Haskell users and potential users where cabal-install was identified as the single greatest issue for professional Haskell development.

The stack tool is not just good by Haskell standards but excellent even compared to other language package managers. Key features include:

  • Excellent project isolation (including compiler isolation)
  • Global caching of shared dependencies to avoid wasteful rebuilds
  • Easily add local repositories or remote GitHub repositories as dependencies

stack is also powered by Stackage, which is a very large Hackage mono-build that ensures that a large subset of Hackage builds correctly against each other and automatically notifies package authors to fix or update libraries when they break the mono-build. Periodically this package set is frozen as a Stackage LTS release which you can supply to the stack tool in order to select dependencies that are guaranteed to build correctly with each other. Also, if all your projects use the same or similar LTS releases they will benefit heavily from the shared global cache.

Educational resources:

Propaganda:


Logging

Haskell has decent logging support. That's pretty much all there is to say.

Rating: Mature

  • fast-logger - High-performance multicore logging system
  • hslogger - Logging library analogous to Python's logging library
  • monad-logger - add logging with line numbers to your monad stack. Uses fast-logger under the hood.
  • katip - Structured logging
  • log - Logging system with ElasticSearch, PostgreSQL and stdout sinks.
  • co-log - Composable contravariant comonadic logging library.

Code formatting

Rating: Mature

Haskell has tools for automatic code formatting:

  • stylish-haskell - Less opinionated code formatting tool that mostly formats imports, language extensions, and data type definitions
  • ormolu - More opinionated formatting tool that uses GHC's own parser
  • brittany - Formats more than stylish-haskell, but less opinionated than ormolu.

Education

Rating: Mature

"Haskell Programming from first principles" has been published. I highly recommend it, for the following reasons:

  • The book does not assume any prior programming experience
  • The book does not have any conceptual gaps or out-of-order dependencies
  • The book is extremely comprehensive

Educational resources:


Databases and data stores

Rating: Immature

This is is not one of my areas of expertise, but what I do know is that Haskell has bindings to most of the open source databases and datastores such as MySQL, Postgres, SQLite, Cassandra, Redis, DynamoDB and MongoDB. However, I haven't really evaluated the quality of these bindings other than the postgresql-simple library, which is the only one I've personally used and was decent as far as I could tell.

The "Immature" ranking is based on the lack of bindings to commercial databases like Microsoft SQL server and Oracle. So whether or not Haskell is right for you probably depends heavily on whether there are bindings to the specific data store you use.

Notable libraries:


Debugging

Rating: Immature

The main Haskell debugging features are:

  • Memory and performance profiling
  • Stack traces
  • Source-located errors, using the assert function
  • Breakpoints, single-stepping, and tracing within the GHCi REPL
  • Informal printf-style tracing using Debug.Trace
  • ThreadScope

The two reasons I still mark debugging "Immature" are:

  • GHC's stack traces require profiling to be enabled
  • There is only one IDE that I know of (leksah) that integrates support for breakpoints and single-stepping and leksah still needs more polish

ghc-7.10 also added preliminary support for DWARF symbols which allow support for gdb-based debugging and perf-based profiling, but there is still more work that needs to be done. See the following page for more details:

Educational resources:


Cross-platform support

Rating: Immature

I give Haskell an "Immature" rating primarily due to poor user experience on Windows:

  • Most Haskell tutorials assume a Unix-like system
  • Several Windows-specific GHC bugs
  • Poor IDE support (Most Windows programmers don't use a command-line editor)

This is partly a chicken-and-egg problem. Haskell has many Windows-specific issues because it has such a small pool of Windows developers to contribute fixes. Most Haskell developers are advised to use another operating system or a virtual machine to avoid these pain points, which exacerbates the problem.

The situation is not horrible, though. I know because I do half of my Haskell programming on Windows in order to familiarize myself with the pain points of the Windows ecosystem and most of the issues affect beginners and can be worked around by more experienced developers. I wouldn't say any individual issue is an outright dealbreaker; it's more like a thousand papercuts which turn people off of the language.

If you're a Haskell developer using Windows, I highly recommend the following installs to get started quickly and with as few issues as possible:

  • Git for Windows - A Unix-like command-line environment bundled with git that you can use to follow along with tutorials
  • MinGHC - Use this for project-independent Haskell experimentation
  • Stack - Use this for project development

Additionally, learn to use the command line a little bit until Haskell IDE support improves. Plus, it's a useful skill in general as you become a more experienced programmer.

For Mac, the recommended installation is:

  • Haskell for Mac - A self-contained relocatable GHC build for project-independent Haskell experimentation
  • Stack - Use this for project development

For other operating systems, use your package manager of choice to install ghc and stack.

Educational resources:


Hot code loading

Rating: Immature

Haskell does provide support for hot code loading, although nothing in the same ballpark as in languages like Clojure.

There are two main approaches to hot code loading:

  • Compiling and linking object code at runtime (i.e. the plugins or hint libraries)
  • Recompiling the entire program and then reinitializing the program with the program's saved state (i.e. the dyre or halive libraries)

You might wonder how Cloud Haskell sends code over the wire and my understanding is that it doesn't. Any function you wish to send over the wire is instead compiled ahead of time on both sides and stored in a shared symbol table which each side references when encoding or decoding the function.

Haskell does not let you edit a live program like Clojure does so Haskell will probably never be "Best in class" short of somebody releasing a completely new Haskell compiler built from the ground up to support this feature. The existing Haskell tools for hot code swapping seem as good as they are reasonably going to get, but I'm waiting for commercial success stories of their use before rating this "Mature".

The halive library has the best hot code swapping demo by far:

Notable libraries:

  • rapid - Code reloading within ghci that persists state across reloads
  • plugins / hint - Runtime compilation and linking
  • dyre / halive - Program reinitialization with saved state

IDE support

Rating: Immature

The best supported editors at the moment appear to be:

I am not the best person to review this area since I do not use an IDE myself. I'm basing this "Immature" rating purely on what I have heard from others. The impression I get is that the biggest pain point is that Haskell IDEs, IDE plugins, and low-level IDE tools keep breaking. The above three editors are the ones that have historically had the fewest setup issues.

Most of the Haskell early adopters have been vi/vim or emacs users so those editors have gotten the most love. Support for more traditional IDEs has improved recently with Haskell plugins for Atom, IntelliJ, and also the Haskell-native leksah IDE.

Also, if you have a Mac then the "Haskell for Mac" development environment is supposed to work really well for learning since it provides an interactive and visual playground for exploring the code.

Notable tools:

  • hoogle — Type-based function search
  • hayoo — Haskell function search covering more libraries
  • hlint — Code linter
  • ghc-mod — editor agnostic tool that powers many IDE-like features
  • ghcid — lightweight background type-checker that triggers on code changes
  • codex — Tags file generator for cabal project dependencies.
  • hdevtools — Persistent GHC-powered background server for development tools
  • ghc-imported-from — editor agnostic tool that finds Haddock documentation page for a symbol
  • haskell-tools - Refactoring tool + library
  • ghcide - A library for building Haskell IDE tooling
  • stan - Haskell Static Analyser

Vim and Neovim Plugins:

The two plugins provide different sets of features. Many plugins installed by haskell-vim-now can be used with haskell-ide-engine as well, but there is no ready made config/installation script yet.

Emacs Plugins:

Language Service Protocol:

The recent movement is to migrate all the editor plugins for Haskell to the Microsoft's Language Protocol (LSP) which allows to support many different editors with one unified plugin. The server part of the protocol is in haskell-ide-engine (not yet on hackage). The client part is different for different editors.

The following editors are tested. See haskell-ide-engine/README.md for installation instructions:

  • Visual Studio Code / VSCode
  • Atom
  • neovim / vim7
  • Sublime Text

More editors probably work, but there is no installation instruction yet: Eclipse, Eclipse Che, Microsoft Monaco,IntelliJ / JetBrains IDEs, Emacs, Theia, Spyder IDE, Oni. See an up to date list of "language clients" at http://langserver.org/

Note that the LSP protocol specification currently only covers navigation/browsing/references, types/symbol info, refactoring, linting and completion/intellisense aspects of IDE. Project management, building, syntax highlighting and debugging are not covered yet, so you need separate editor-specific support for that. Sometimes when you install HIE as per instruction, these components are bundled as well, but sometimes they don't.

Non-LSP IDE plugins:

IDEs:

Educational resources:

Contributors

  • Aaron Levin
  • Alois Cochard
  • Ben Kovach
  • Benno Fünfstück
  • Carlo Hamalainen
  • Chris Allen
  • Curtis Gagliardi
  • Deech
  • David Howlett
  • David Johnson
  • Edward Cho
  • Greg Weber
  • Gregor Uhlenheuer
  • Juan Pedro Villa Isaza
  • Kazu Yamamoto
  • Kevin Cantu
  • Kirill Zaborsky
  • Liam O'Connor-Davis
  • Luke Randall
  • Marcio Klepacz
  • Mitchell Rosen
  • Nicolas Kaiser
  • Oliver Charles
  • Pierre Radermecker
  • Rodrigo B. de Oliveira
  • Stephen Diehl
  • Tim Docker
  • Tran Ma
  • Yuriy Syrovetskiy
  • @bburdette
  • @co-dan
  • @ExternalReality
  • @GetContented
  • @psibi
  • @newswim