Friday, March 5, 2010

Is Go suitable for building DSL?

Computer problems can be approached easier once they are expressed in an effective way. There are modern and general-purpose languages like Ruby and Python that are suitable for building a sort of internal language that is specific for the problem domain: a Domain Specific Language (DSL). A DSL tries to describe the problem with expressivity. But of course, the mean of expressivity is relative to the problem domain. In this blog post I'll explore Go's capability in building DSL.

Rake, a popular Ruby DSL


Rake, like make or ant, is a DSL for defining tasks.

Tasks definitions above are written using Rake DSL. In particular, they define two dependent tasks. They state that the :world task is dependent from the :hello task. In other words, when the :world task is executed, the :hello task will be executed first. This will result in the string "Hello World!" printed on the console. The neat fact about the code above is that it is perfectly legal ruby code. In particular,
  • task is a ruby method that accepts an hash argument
  • :hello and :world are ruby symbols
  • :hello => :world is a ruby hash
  • the do ... end part is a ruby block
Since rake DSL uses valid ruby code to express itself we name it an internal DSL.

Makengo: an internal DSL in Go


When I approached to Go the first time I was intrigued by what is stated in its homepage:

"It feels like a dynamic language but has the speed and safety of a static language. It's a joy to use."

Belonging from years of experience with dynamic languages such as ruby and enjoying its capability to build DSL, I was curious to see how much the compiled and strongly typed nature of Go can be bended in this sense. So, I started writing a small experimental project named makengo.

Using makengo DSL, the previous tasks definitions become:

How does it look compared to the ruby version? First of all, the Go version takes 117 chars against the 75 chars of the Ruby version. So, it's ~56% more verbose. In particular
  • There are too many parentheses
  • String names are longer than symbol names (:hello vs "hello", :world vs "world")
  • The block syntax is burdened by the func keyword
  • In order to express the dependency we need the long named DependsOn function rather than the concise key => value hash syntax
I could shortening the DependsOn() function name and gain some chars but what about expressivity? Rake DSL use Ruby's hash syntax to express dependency but I can't find nothing so effective in Go.

Go DSL capabilities


The retrospection on Makengo raises the main question of this post: is Go suitable for building internal DSLs? Let's try answering, exploring Go's DSL capability in more detail.

Closures


Go has many features that help building DSL. The first thing that comes to mind are anonymous functions that act as closures. In makengo, I use anonymous an function to define what a task does.

Also, anonymous functions can take arguments in order to push values in the block. For example:


Compared with C, this a great step towards expressivity. However, compared with Ruby, Go's anonymous functions syntax appears to be a lot more verbose. This is due to the necessity of a func keyword and to the use of many parentheses. Also, considering that since Go is a strongly typed language, anonymous function arguments need their own type declaration.

Method chaining


In Go, types can receive methods. In Makengo, this fact is exploited to express dependencies among tasks. Actually, the Task function returns a task object and task object can receive the DependsOn method allowing for method chaining:



Dynamic reception and metaprogramming


Go lacks of a ruby-like method_missing method. In ruby, when an unknown method is sent to an object the object responds executing method_missing. This language feature is known as Dynamic Reception. Dynamic Reception, combined with the metaprogramming magic of ruby, allows for very readable DSL (taken from Machinist):

Regarding metaprogramming, Go has an interesting reflect package that allows the manipulation of objects with arbitrary types. This comes in handy, for example, in variadic functions (see below).

Literal maps


In Ruby, a literal map could be passed as argument to a function allowing for more expressive function calls inside the DSL:

Moreover, Ruby allows you omit the delimiters for a literal map. So you can shorten it to:

In Go, a literal map looks like this:

but there is no way to shorten it.

Literal lists and varargs


Like maps, literal lists in Go have not optional delimiters so it's not convenient to use them as function arguments inside DSL. BTW, Go has good support for varargs that combined with types reflection, types assertion and interfaces, are good substitutes for literal lists. For example, in Makengo I can define tasks that depend on more than a single task passing multiple arguments to DependsOn:


A comparison table


Let's try summarize in a table what we discussed until this point.
















































DSL-related featuresGoRuby
Method chainingYesYes
Optional delimiters in mapsNoYes
Optional delimiters in listsNo but you can use varargsYes
MetaprogrammingLimitedYes
Dynamic receptionNoYes
Short signature for blocksNoYes


The answer


So what about the answer to the original question? Is Go suitable for building internal DSL? The answer of course is ... it depends!

As a compiled language, Go can't compete with dynamic languages in building complex and general purpose DSL. However, it's enough flexible to build simple DSL. Moreover, as said at the beginning of this post, the mean of expressivity is relative to the problem domain. Go can be very expressive when dealing with problems that involve network and concurrency issues.

Another road to explore is the implementation of a light VM in Go exploiting some of its advantages against C: built-in concurrency and GC. On top of this VM could live a set of dynamic languages more suited to implement external DSLs. It seems that Eleanor McHugh is on the right track with her GoLightly VM.

Go DSL capability need more investigation, of course. But a fact is clear to me: I should not think at Go as a Panacea trying to force its nature: Go it's not a scripting language. I should rather taking advantage of its peculiarities in order to fully exploit all of its expressivity.

5 comments:

  1. | | map[string]string { "world":"hello" }
    |
    | but there is no way to shorten it.

    package main

    import "fmt"

    type dp map[string]string

    func main() {
    fmt.Print( dp{ "hello":"world" } )
    }

    You'll have to gofmt it as apparently the blogspot comment system is adverse to code.

    ReplyDelete
  2. Thank you Michael. Actually I was wrong, you found a shorter way :) BTW, dp type name doesn't seem so much explicative. I'd rather prefer something like

    deps{ "hello":"world" }

    or

    dependson{ "hello":"world" }

    But now the problem is: how to build a full task definition using the map type?

    ReplyDelete
  3. What about this approach (look at main):

    http://pastie.org/pastes/867451


    Really, Go has pretty decent literal syntax for most things, so just using them w/ some helper functions is the right way for pretty much everything, in my experience.

    ReplyDelete
  4. Very nice piece of code @Anonymous, thank you! You proposed another interesting DSL pattern: a sort of type aliasing supported by helper functions. Very instructive. I'd like to update my post adding your example and giving you credits. If it's possible and if you like, get out from anonymity :)

    ReplyDelete
  5. Anonymity is preferred, thanks. :)

    ReplyDelete