The Julia Programming Language

Update 2020: This is an old, outdated review. With the benefit of more experience, I no longer agree with my criticsms in this article.

Julia is a new programming language primarily intended for scientific computing. It’s attempting to take on roles that are currently occupied by Matlab, its clones, and R. “Matlab done right” could very well be its tag-line, but it’s more than that. It has a beautiful type system, it’s homoiconic, and its generic function support would make a Lisp developer jealous. It still has a long ways to go, but, except for some unfortunate issues, it’s off to a great start.

Speaking strictly in terms of the language, doing better than Matlab isn’t really a significant feat. Among major programming languages, Matlab’s awfulness and bad design is second only to PHP. Octave fixes a lot of the Matlab language, but it can only go so far.

For both Matlab and R, the real strength is the enormous library of toolboxes and functionality available to help solve seemingly any scientific computing task. Plus the mindshare and the community. Julia has none of this yet. The language is mostly complete, but it will take years to build up its own package library to similar standards.

If you’re curious about learning more, the Julia manual covers the entire language as it currently exists. Unfortunately anything outside the language proper and its standard library is under-documented at this time.

A Beautiful Type System

One of the first things you’ll be told is that Julia is dynamically typed. That is, statically typed (C++, Java, Haskell) versus dynamically typed (Lisp, Python, JavaScript). However, Julia has the rather unique property that it straddles between these, and it could be argued to belong to one or the other.

The defining characteristic of static typing is that bindings (i.e. variables) have types. In dynamic typing, only values and objects have types. In Julia, all bindings have a type, making it like a statically typed language. If no type is explicitly declared, that type is Any, an abstract supertype of all types. This comes into play with generic functions.

Both abstract and concrete types can be parameterized by other types, and certain values. The :: syntax it used to declare a type.

type Point {T}
  x::T
  y::T
end

This creates a Point constructor function. When calling the constructor, the parameter type can be implicit, derived from the type of its arguments, or explicit. Because both x and y have the same type, so must the constructor’s arguments.

# Implicit type:
Point(1, -1)
# => Point{Int64}(1,-1)

# Explicit type:
Point{Float64}(1.1, -1.0)
# => Point{Float64}(1.1,-1.0)

Point(1, 1.0)
# ERROR: no method Point{T}(Int64,Float64)

The type can be constrained using <:. If Point is declared like the following it is restricted to real numbers. This is just like Java’s Point<T extends Number>.

type Point {T <: Real}
  x::T
  y::T
end

Unlike most languages, arrays aren’t built directly into the language. They’re implemented almost entirely in Julia itself using this type system. The special part is that they get literal syntax.

[1, 2, 3]
# => Array{Int64,1}

[1.0 2.0; 3.0 4.0]
# => Array{Float64,2}

Each Array is parameterized by the type of value it holds and by an integer, indicating its rank.

The Billion Dollar Mistake

Julia has avoided what some call The Billion Dollar Mistake: null references. In languages such as Java, null is allowed in place of any object of any type. This allowance has lead to many run-time bugs that, if null didn’t exist, would have been caught at compile time.

Julia has no null and so there’s no way to make this mistake, though some kinds of APIs are harder to express without it.

Generic Functions

All of Julia’s functions are generic, including that Point constructor above. Different methods can be defined for the same function name, but for different types. In Common Lisp and Clojure, generic functions are an opt-in feature, so most functions are not generic.

Note that this is significantly different than function overloading, where the specific function to call is determined at compile time. In multimethods, the method chosen is determined by the run-time type of its arguments. One of Julia’s notable achievements is that its multimethods have very high performance. There’s usually more of a trade-off.

Julia’s operators are functions with special syntax. For example, the + function,

+(3, 4)
# => 7

A big advantage is that operators can be passed around as first-class values.

map(-, [1, 2, 3])
# [-1, -2, -3]

Because all functions are generic, operators can have methods defined for specific types, effectively becoming operator overloading (but better!).

function +(p1::Point, p2::Point)
  return Point(p1.x + p1.y, p2.x + p2.y)
end

Point(1,1) + Point(1, 2)
# => Point{Int64}(2,3)

(Note that to write this method correctly, either Point or the method should probably promote its arguments.)

Foreign Function Interface

Julia has a really slick foreign function interface (FFI). Libraries don’t need to be explicitly loaded and call interfaces don’t have to be declared ahead of time. That’s all taken care of automatically.

I’m not going to dive into the details, but basically all you have to do is indicate the library, the function, the return type, and then pass the arguments.

ccall((:clock, "libc"), Int32, ())
# => 2292761

Generally this would be wrapped up nicely in a regular function and the caller would have no idea an FFI is being used. Unfortunately structs aren’t yet supported.

Julia’s Problems

Not everything is elegant, though. There are some strange design decisions. The two big ones for me are strings and modules.

Confused Strings

Julia has a Char type that represents a Unicode code point. It’s a 32-bit value. So far so good. However, a String is not a sequence of these. A Julia string is a byte-array of UTF-8 encoded characters.

Indexing into a string operates on bytes rather than characters. Attempting to index into the middle of a character results in an error. Yuck!

"naïvety"[4]
# ERROR: invalid UTF-8 character index

I don’t understand why this behavior was chosen. This would make sense if Julia was an old language and this was designed before Unicode was established (e.g. C). But, no, this is a brand new language. There’s no excuse not to get this right the first time. I suspect it has to do with Julia’s FFI.

Clunky, Closed Modules

Julia’s module system looks like it was taken right out of Scheme’s R6RS. This isn’t a good thing.

The module definition that wraps the entire module up in a single syntactic unit. Here’s an example from the documentation. According to the style guide, the body of the module is not indented out.

module MyModule
using Lib
export MyType, foo

type MyType
  x
end

bar(x) = 2x
foo(a::MyType) = bar(a.x) + 1

import Base.show
show(io, a::MyType) = print(io, "MyType $(a.x)")
end

That final end seals the module for good. There’s no opening the module back up to define or redefine new functions or types. If you want to change something you have to reload the entire module, which will obsolete any type instances.

Compare this to Clojure, where the module isn’t wrapped up in a syntactical construct.

(ns my.module
  (require : [clojure.set :refer [rename-keys]]))

Common Lisp’s defpackage also works like this. At any time you can jump into a namespace and make new definitions.

(in-ns 'my.module)

This is absolutely essential to interactive development. The lack of this makes Julia far less dynamic than it should be. Combined with the lack of a printer, Julia is not currently suitable as an interactive interpreter subprocess (Slime, Cider, Skewer, etc.).

This is a real shame, because I’d like to start playing around with Julia, but right now it feels like a chore. It’s needlessly restricted to a C++/Java style workflow.

I’ll probably revisit Julia once it’s had a few more years to mature. Then we’ll see if things have improved enough for real use.

Have a comment on this article? Start a discussion in my public inbox by sending an email to ~skeeto/public-inbox@lists.sr.ht [mailing list etiquette] , or see existing discussions.

null program

Chris Wellons

wellons@nullprogram.com (PGP)
~skeeto/public-inbox@lists.sr.ht (view)