# Web development in Julia: A progress report (Warning: Contains benchmarks)

Continuing my quest to explore the idea of using Julia for web development, I wanted to address some of my own questions around performance and implementation. My two biggest concerns were:

1. Should Julia web pages be served by a Julia HTTP server (such as HttpServer.jl) – or would it be better to have Julia work with existing software such as Apache and nginx?
2. How would Julia perform on the web compared to the competition?

### Addressing the HTTP server question

After some consideration, my personal conclusion is that a server implemented in Julia would be another codebase that would need to be maintained; would mean missing out on tools available to existing server software, such as .htaccess, modules and SDKs; and would ultimately feel like reinventing the wheel. I feel it would be more sensible to leverage existing software that already has active development and has been tried and tested in the wild.

Following from this, I knew that my primary performance concern should be the interface between the server and Julia. In my previous posts, I was using Apache and running Julia via CGI. CGI is slow enough, but a known fact of Julia is that the binary is somewhat slow to start due to internal processes/compilation. I figured that FastCGI would be the next best option – and as there are no existing solutions (except for an incomplete FastCGI library), I set about creating a FastCGI process manager for Julia.

FYI: I’ve decided to release all of my web-Julia-related code under the GitHub organisation Jaylle, which can be found at https://github.com/Jaylle. Currently only the FPM and CGI module are available, but in future that’s where I’ll add the web framework and whatever else gets developed.

I plan to elaborate on the process manager more in a future post, but in short there are two parts:

• The FastCGI server / process manager (coded in C). This accepts requests and manages and delegates to the workers.
• The worker (coded in Julia). This listens for TCP connections from the FPM, accepts a bunch of commands and then runs the requested Julia page/code.

This way, there’s always a pre-loaded version of Julia in memory, circumventing any startup concerns (unless a worker crashes, of course).

### Some early benchmarks

Now that the FPM is in a usable prerelease state, I wanted to see how it could perform compared to the alternatives. In this case, I chose PHP (obvious) and Python. I chose Python because the name often crops up in Julia discussions and there’s a FastCGI module available for it.

To run these tests, I used the Apache ab tool from my Windows machine. The server is a cheap 1-core VPS running CentOS 6 64-bit.

In all tests, the server software used was nginx. For the languages, I used PHP-FPM for PHP, Web.py for Python and the Jaylle FPM for Julia.

The individual tests are superficial and the results anecdotal, but I just wanted something to give me an idea of how my FPM performed by comparison. To elaborate:

• Basic output: Printed “Hello, [name]” – with [name] taken from the query string (?name=…)
• Looped arithmetic: Adding and outputting numbers in a loop with 7000 iterations.
• Looped method calls: Calling arithmetic-performing methods from within a loop with 7000 iterations.

Below is a table of the results. The numbers shown are requests per second; higher is better.

 Basic output Looped arithmetic Looped method calls PHP 28.17 11.29 10.92 Web.py (Python) 24.61 7.92 7.25 Jaylle (Julia) 24.85 5.27 5.12

The only thing that I can say from these results is that I’m comforted seeing that my FPM’s performance isn’t obviously terrible compared to the others, but that there’s probably some work that does need to be done to at least get it up to the same level as Python, if not PHP.

In other news, I’ve realised (4 years late) that all the cool people use Twitter now. I therefore have started actively using my account. I can’t promise that following me will improve your quality of life, but feel free to give it a chance: @phollocks

Coming soon: FPM documentation + writeup (as soon as I’m comfortable enough to tag a release).

# Revisiting emulated OOP behaviour and multiple dispatch in Julia

In an earlier post, I explored one approach to emulating bundling functionality with the data on which it operates, akin to object methods in OOP languages such as C# and PHP. A comment posted by Matthew Browne questioned whether this approach was compatible with Julia’s multiple dispatch.

This is something I thought about at the time of writing the original article, but I had assumed it wouldn’t be possible due to the way in which the anonymous functions are assigned to variables i.e. assigning one definition would overwrite the previous. However, Matthew’s question prompted me to reconsider – and after some brief experimentation and some small alterations, I found that there is indeed a way to maintain compatibility with multiple dispatch.

Below is an updated example type definition:

type MDTest
method::Function

function MDTest()
this = new()

function TestFunction(input::String)
println(input)
end

function TestFunction(input::Int64)
println(input * 10)
end

this.method = TestFunction

return this
end
end

The theory is basically the same, with the constructor assigning the methods to their respective fields within the type. The difference is in how the functions are defined and assigned.

On lines 7 and 11, methods are defined with different argument types. These methods could be defined outside of the type definition without error, but defining them within the constructor has the advantage of not polluting the global scope.

On line 15, the function is assigned to its field using some slightly different syntax, which allows both methods to be called.

With this, the example code below:

test = MDTest()

test.method("String")

test.method(5)

Produces the output:

String
50

Another advantage to this approach is the absence of anonymous functions – which, according to benchmarks and GitHub issues, have significantly worse performance compared to named functions.

# Julia variable gotchas

As is typical for many languages, assigning one variable to another in Julia does not create a copy of the variable data, but rather a reference to the existing data. However, I learned the hard way whilst working on the CGI module* that Julia does not currently support a copy-on-write mechanism for collections.

Take the example code below:

n = [ 1, 2, 3 ]

m = n

As expected, m becomes a reference to the collection referenced by n. Working with any number of mainstream languages, one might expect a copy to be made of the data referenced by n if either n or m is modified, for example:

n = [ 1, 2, 3 ]

m = n

push!(n, 4)

# Expect n = [ 1, 2, 3, 4] and m = [ 1, 2, 3 ]

This is not the case for Julia. When the array pointed to by n is modified, m maintains its reference to that same array, giving both a value of [ 1, 2, 3, 4 ].

### Problems in the wild

I encountered this quirk when working with binary data and UTF-8 strings.

n = Uint8[ 0x32, 0x33, 0x34, 0x61 ]

m = utf8(n)

empty!(n)

Having created a string using the utf8 function, I wanted to empty the original byte array to free those resources. After a few minutes of trying to figure out how a bounds error had crept in to my app, I narrowed it down to this deletion of the byte array.

Digging deeper into the Julia source, the utf8 function is just an alias for a conversion function.

utf8(x) = convert(UTF8String, x)
...
convert(::Type{UTF8String}, a::Array{Uint8,1}) = is_valid_utf8(a) ? UTF8String(a) : ...

You can see here that passing an array of Uint8 bytes to utf8() creates an instance of UTF8String with the Uint8 array as its data. The type definition for UTF8String is:

immutable UTF8String <: String
data::Array{Uint8,1}
end

As was covered above, the UTF8String’s data field will be only a reference to the collection passed to the utf8 function. If that collection is modified in any way at any point during the program’s runtime, so too will be the returned string.

### In closing

It seems that the solution at this time is to explicitly use the copy or deepcopy functions, where copies of data are required by the program logic.

The issue is explored in this Google Groups thread. If I’ve understood correctly, the gist of it is that Julia makes this sacrifice for the sake of performance. As this is a feature wanted by many, there’s a possibility of it being implemented in a later version of the language.

* Write-up to follow at a later date