The Living Thing / Notebooks :

Julia, the programming language

The hippest way to get your IEEE754 on. Hngh!

Julia: A JIT-compiled language with especial focus on high performance scientific computation.


tl;dr No magic bullet, but a handy arrow for the quiver.

Some of Julia’s community makes ambitious claims about julia being the fastest and bestest thing ever.

Unsurprisingly, julia is no panacea for programming ills, nor will it ever be. It is well designed for numerical computation. In my non-rigorous experiments it seems to do better than other scripting+compilation options, such as cython, on certain tasks. In particular, if you have dynamically defined code in the inner loop it does very well e.g you are doing a Monte Carlo simulation, but don’t know the users’s desired density ahead of time. This is more or less what you’d expect from doing compilation as late as possible rather than shipping a compiled library, but it (usually) involves less messing around with compilation toolchains, platform libraries, ABIs, makefiles etc. OTOH it uses lots of memory, so don’t be running this on your raspberry pi. (Although my colleague Rowan assures me he runs serious julia code on Raspberry Pi all the time, so maybe I need to tighten up my algorithms.)

Julia has its own idiosyncratic frictions. The community process is problematic (see also Giving up on Julia). The debugging options are not impressive, and library support is patchy, especially at this early stage.

That said, the idea of a science-users-first JIT language is timely, and Julia is that. Python, for example, has clunky legacy issues in the numeric code and a patchy API, and is ill-designed for jit-compilation, despite various projects that attempt to facilitate it. Matlab is expensive and nasty for non-numerics. Lua has some good science libraries and could likely have filled this niche but for AFAICT sociological reasons has not acquired the hipness or critical mass of Julia.

The language


In order of increasing depth

A Mental Model for Julia: Talking to a Scientist

If you read the comments to the blog, where Chris weighs in, you may notice that seems similar to talking to Chris himself. (No disrespect, Chris, I think that your contribution here is wonderful.) Who was it who said that we shape our tools, and thereafter they shape us?

Here’s something that wasn’t obvious to me: What are Symbols?

Pitfalls, gotcha

Chris Rackauckas mentions 7 Julia gotchas.

Here are some more.

Implementing methods for custom types is clunky

You want to implement a standard interface on your type so you can, e.g. iterate over it, which commonly looks like this:

for element in iterable
    # body

or equivalently

iter_result = iterate(iterable)
while iter_result !== nothing
    (element, state) = iter_result
    # body
    iter_result = iterate(iterable, state)

Here is a nice example of that: A range iterator which yields every nth element up to some number of elements looks like

struct EveryNth

julia> function Base.iterate(iter::EveryNth, state=(iter.start, 0))
           element, count = state

           if count >= iter.length
               return nothing

           return (element, (element + iter.n, count + 1))

julia> Base.length(iter::EveryNth) = iter.length

julia> Base.eltype(iter::EveryNth) = Int

(If you are lucky you might be able to inherit from AbstractArray.)

The weird thing for me is that this requires injecting your methods into another namespace; in this case, Base. That might feel gross, and it certainly does lead to surprising behaviour that this is how things are done. Importing one package can magically change the behaviour of another. This “meta-side-effect” is everywhere in Julia, and is clearly marked when you write a package, but not when you use the package. NB this injection works in many other languages, e.g. Python, but there the convention is to avoid it.

The type system is logical, although it’s not obvious if you are used to classical OOP. (This last part is not a criticism.)

It allocates for and copies array slices per default

You are using a chunk of an existing array and don’t want to copy? Consider using views for slices, they say, which means not using slice notation but rather the view function, or the handy @views macro.

Custom types and containers are scruffy

If you want fast numerics but you are not sure of your float precision, you should not use AbstractFloat for your ~~argument~~ type definitions ~~(although it works fine for simple types)~~. EDIT: You don’t need to worry about this in method definitions, as Chris Rackauckas points out.

If you do need well-typed containers, the idiomatic way to do this is using parametric types and parametric methods and so-called orthogonal design.

AFAICT a parametric declaration is the goods if, e.g. you want your method to work with an array of Float32, or of Float64, or of Float16, or of Decimal, but not an array which mixes those types. But also it should work on dense arrays, or view of dense arrays, or sparse arrays thereof, right? Then the correct type would look like one of these:

RealMatrixType = Union{AbstractMatrix{F},SparseMatrixCSC{F}} where {F<:Real}
RealVectorType = Union{AbstractVector{F},SparseVector{F}} where {F<:Real}

EDIT: you usually don’t need to bother with this kind of declaration; the compiler will work it dynamically out so long as the arguments’ types are sensible. The rule is: let the compiler work out the argument types in functions, but you should choose the types in definitions.

It’s unstable and hangs all the time


Especially when it’s doing things that are supposed to be julia specialities, such as JIT-compiling dynamic inner functions. Use the recommended julia command:

killall -9 julia

FWIW this problem has mostly occurred for me using the JunoPro Intel MKL builds. Vanilla Juno is much more stable, and also smaller, and not noticeably slower on my particular code.

Argument syntax is only OK

Keyword arguments exist but do not participate in method dispatch. Basically, keyword arguments are second-class citizens and might make things slow or stupid if you need to specialise your code based on them. So… design your functions around that. This is usually OK with careful thought, but often leads to many irritating helper functions to handle default arguments.

Going fast isn’t magical

Don’t believe that wild-eyed evangelical gentleman at the Julia meetup. He doesn’t write actual code; he’s too busy talking about it. You still need to optimize and hint.

Here is a convenient overview of timer functions to help you work out what to optimize.

Active Julia contributor Chris Rackauckas, in the comments of this blog, is at pains to point out that type-hinting is rarely required, and that often even threaded vectorisation annotation is not needed, which is amazing. (The manual is more ambivalent, but anything here is a mild bit of magic.)

Not to minimize how excellent that is, but bold coder, still don’t get lazy. This doesn’t eliminate role that performance annotations and careful code structuring typically play, and I also do not regard it as a bug that you occasionally annotate code to inform the computer what you know about what can be done with it. Julia should not be spending CPU cycles proving non-trivial correctness results for you that you already know. That would likely be solving redundant, possibly superpolynomially hard problems to speed up polynomial problems, which is a bad bet. It is a design feature of julia that it allows you to be clever; it would be a bug if it tried to be clever in a stupid way. You will need to think about array views versus copies. And coding so that, e.g. the broadcast mechanism can do it’s best job optimising requires some thought. (See also the blog post on how you structure algorithms for fancy environments and GPU computing.)

That said, you can do some non-magically acceleration by using cuarrays to push computation onto GPUs. There is also crazy-idealistic support for goole cloud TPU compilation.

Does this help?

Unsubstantiated rumour, (see Chris Rackauckas’ comments for substantiation) is that the following optimises system julia 0.6 installation for local CPU:

        dirname(JULIA_HOME), "share", "julia", "build_sysimg.jl"

It is not needed for Julia 0.7+.

Arrays are column-major

The first index should change fastest.

Multidimensional arrays in Julia are stored in column-major order. This means that arrays are stacked one column at a time. This can be verified using the vec function or the syntax [:] … Notice that copy_cols is much faster than copy_rows. This is expected because copy_cols respects the column-based memory layout of the Matrix and fills it one column at a time. Additionally, copy_col_row is much faster than copy_row_col because it follows our rule of thumb that the first element to appear in a slice expression should be coupled with the inner-most loop. This is expected because copy_cols respects the column-based memory layout of the Matrix and fills it one column at a time. Additionally, copy_col_row is much faster than copy_row_col because it follows our rule of thumb that the first element to appear in a slice expression should be coupled with the inner-most loop.



For Julia > 0.7, you should use Revise.jl, and Project environments.

(v1.0) pkg> generate PatternMachine
Generating project PatternMachine:

shell> cd PatternMachine

(v1.0) pkg> activate .

julia> import PatternMachine
[ Info: Precompiling PatternMachine [93e276e4-e6da-11e8-1ff8-9f9dfed081b5]

(PatternMachine) pkg> add DSP

Revise auto init is complicated.: To .julia/config/startup.jl add

atreplinit() do repl
        @eval using Revise
        @async Revise.wait_steal_repl_backend()

and (they claim) that to julia/config/startup_ijulia.jl you must add

    @eval using Revise

But for me this leads to the kernel hanging shortly after startup, and the non-Ijulia setup is sufficient.

For VS code users .julia/config/startup.jl should purportedly be

atreplinit() do REPL
    @schedule begin
            @eval using Revise
        catch err
            warn("Could not load Revise.")

Here is Erik Engheim’s modern workflow walk-through.

crappy 0.6 workflow

You are using Julia because it is dynamic and because it is fast, but if you try to use code in a dynamic fashion, in the global namespace, it is no longer fast. Many complicated interactions ensue with the module system, and the recommended workarounds keep changing, and then suddenly it takes 90 seconds to reload all your module dependencies and you may as well have a compile/link cycle and use c++. This is slightly less awful in 0.7, but definitely needs thought.

Julia 0.6 workflow tips, summarised for you:

Put code under development in a temporary module. Create a file, say Tmp.jl, and include within it

module Tmp
  <your definitions here>

Create another file, say tst.jl, which begins with import Tmp and includes tests for the contents of Tmp. reflect updates by using reload("Tmp").


What works more easily for me is running, in Juno or jupyter console:



There is a reasonable IDE called juno, built on atom. There is jupyter integration through IJulia.

Both these have their own annoyances.


Juno is single-window only so you can’t use multiple monitors, and thus you end up squinting at tiny windows of code hidden between all the outputs. Atom’s panes aren’t well-designed for this use-case. For me that means there are a million tiny frictions distracting me from writing code in Juno. I can’t stand it.

If you install Juno as an app, but you also already use Jupyter, there is an additional annoyance in that it hijacks your atom install in a confusing way and mangles your various package preferences. If you love juno, I recommend installing it from within atom via the uber-juno package.

Possibly you can bypass this using homebrew? I didn’t try. But maybe give this a burl:

brew cask install juno

I personally don’t find juno worth the irritations, and never use it. Instead…


IJulia isn’t an IDE er se, it’s a sort-of light interactive notebook. You dont’ want to edit code in Julia; for that you’ll need a text editor. There are loads of those. I use visual studio code. These two components have have a less annoying, zippier and more stable workflow for my style, but not trying to do everything badly, which is what my experience of Juno reminds me of.

IJulia is also not flawless. For a start, it’s built on Jupyter, which I also don’t love.

For another thing, it does its own overzealous installs per default, profligately installing another copy of jupyter, which you then have to maintain separately. Boring. You can bypass this by commanding it to use the perfectly good jupyter you already have installed:

ENV["JUPYTER"] = "/usr/local/bin/jupyter"

Now julia appears as a normal kernel in your usual jupyter setup.

If you want particular upyter kernels for particular julia environments, it is possible: by using the following kernel interpreter template:

julia -e 'import Pkg; Pkg.activate("myproject")' -L /path/to/kernel.jl


There is a package called Weave.jl which is inspired by R‘s knitr but compatible with jupyter, which is asl not an IDE, but a good way of invoking reproducible self-documenting code. It could probably be used to fashion a working academic paper if you used the right pandoc hacks.


NB you can also use RMarkdown in julia mode It’s not clear to me how graphing works in this setup.


See the API list


Sort of easy, but there is a tedious need to define the call signature at call time. Survivable.



This package provides an interface from R to Julia, based on the XR structure, as implemented in the XR package, in this repository.


rjulia provides an interface between R and Julia. It allows a user to run a script in Julia from R, and maps objects between the two languages.



conda create -n conda_jl python
conda activate conda_jl
env CONDA_JL_HOME="$HOME/anaconda3/envs/conda_jl" \
   PYTHON=(which python) \
   JUPYTER=(which jupyter) \

PyCall.jl invokes python. Per default it installs yet another conda python, via Conda.jl, and defaults to the elderly python 2.7. This is irritating for various reasons, such as being ancient, and eating your diskspace with yet more versions of the same shit that you already have installed if you use python, except shittier versions.

Here is how to use an existing version

ENV["PYTHON"] = (which python3)"PyCall")

Here is how you use Conda, but with python 3:

ENV["CONDA_JL_VERSION"] = "3""Conda")

Here is how you use an existing environment

conda create -n conda_jl python
export CONDA_JL_HOME=~/anaconda3/envs/conda_jl
julia -e '"Conda")'

Either way you should regularly run conda clean to stop your disk filling up with obsolete versions of obscure dependencies for that package you tried out that one time because conda is weird.

conda clean -pt

Data loading/saving/exchange

The names are nearly all self explaining

Debugging and coding

Debugger, Gallium.jl:

This is the julia debugger. Please note that the 0.6 version of this package currently does not support breakpointing, C/C++ debugging or native code inspection. These features are being rebuilt, but were never particularly reliable in prior versions of this package and a cause of instability for the more mature features. In exchange, this package features a significantly more robust pure julia debug prompt, provided by ASTInterpreter2. Please file interpreter issues against that package.

Linter, Lint.jl (also has an atom linter plugin):

Signal processing, esp audio has been split off from core and now needs to be installed.

Numerical tours of signal processing in Julia

JuliaAudio processes audio. They recommend PortAudio.jl as a realtime soundcard interface, which looks sorta simple. See rkat’s example of how this works. They have nice abstractions like SampledSignals. This stuff is used, at, e.g. Dolby for algorithms prototyping.

Images.jl processes images.

Linear algebra

The default linear algebra baked in to julia is good, but there are a lot of specialisations and optimisations for particular things.

Einstein summation

Einstein summation is a handy way of thinking about lots of tensor operations which is not as esoteric as it sounds, inclduing, e.g. batch matrix multiply. TensorOperations.jl does some optimised einstein summations. Einsim.jl does reputedly more einstein summations but with possibly less optimistaion.

Structured matrices

Chris says

One of the best scientific computing packages is BandedMatrices.jl. It lets you directly define banded matrices and then overloads things like * and \ to use the right parts of BLAS. Compared to just using a sparse matrix (the standard MATLAB/Python way), this is SCREAMING FAST (the QR factorization difference is pretty big). Banded matrices show up everywhere in scientific computing (pretty much every PDE has a banded matrix operator). If it’s not a banded matrix, then you probably have a block-banded matrix, and thank god that @dlfivefifty also wrote BlockBandedMatrices.jl which will utilize the same speed tricks but expanded to block banded matrices (so, discretizations of systems of PDEs).

Statistics and dataframes

Data frames are provided by DataFrames.jl, and also DataTables.jl (deprecated, but used by many packages). The two are subtly incompatible in completely boring ways which you can hopefully ignore soon because DataTables is deprecated For now, use IterableTables.jl to translate where needed between these and many other data sources. One can access DataFrames (And DataTables and SQL databases and streaming data sources) using Query.jl. You can load a lot of the r standard datasets using RDatasets.

using RDatasets
iris = dataset("datasets", "iris")
neuro = dataset("boot", "neuro")

Another notable product of JuliaStats organisation is Distributions.jl, a probability distribution toolkit providing densities, sampling etc. It is complemented by the nifty Bridge.jl which simulates random processes with univariate indices.

Turing.jl does posterior inference, in a compsotional model setting.

Turing.jl is a Julia library for (universal) probabilistic programming. Current features include:

DynamicHMC does Hamiltonian/NUTS sampling in a raw likelihood setting.

Possibly it is a competitor of Klara.jl, the Juliastats MCMC.

Lasso and other sparse regressions are available in Lasso.jl which reimplements the lasso algorithm in pure julia, GLMNET.jl which wrap the classic Friedman FORTAN code for same. There is also (functionality unattested) an orthogonal matching pursuit one called OMP.jl but that algorithm is simple enough to bang out yourself in an afternoon, so no stress if it doesn’t work. Incremental/online versions of (presumably exponential family) statistics are in OnlineStats. Mixedmodels.jl

is a Julia package providing capabilities for fitting and examining linear and generalized linear mixed-effect models. It is similar in scope to the lme4 package for R.


The julia wiki includes worked examples using various engines.

A curious feature of the julia toolkits is that they all, AFAICT produce SVG graphics per default, which easily become gigantic for even medium-large data sets. For exploratory data inspection you will probably want to disable this, unless you are careful to make sure your datasets are small. Same goes for interactive JS stuff - it’s only interactive if your browser can render it without crashing. Work that out by rendering to PNG before you try the fancy SVG stuff, whcih will fill up your repo and hard drive.


The aspirational ggplot clone is Gadfly. It’s elegant, pretty, well integrated into the statistics DataFrame ecosystem, but missing some stuff, and has some weird gotchas.

Switch Gadfly from SVG to PNG or similar rasterised format:

First you will need to install the Fontconfig and Cairo bonus libraries


Now you can force PNG output:

  PNG(150, 150),
  plot(data,  x=:gamma, y=:p, Geom.point, Geom.line)

Another weird characteristic is that Gadfly is slow, especially on intial startup; But even after startup, if I histogram a million data points it takes 30 seconds for each plot of that histogram. It saps the utility of an elegant graph grammar if it’s not responsive to your elegant adjustments.

Gadfly is based on a clever declarative vector graphics system called Compose.jl, which is independently useful.


Plots.jl wraps many other plotting packages as backends, although notably not Gadfly. The author explains this as being because Gadfly does not support 3d or interactivity, although since I want neither of those things in general, especially for publication, this is a contraindication for the compatibilities of our mutual priorities. I have had enough grief trying to persuade various mathematics department that this whole “internet” thing is anything other than a PDF download engine; I don’t need to compromise my print output for animation of all the useless fripperies. We tell you, this “multimedia” thing is a fad, and the punters will all be going back to quill and vellum soon enough.

I’m sure you can shoehorn it into print-quality CMYK plots. As an example, if you want fast plots of big things it wraps GR.jl which in turn wraps GR, a cross-platform visualisation framework:

GR is a universal framework for cross-platform visualization applications. It offers developers a compact, portable and consistent graphics library for their programs. Applications range from publication quality 2D graphs to the representation of complex 3D scenes.[…]

GR is essentially based on an implementation of a Graphical Kernel System (GKS) and OpenGL. […] GR is characterized by its high interoperability and can be used with modern web technologies and mobile devices. The GR framework is especially suitable for real-time environments.

Switch to PNG from SVG:

using Plots: gr, default

For all its smoothness and speed, Plots is not very pretty and its not clear how to make it look nicer, since nice-looking-ness is hidden down at the toolkit layer.

Plots also targets various Javascript online backends like Plotly, which is neat. I sadly have no use for this at the moment. As I mentioned, in my department, mentioning you can output to this “online” nonsense is about as popular as claiming that it outputing your data as modulated flatulence, since the two are approximately equivalent in terms of our professional performance metrics.

Plots has a rich extentions ecosystem. PlotRecipes and StatPlots use a “Recipes” system to provide macro-based data-specific plotting tools.

For example, StatPlots causes Plots to have sensible DataFrame support.

Table-like data structures, […] are supported thanks to the macro @df which allows passing columns as symbols.

using StatPlots
using DataFrames, IndexedTables
df = DataFrame(a = 1:10, b = 10 .* rand(10), c = 10 .* rand(10))
@df df plot(:a, [:b :c], colour = [:red :blue])

There are many other examples.


Vega.jl binds to a javascript ggplot-alike, vega. There is a “higher level” binding also called VegaLite.jl.

It’s unclear to me how either of these work with print graphics, and they are both lagging behind the latest version of the underlying Vega library, so I’m nervous of them.


Winston.jl has a brutalist simple approach but seems to go.

Differentiating, optimisation, differentiable learning

The deep learning toolkits are lacking the overblown feature lists of the fancy python/C++ libraries (e.g. mobile app building, cuDNN-backed optimisations) But maybe the native performance is OK without bells and whistles. Also, no-one runs Julia on their phone anyway.

Tensorflow.jl. (Surely one misses the benefit of Julia by using tensorflow, since there are two different array-processing infrastructures to pass between, and a very different approach to JIT versus pre-compiled execution. Or is it a match made in heaven?)

Flux.jl sounds like a reimplementation of tensorflow-style differentiable programming inside Julia, which strikes me as the way you’d actually do this right, given the end-to-end-optimised design philosophy of julia.

Flux is a library for machine learning. It comes “batteries-included” with many useful tools built in, but also lets you use the full power of the Julia language where you need it. The whole stack is implemented in clean Julia code (right down to the GPU kernels) and any part can be tweaked to your liking.

It’s missing many features of tensorflow, but has nice dynamism like pytorch. Its GPU support is real, and elegant, but AFAICS more proof-of-concept than production-ready, depending on the supposition that cuarrays can represent all the operations you need and will perform them optimally, and that you don’t need any fancy DNN-specific GPU optimizations.

Alternatively, Mocha.jl is a belt-and-braces deep learning thing, with a library of pre-defined layers. Swap out some of the buzzwords of Flux with newer ones, and skip some older ones, and there you are. However, it doesn’t appear to be nearly as flexible, missing state filters and recurrent nets and many other models that are made less holy by their distance from the platonic ideal of deep learning, the convnet, and therefore it is not much use to me, despite promising clever things about using CUDA etc. YMMV.

Knet.jl is another deep learning library that claims to show the ease of implementing deep learning frameworks. Their RNN support looks a little limp, and they are using the less-popular autodiff frameworks, so I won’t be using ‘em but point is well taken.

If you are going to do that, why not do something left-field like use the dynamical systems approach to deep learning, popularised by Haber and Ruthotto et al, who have released their models as meganet.jl. I’m curious to see how they work.

JuMP support many types of optimisation, including over non-continuous domains, and is part of the JuliaOpt family of confusingly diverse optimizers, whcih invoke veraious ub-families of optimizers. The famous NLOpt solvers comprise one such class, and they can additionally be invoked separately.

Unlike NLPT and the JuMP family, Optim (part of JuliaNLSolvers) solves optimisation problems purely inside Julia. It will automatically invoke ForwardDiff. Assumes mostly unconstrained problems.

The autodiff stuff is getting intersting enough to deserve its own section


The juliadiff project produces ForwardDiff.jl and ReverseDiff.jl which do what you would expect, namely autodiff. Julia has an embarrassement of different methods of autodiff and it’s not always clear the selling points of each. When do you use DualNumbers.jl to differentiate? Probalby never; It seems to be the same as ForwardDiff.jl. OTOH, HyperDualNumbers, for example, promises cheap 2nd order derivatives. Zygote.jl, an alternative AD library from the creators of the aforementioned Flux has fancy features including compiling to Google Cloud TPUs. (Flux itself does something different) See also XGrad.

Julia’s homoiconicity and introspection makes this intersting.

Look at how HyperDualNumbers handles it, from the homepage example, where we are eva;luating derivatives of

$$ f(x) = e^x / (sqrt(sin(x)^3 + cos(x)^3))$$

To calculate these derivatives at a location x, evaluate your function at hyper(x, 1.0, 1.0, 0.0). For example:


t0 = Hyper(1.5, 1.0, 1.0, 0.0) y = f(t0) 4.497780053946162 + 4.053427893898621ϵ1 + 4.053427893898621ϵ2 + 9.463073681596601ϵ1ϵ2 “` The first term is the function value, the coefficients of both ϵ1 and ϵ2 (which correspond to the second and third arguments of hyper) are equal to the first derivative, and the coefficient of ϵ1ϵ2 is the second derivative.


ApproxFun.jl does Chebychev and Fourier interpolations.

Laplacians.jl by Dan Spielman et al is an advanced matrix factorisation toolkit.

UIs and servers

HttpServer.jl does basic http protocol serving; this is made modular and composable by Mux.jl. Fancy caching and templating etc come from Genie.jl.

Escher.jl goes further, rendering HTML UI widgets etc.

Various other options are listed in aviks’ stackoverflow answer.

QML.jl magically plugs into Qt5.