Julia is a nice and promising scientific language for high performance computing, with the central paradigm of multiple dispatch. However, Julia does only partially support object oriented programming (OOP) with dot notation. For example, "objects" can not have their own methods. Unfortunately, this happens to be the style of programming that I prefer.
This StackOverflow question
discusses the matter, however, the fields (if exposed) in the solution get wrapped in Core.Box
, which kind of defeats the purpose of fields, as they can no longer be accessed in the manner that one would expect.
I figured out an undocumented way in which object oriented programming with dot notation, including methods and without boxing, can be implemented and decided to write this post so that the knowledge can be passed on and the relative common question of dot notational OOP in Julia could be answered with more than "No, it does not work". Additionally, I hope that this way of programming would become more widely supported in Julia.
Python Vs Julia Classes
Consider the following class in Python and how it's methods are called:
# Python
class ExampleClass:
def __init__(self, field_0, field_1):
self.field_0 = field_0
self.field_1 = field_1
def method_0(self):
return self.field_0 * self.field_1
def method_1(self, n):
return (self.field_0 + self.field_1) * n
def method_2(self, val_0, val_1):
self.field_0 = val_0
self.field_1 = val_1
ex = ExampleClass(10, 11)
ex.method_0()
ex.method_1(1)
ex.method_2(20, 22)
The common way to implemented this in Julia would look like this:
# Julia
mutable struct ExampleClass
field_0
field_1
end
function method_0(example_class)
return example_class.field_0 * example_class.field_1
end
function method_1(example_class, n)
return (example_class.field_0 + example_class.field_1) * n
end
function method_2(example_class, val_0, val_1)
example_class.field_0 = val_0
example_class.field_1 = val_1
end
ex = ExampleClass(10, 11)
method_0(ex)
method_1(ex, 1)
method_2(ex, 20, 22)
The key difference between these two examples is that the methods in Python belong to the object and the functions in Julia belong to the global scope. I do not like the concept of global scope and would like to avoid it as much as possible.
Julia OOP Methods With Dot Notation
By implementing the class in a different way we can replicate the Pythonic dot notation. Consider the following:
# Julia
mutable struct ExampleClass
field_0
field_1
method_0
method_1
method_2
function ExampleClass(field_0, field_1)
this = new()
this.field_0 = field_0
this.field_1 = field_1
this.method_0 = function()
return this.field_0 * this.field_1
end
this.method_1 = function(n)
return (this.field_0 + this.field_1) * n
end
this.method_2 = function(val_0, val_1)
this.field_0 = val_0
this.field_1 = val_1
end
return this
end
end
ex = ExampleClass(10, 11)
ex.method_0()
ex.method_1(1)
ex.method_2(20, 22)
Now it works like in Python, just as I like it!
Private Variables
What about private variables? I see that there are two ways in which private variables can be implemented. One is to add new fields to the
struct, but prefix them with an underscore like this:
# Julia
mutable struct ExampleClass
field_0
field_1
method_0
method_1
method_2
_private_var
function ExampleClass(field_0, field_1)
this = new()
this._private_var = 0
...
end
end
Unfortunately, the variables are not really private and can still be accessed from the outside. Another way, with truly private variables
is to include them in the constructor like this:
# Julia
mutable struct ExampleClass
field_0
field_1
method_0
method_1
method_2
function ExampleClass(field_0, field_1)
this = new()
private_var = 0
...
end
end
Now the private variables are really private, however, they can not be accessed with dot notation from this
, which is unfortunate.
I prefer the underscore method.
How Does it Work?
The revised Julia code relies primarily on three things: closures, variable capturing and anonymous functions.
Caveats
Unfortunately this way of programming has some caveats, at least when compared to the "idiomatic" Julia way.
- The size of the classes grow, with each method increasing the size by 8 bytes (at least on 64 bit systems) and thus allocation becomes slightly slower.
- Calling the anonymous functions is slightly slower than calling the global functions.
- Capturing variables in Julia has some performance implications. See the Julia docs.
Using
let this = this
may have some performance benefit when applied to the anonymous functions, but I am uncertain of its effects in this case. - The classes (or structs) must be mutable, since they are modified in the constructor. This can perhaps be circumvented in some fashion.
Performance Evaluation
This section provides a quick performance comparison between the both styles in Julia. Note that the results are highly dependent on the fields and
methods of the classes and as such this test should not be considered to be representative of every situation. This test focuses on allocation
and method call performance. You should benchmark your own code for your self, which probably has other underlying assumptions. The benchmarks are executed
with these two types:
# Julia
mutable struct DotNotationType
field_0::Int
field_1::Int
field_2::Int
field_3::Int
do_work::Function
function DotNotationType(field_0::Int, field_1::Int, field_2::Int, field_3::Int)
this = new()
this.field_0 = field_0
this.field_1 = field_1
this.field_2 = field_3
this.field_2 = field_3
this.do_work = function(n::Int)
sum = 0
for i in 1:n
sum += this.field_0 * this.field_1 + this.field_2 * this.field_3 ^ n
end
return sum
end
return this
end
end
mutable struct GlobalScopeType
field_0::Int
field_1::Int
field_2::Int
field_3::Int
end
function do_work(gst::GlobalScopeType, n::Int)
sum = 0
for i in 1:n
sum += gst.field_0 * gst.field_1 + gst.field_2 * gst.field_3 ^ n
end
return sum
end
Boxing & Class Sizes
println(DotNotationType(1, 2, 3, 4).field_0)
println(DotNotationType(1, 2, 3, 4).method_0)
println(sizeof(DotNotationType))
println(sizeof(GlobalScopeType))
1
#61
40
32
The fields of DotNotationType
are not boxed within Core.Box
, as in this question on StackOverflow.
Additionally, the size of DotNotationType
is 8 bytes, or 1.25x, larger than the size of GlobalScopeType
. Note that in "real" OOP languages, methods on a class do not generally make its memory imprint larger. This is due to the fact that the methods are conceptually placed in the global scope at compile time (not accounting for anonymous functions in other languages).
Allocation Performance
using Pkg
Pkg.add("BenchmarkTools")
using BenchmarkTools
function allocate_DotNotationType()
types = DotNotationType[]
for i in 1:100000
push!(types, DotNotationType(i, i * i - i))
end
return types
end
function allocate_GlobalScopeType()
types = GlobalScopeType[]
for i in 1:100000
push!(types, GlobalScopeType(i, i * i - i))
end
return types
end
# Assuming execution from a Jupyter Notebook
display(@benchmark allocate_DotNotationType())
display(@benchmark allocate_GlobalScopeType())
BenchmarkTools.Trial:
memory estimate: 78.89 KiB
allocs estimate: 2010
--------------
minimum time: 23.629 μs (0.00% GC)
median time: 26.958 μs (0.00% GC)
mean time: 34.836 μs (17.34% GC)
maximum time: 3.753 ms (96.40% GC)
--------------
samples: 10000
evals/sample: 1
BenchmarkTools.Trial:
memory estimate: 63.27 KiB
allocs estimate: 1010
--------------
minimum time: 17.818 μs (0.00% GC)
median time: 20.827 μs (0.00% GC)
mean time: 26.720 μs (15.82% GC)
maximum time: 3.183 ms (95.79% GC)
--------------
samples: 10000
evals/sample: 1
Allocating DotNotationType
is about 1.29x (comparing medians) slower than allocating GlobalScopeType
. This seems to be directly dependent on the
fact that the size of DotNotationType
is 1.25x larger than the size of GlobalScopeType
.
Method Performance
using Pkg
Pkg.add("BenchmarkTools")
using BenchmarkTools
dnts = allocate_DotNotationType()
gsts = allocate_GlobalScopeType()
function call_DotNotationType(dnts)
sum = 0.0
for d in dnts
sum += d.do_work(4)
end
return sum
end
function call_GlobalScopeType(gsts)
sum = 0.0
for g in gsts
sum += do_work(g, 4)
end
return sum
end
# Assuming execution from a Jupyter Notebook
display(call_DotNotationType(dnts))
display(call_GlobalScopeType(gsts))
display(@benchmark call_DotNotationType(dnts))
display(@benchmark call_GlobalScopeType(gsts))
-6.042282548123895e20
-6.042282548123895e20
BenchmarkTools.Trial:
memory estimate: 31.25 KiB
allocs estimate: 2000
--------------
minimum time: 72.623 μs (0.00% GC)
median time: 74.246 μs (0.00% GC)
mean time: 79.119 μs (1.95% GC)
maximum time: 2.326 ms (96.27% GC)
--------------
samples: 10000
evals/sample: 1
BenchmarkTools.Trial:
memory estimate: 16 bytes
allocs estimate: 1
--------------
minimum time: 24.066 μs (0.00% GC)
median time: 31.175 μs (0.00% GC)
mean time: 30.374 μs (0.00% GC)
maximum time: 377.014 μs (0.00% GC)
--------------
samples: 10000
evals/sample: 1
Both methods get the same result, which implies that both methods do the same work. However, calling the method of DotNotationType
is about 2.38x (comparing medians) slower than calling the function of GlobalScopeType
. I am not entirely sure what is causing the slow down, but it may have something to do with the issues mentioned in the Julia docs about the performance of capturing variables or it may have something to do with the optimization of anonymous functions.
Conclusions
- OOP is possible in Julia.
- OOP with dot notation methods is possible in Julia.
- OOP without boxing of fields is possible.
- OOP with dot notation does have caveats and generally worse performance.
- OOP dot notation methods in Julia are implemented in a generally inferior approach than that of most mainstream programming languages. Note that this is due to the fact that the way of programming presented in this post declares the functions as anonymous functions. This is not how the most common "real" OOP languages implement class methods.
Great, we got a working solution for object oriented programming with dot notational methods! The issues with the dot notational way of programming may or may not affect the general performance of your application, but I expect that this is highly domain specific. For example, in the "Method Performance" section I specifically focused on measuring the performance of method calls. Measuring the performance of actual work within a method may behave drastically different. Regardless, this may at least be a good way of porting dot notational code from, for example, Python to Julia and hopefully this will encourage Julia to implement true classes with true dot notation!
Top comments (0)