Re-posted from: https://bkamins.github.io/julialang/2023/06/30/namedtuple.html
Introduction
Named tuples are one of the most commonly used data types in Julia.
In this post I want to discuss how I think of this type and use it in practice.
This post was written under Julia 1.9.1.
The basics
The Julia Manual in the section Named Tuples specifies:
The components of tuples can optionally be named, in which case a named tuple is constructed (…)
Named tuples are very similar to tuples, except that fields can additionally
be accessed by name using dot syntax (x.a) in addition to the regular indexing syntax (x[1]).
And gives the following example:
julia> x = (a=2, b=1+2)
(a = 2, b = 3)
julia> x[1]
2
julia> x.a
2
This definition highlights that NamedTuple is similar to a Tuple except that its
components can have names.
However, I find it more natural to think of a NamedTuple as an anonymous struct,
that optionally can be integer-indexed to get access to its components.
Let me give a few reasons why I think it is a better mental model.
Subtyping behavior
Let us have a look at the Tuple type first. Construct a basic tuple:
julia> t = (1, "a")
(1, "a")
julia> typeof(t)
Tuple{Int64, String}
Note that type of t is a subtype of e.g. Tuple{Integer, AbstractString}:
julia> t isa Tuple{Integer, AbstractString}
true
This subtyping behavior is called covariant.
Moreover Tuple{Integer, AbstractString} is not concrete:
julia> isconcretetype(Tuple{Integer, AbstractString})
false
This means that you cannot create a tuple that has this type.
If you try the following call:
julia> t2 = Tuple{Integer, AbstractString}(t)
(1, "a")
julia> typeof(t2)
Tuple{Int64, String}
Nothing changes. The resulting t2 tuple has concrete parameters Int64 and String.
You might ask how much of these features a NamedTuple shares? The answer is none.
Let us check:
julia> nt = (a=1, b="a")
(a = 1, b = "a")
julia> typeof(nt)
NamedTuple{(:a, :b), Tuple{Int64, String}}
Type of nt is not a subtype of e.g. NamedTuple{(:a, :b), Tuple{Integer, AbstractString}}:
julia> nt isa NamedTuple{(:a, :b), Tuple{Integer, AbstractString}}
false
However, it is a subtype of UnionAll type NamedTuple{(:a, :b), <:Tuple{Integer, AbstractString}}.
julia> nt isa NamedTuple{(:a, :b), <:Tuple{Integer, AbstractString}}
true
This subtyping behavior is called invariant and is shared with struct types.
Moreover, you can construct a value that has NamedTuple{(:a, :b), Tuple{Integer, AbstractString}} type (so it is concrete):
julia> isconcretetype(NamedTuple{(:a, :b), Tuple{Integer, AbstractString}})
true
julia> nt2 = NamedTuple{(:a, :b), Tuple{Integer, AbstractString}}((1, "a"))
NamedTuple{(:a, :b), Tuple{Integer, AbstractString}}((1, "a"))
julia> typeof(nt2)
NamedTuple{(:a, :b), Tuple{Integer, AbstractString}}
Again, this behavior is the same as for struct types.
The fact that you can flexibly set parameters of a named tuple is mostly useful
if you have heterogeneous data. Here is an example.
Permissive code:
julia> [(a=1, b=missing), (a=missing, b="x")]
2-element Vector{NamedTuple{(:a, :b)}}:
(a = 1, b = missing)
(a = missing, b = "x")
julia> typeof.([(a=1, b=missing), (a=missing, b="x")])
2-element Vector{DataType}:
NamedTuple{(:a, :b), Tuple{Int64, Missing}}
NamedTuple{(:a, :b), Tuple{Missing, String}}
Here we created a vector with elements having different types.
The code using it would not be type stable. However, we could make the types the same
(which would be more compiler friendly; this would be relevant if you processed really large data):
julia> @NamedTuple{a::Union{Int,Missing}, b::Union{String,Missing}}.([(a=1, b=missing), (a=missing, b="x")])
2-element Vector{NamedTuple{(:a, :b), Tuple{Union{Missing, Int64}, Union{Missing, String}}}}:
NamedTuple{(:a, :b), Tuple{Union{Missing, Int64}, Union{Missing, String}}}((1, missing))
NamedTuple{(:a, :b), Tuple{Union{Missing, Int64}, Union{Missing, String}}}((missing, "x"))
The difference is that now the compiler will know the type of every field of our named tuple
(and Julia gracefully handles small Unions).
Note that I used the @NamedTuple{a::Union{Int,Missing}, b::Union{String,Missing}} macro call
for a convenient way to specify NamedTuple type. Let us check it (with one more way to write it):
julia> @NamedTuple{a::Union{Int,Missing}, b::Union{String,Missing}}
NamedTuple{(:a, :b), Tuple{Union{Missing, Int64}, Union{Missing, String}}}
julia> @NamedTuple begin
a::Union{Int,Missing}
b::Union{String,Missing}
end
NamedTuple{(:a, :b), Tuple{Union{Missing, Int64}, Union{Missing, String}}}
Now, even on the syntax level, the similarity to struct definition is apparent.
Let us define two types:
julia> const S1 = @NamedTuple begin
a::Union{Int,Missing}
b::Union{String,Missing}
end
NamedTuple{(:a, :b), Tuple{Union{Missing, Int64}, Union{Missing, String}}}
julia> struct S2
a::Union{Int,Missing}
b::Union{String,Missing}
end
The S1 and S2 types are almost undistinguishable. The only differences are:
S1is a subtype ofNamedTupleUnionAll, whileS2is just a subtype ofAny;S1accepts a tuple as a single argument, whileS2accepts two arguments;S1is indexable, whileS2is not;- adding methods to foreign functions for
S1is type piracy.
How named tuples are different from structs
Let us create instances of S1 and S2 and discuss the differences mentioned above.
First the difference in the constructor:
julia> x1 = S1((1, "a"))
NamedTuple{(:a, :b), Tuple{Union{Missing, Int64}, Union{Missing, String}}}((1, "a"))
julia> x2 = S2(1, "a")
S2(1, "a")
We can see the difference in the constructor. An extra pair of parentheses is needed for S1.
It would be tempting to define:
julia> S1(x, y) = S1((x, y))
NamedTuple{(:a, :b), Tuple{Union{Missing, Int64}, Union{Missing, String}}}
julia> S1(1, "a")
NamedTuple{(:a, :b), Tuple{Union{Missing, Int64}, Union{Missing, String}}}((1, "a"))
This would work. But if you are writing library code you should never do this.
The reason is that you have just added a method for
NamedTuple{(:a, :b), Tuple{Union{Missing, Int64}, Union{Missing, String}}} type
and this type is part of Base Julia. Such behavior is called type piracy
and potentially could break some third-party code (in this case the risk is minimal,
but it is worth haveing this habit).
The difference is that with your custom type S2 you are free to add methods to
any functions you like (even third party), since you own S2. Therefore
(provided you make a correct method definition) you will not break any code by doing so.
Let us give an example. We have discussed the benefit of NamedTuple over
struct is that it is indexable:
julia> x1[1]
1
julia> x2[1]
ERROR: MethodError: no method matching getindex(::S2, ::Int64)
However, you can easily fix this:
julia> Base.getindex(x::S2, i::Int) = getfield(x, i)
julia> x2[1]
1
What we did by adding a method to Base.getindex is not type piracy since
S2 is a type that we own.
The last difference between S1 is S2 is that, as we mentioned,
S1 is subtype of NamedTuple, so it will get accepted by functions
that work with named tuples, as opposed to S2. Here is a small example:
julia> empty(x1)
NamedTuple()
julia> empty(x2)
ERROR: MethodError: no method matching empty(::S2)
Again – we could easily add a method for Base.empty for S2 if we wanted, just as we did
with Base.getindex above.
Conclusions
Given these examples named tuple shares some features with tuples, and some with structs.
My experience is that it is more similar to a struct in practice. There are two reasons:
- I rarely index or iterate named tuples, most often I access their fields by name.
- In dispatch named tuple does behave like a struct (this is quite relevant when writing your code).
Still I like to think of named tuple as an anonymous struct. Although we could give it an alias
like we did with S1 definition, this is rarely done. Typically the concrete type of a named tuple
is left dynamic, and I just define my functions for NamedTuple UnionAll.
Happy hacking!