In my recent post I have discussed what interfaces Julia defines for working with containers. Today I want to make a closer look at data frame objects that are defined in DataFrames.jl.

Before I move forward I want to make a small announcement. On my blog I have recently added Learning section, where I collect a list of learning materials that I find useful for doing data science with Julia. If you would like to have some position added to this list please contact me.

This post was written under Julia 1.9.0-beta4 and DataFrames.jl 1.5.0.

Interfaces refresher

Let me start with recalling the discussion we had in this post about data frame design:

  1. Data frame is not iterable. Instead, if you want to iterate its rows use the eachrow wrapper, and if you want to iterate its columns use the eachcol wrapper.
  2. You can index data frame like a matrix, but you are always required to pass both row and column indices (in other words: linear indexing is not supported).
  3. In broadcasting data frame behaves as a matrix (two dimensional container).
  4. You can get columns of a data frame by their name using property access.

For this reason often users are surprised when they read that data frame is considered to be a collection of rows. From the way it supports the standard interfaces it does not seem that it is the case.

What we are missing from the whole picture is that the four above interfaces are related to only a limited number of functions:

  1. Iteration: iterate.
  2. Indexing: getindex, setindex!, firstindex, lastindex, size.
  3. Broadcasting: axes, broadcastable, BroadcastStyle, similar, copy , copyto!.
  4. Property access: getproperty and setproperty!.

In general, DataFrames.jl exposes over 120 functions that work on data frame objects and out of them 38 are methods that are extensions to functions that are defined in Base Julia and work on collections. All these functions consider data frame to be a collection of rows.

In what follows I will go through all of them so that DataFrames.jl users have an easy reference to them in one place.

Row sorting and reordering

We support sort, sort!, sortperm, issorted, permute!, invpermute!, reverse!, reverse, shuffle!, and shuffle functions that work on data frame rows.

Here let me remark that in particular shuffling functions are often quite handy when preparing data to be passed to various statistical models.

Dropping rows

We support deleteat!, keepat!, empty!, empty, filter!, filter, first, last, and resize!.

Let me mention that resize! allows not only to drop rows form a data frame but also add them (although it is not often used).

In addition there is an isempty function that checks if data frame has zero rows. It is a important to remember that it is not required that data frame has no columns:

julia> using DataFrame

julia> df = DataFrame(a=1, b=2)
1×2 DataFrame
 Row │ a      b
     │ Int64  Int64
   1 │     1      2

julia> isempty(df)

julia> empty!(df)
0×2 DataFrame
 Row │ a      b
     │ Int64  Int64

julia> isempty(df)

Remember that if the DataFrames.jl documentation says that data frame is empty it means that it has zero rows (but it does not say anything about number of columns).

Adding rows

You can add single rows to a data frame using push!, pushfirst!, and insert! or collections of rows (in general Tables.jl tables) using append! and prepend!.

Related functions are repeat! and repeat to repeat rows in a data frame.

Row extraction

There are four functions that allow you to extract one row from a data frame only, pop!, popat!, and popfirst!.

I often find the only function useful, when I want to explicitly verify the contract that some operation returned a data frame with only one row.

Identification of missing values in rows

You can use completecases to find which rows do not contain missing values and dropmissing! and dropmissing to drop them.

Identification of unique rows

There are unique and unique! functions that return unique rows in a source data frame. To get an indicator vector which rows are non-unique use the nonunique function, and allunique checks if all rows in a data frame are unique.

Here let me show an example of nonunique functionality that allows you to choose which duplicates are highlighted (it was added in 1.5 release):

julia> df = DataFrame(a=[1, 2, 1, 3, 1])
5×1 DataFrame
 Row │ a
     │ Int64
   1 │     1
   2 │     2
   3 │     1
   4 │     3
   5 │     1

julia> nonunique(df) # by default first duplicate is kept
5-element Vector{Bool}:

julia> nonunique(df, keep=:last) # keep last duplicate
5-element Vector{Bool}:

julia> nonunique(df, keep=:noduplicates) # do not keep any duplicates
5-element Vector{Bool}:

The keep keyword argument name is used because often false in the returned vector is meant to indicate which rows should be later kept in a data frame (the same keyword argument name is consistently used in unique and unique!).


As you could see in this post there are many functions in Base Julia that support working with collections. In DataFrames.jl we wanted users to be able to reuse these functions when working with data frames. Therefore all of them are supported and they consider data frames to be collections of rows.

Sometimes it is useful to have an iterable and indexable collection of, respectively, rows and columns of a data frame. For this reason we provide the eachrow and eachcol wrappers that have this functionality. As a consequence, for clarity and to minimize the risk of error on user’s side, without being wrapped data frame is not iterable and behaves like a matrix in indexing and broadcasting.