Access APIs

By nature, data frames are dynamic objects; column labels depend on the input source and new columns can be added or deleted while wrangling. Kotlin, in contrast, is a statically typed language where all types are defined and verified ahead of execution.

That's why creating a flexible, handy, and, at the same time, safe API to a data frame is tricky.

In the Kotlin DataFrame library, we provide four different ways to access columns, and, while they are essentially different, they look pretty similar in the data wrangling DSL.

DataFrame.read("titanic.csv")
    .add("lastName") { "name"<String>().split(",").last() }
    .dropNulls("age")
    .filter {
        "survived"<Boolean>() &&
            "home"<String>().endsWith("NY") &&
            "age"<Int>() in 10..20
    }

val survived by column<Boolean>()
val home by column<String>()
val age by column<Int?>()
val name by column<String>()
val lastName by column<String>()

DataFrame.read("titanic.csv")
    .add(lastName) { name().split(",").last() }
    .dropNulls { age }
    .filter { survived() && home().endsWith("NY") && age()!! in 10..20 }

data class Passenger(
    val survived: Boolean,
    val home: String,
    val age: Int,
    val lastName: String
)

val passengers = DataFrame.read("titanic.csv")
    .add(Passenger::lastName) { "name"<String>().split(",").last() }
    .dropNulls(Passenger::age)
    .filter {
        it[Passenger::survived] &&
            it[Passenger::home].endsWith("NY") &&
            it[Passenger::age] in 10..20
    }
    .toListOf<Passenger>()

val df /* : AnyFrame */ = DataFrame.read("titanic.csv")

df.add("lastName") { name.split(",").last() }
    .dropNulls { age }
    .filter { survived && home.endsWith("NY") && age in 10..20 }

The String API is the simplest and unsafest of them all. The main advantage of it is that it can be used at any time, including when accessing new columns in chain calls. So we can write something like:

df.add("weight") { ... } // add a new column `weight`, calculated by some expression
    .sortBy("weight") // sorting dataframe rows by its value

We don't need to interrupt a function call chain and declare a column accessor or generate new properties.

In contrast, generated extension properties form the most convenient and the safest API. Using them, you can always be sure that you work with correct data and types. However, there's a bottleneck at the moment of generation. To get new extension properties, you have to run a cell in a notebook, which could lead to unnecessary variable declarations. Currently, we are working on a compiler plugin that generates these properties on the fly while typing!

The Column Accessors API is a kind of trade-off between safety and needs to be written ahead of the execution type declaration. It was designed to better be able to write code in an IDE without a notebook experience. It provides type-safe access to columns but doesn't ensure that the columns really exist in a particular data frame.

The KProperties API is useful when you already have declared classed in your business logic with fields that correspond to columns of a data frame.