In mgirlich/tibblify: Rectangle Nested Lists

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>"
)

Introduction

With tibblify() you can rectangle deeply nested lists into a tidy tibble. These lists might come from an API in the form of JSON or from scraping XML. The reasons to use tibblify() over other tools like jsonlite::fromJSON() or tidyr::hoist() are:

It can guess the output format like jsonlite::fromJSON().
You can also provide a specification how to rectangle.
The specification is easy to understand.
You can bring most inputs into the shape you want in a single step.
Rectangling is much faster than with jsonlite::fromJSON().

Example

Let's start with gh_users, which is a list containing information about four GitHub users.

library(tibblify)

gh_users_small <- purrr::map(gh_users, ~ .x[c("followers", "login", "url", "name", "location", "email", "public_gists")])

names(gh_users_small[[1]])

Quickly rectangling gh_users_small is as easy as applying tibblify() to it:

tibblify(gh_users_small)

We can now look at the specification tibblify() used for rectangling

guess_tspec(gh_users_small)

If we are only interested in some of the fields we can easily adapt the specification

spec <- tspec_df(
  login_name = tib_chr("login"),
  tib_chr("name"),
  tib_int("public_gists")
)

tibblify(gh_users_small, spec)

Objects

We refer to lists like gh_users_small as collection and objects are the elements of such lists. Objects and collections are the typical input for tibblify().

Basically, an object is simply something that can be converted to a one row tibble. This boils down to a condition on the names of the object:

the object must have names (the names attribute must not be NULL),
every element must be named (no name can be NA or ""),
and the names must be unique.

In other words, the names must fulfill vec_as_names(repair = "check_unique"). The name-value pairs of an object are the fields.

For example list(x = 1, y = "a") is an object with the fields (x, 1) and (y, "a") but list(1, z = 3) is not an object because it is not fully named.

A collection is basically just a list of similar objects so that the fields can become the columns in a tibble.

Specification

Providing an explicit specification has a couple of advantages:

you can ensure type and shape stability of the resulting tibble in automated scripts.
you can give the columns different names.
you can restrict to parsing only the fields you need.
you can specify what happens if a value is missing.

As seen before the specification for a collection is done with tspec_df(). The columns of the output tibble are describe with the tib_*() functions. They describe the path to the field to extract and the output type of the field. There are the following five types of functions:

tib_scalar(ptype): a length one vector with type ptype
tib_vector(ptype): a vector of arbitrary length with type ptype
tib_variant(): a vector of arbitrary length and type; you should barely ever need this
tib_row(...): an object with the fields ...
tib_df(...): a collection where the objects have the fields ...

For convenience there are shortcuts for tib_scalar() and tib_vector() for the most common prototypes:

logical(): tib_lgl() and tib_lgl_vec()
integer(): tib_int() and tib_int_vec()
double(): tib_dbl() and tib_dbl_vec()
character(): tib_chr() and tib_chr_vec()
Date: tib_date() and tib_date_vec()
Date encoded as character: tib_chr_date() and tib_chr_date_vec()

Scalar Elements

Scalar elements are the most common case and result in a normal vector column

tibblify(
  list(
    list(id = 1, name = "Peter"),
    list(id = 2, name = "Lilly")
  ),
  tspec_df(
    tib_int("id"),
    tib_chr("name")
  )
)

With tib_scalar() you can also provide your own prototype

Let's say you have a list with durations

x <- list(
  list(id = 1, duration = vctrs::new_duration(100)),
  list(id = 2, duration = vctrs::new_duration(200))
)
x

and then use it in tib_scalar()

tibblify(
  x,
  tspec_df(
    tib_int("id"),
    tib_scalar("duration", ptype = vctrs::new_duration())
  )
)

Vector Elements

If an element does not always have size one then it is a vector element. If it still always has the same type ptype then it produces a list of ptype column:

x <- list(
  list(id = 1, children = c("Peter", "Lilly")),
  list(id = 2, children = "James"),
  list(id = 3, children = c("Emma", "Noah", "Charlotte"))
)

tibblify(
  x,
  tspec_df(
    tib_int("id"),
    tib_chr_vec("children")
  )
)

You can use tidyr::unnest() or tidyr::unnest_longer() to flatten these columns to regular columns.

Object Elements

For example in gh_repos_small

gh_repos_small <- purrr::map(gh_repos, ~ .x[c("id", "name", "owner")])
gh_repos_small <- purrr::map(
  gh_repos_small,
  function(repo) {
    repo$owner <- repo$owner[c("login", "id", "url")]
    repo
  }
)

gh_repos_small[[1]]

the field owner is an object itself. The specification to extract it uses tib_row()

spec <- guess_tspec(gh_repos_small)
spec

and results in a tibble column

tibblify(gh_repos_small, spec)

If you don't like the tibble column you can unpack it with tidyr::unpack(). Alternatively, if you only want to extract some of the fields in owner you can use a nested path

spec2 <- tspec_df(
  id = tib_int("id"),
  name = tib_chr("name"),
  owner_id = tib_int(c("owner", "id")),
  owner_login = tib_chr(c("owner", "login"))
)
spec2

tibblify(gh_repos_small, spec2)

Required and Optional Fields

Objects usually have some fields that always exist and some that are optional. By default tib_*() demands that a field exists

x <- list(
  list(x = 1, y = "a"),
  list(x = 2)
)

spec <- tspec_df(
  x = tib_int("x"),
  y = tib_chr("y")
)

tibblify(x, spec)

You can mark a field as optional with the argument required = FALSE:

spec <- tspec_df(
  x = tib_int("x"),
  y = tib_chr("y", required = FALSE)
)

tibblify(x, spec)

You can specify the value to use with the fill argument

spec <- tspec_df(
  x = tib_int("x"),
  y = tib_chr("y", required = FALSE, fill = "missing")
)

tibblify(x, spec)

Converting a Single Object

To rectangle a single object you have two options: tspec_object() which produces a list or tspec_row() which produces a tibble with one row.

While tibbles are great for a single object it often makes more sense to convert them to a list.

For example a typical API response might be something like

api_output <- list(
  status = "success",
  requested_at = "2021-10-26 09:17:12",
  data = list(
    list(x = 1),
    list(x = 2)
  )
)

To convert to a one row tibble

row_spec <- tspec_row(
  status = tib_chr("status"),
  data = tib_df(
    "data",
    x = tib_int("x")
  )
)

api_output_df <- tibblify(api_output, row_spec)
api_output_df

it is necessary to wrap data in a list. To access data one has to use api_output_df$data[[1]] which is not very nice.

object_spec <- tspec_object(
  status = tib_chr("status"),
  data = tib_df(
    "data",
    x = tib_int("x")
  )
)

api_output_list <- tibblify(api_output, object_spec)
api_output_list