---
title: "Column type guessing"
output: rmarkdown::html_vignette
vignette: >
  %\VignetteIndexEntry{Column type guessing}
  %\VignetteEngine{knitr::rmarkdown}
  %\VignetteEncoding{UTF-8}
---

```{r, include = FALSE}
knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>"
)
```

readr will guess column types from the data if the user does not specify the types.
The `guess_max` parameter controls how many rows of the input file are used to form these guesses.
Ideally, the column types would be completely obvious from the first non-header row and we could use `guess_max = 1`.
That would be very efficient!
But the situation is rarely so clear-cut.

By default, readr consults 1000 rows when type-guessing, i.e. `guess_max = 1000`.
Note that readr never consults rows that won't be part of the import, so the actual default is `guess_max = min(1000, n_max)`.

Sometimes you want to convey "use all of the data we're going to import to guess the column types", often without even knowing or specifying `n_max`.
How should you say that?
It's also worth discussing the possible downsides of such a request.

```{r setup}
library(readr)
```

## readr >= 2.0.0

readr got a new parsing engine in version 2.0.0 (released July 2021).
In this so-called second edition, readr calls `vroom::vroom()`, by default.
The vroom package and, therefore, the second edition of readr supports a very natural expression of "use all the data to guess": namely `guess_max = Inf`.

```{r, eval = FALSE}
read_csv("path/to/your/file", ..., guess_max = Inf)
```

Why isn't this the default?
Why not do this all the time?
Because column type guessing basically adds another pass through the data, in addition to the main parsing.

If you routinely use `guess_max = Inf`, you're basically processing every file twice, in its entirety.
If you only work with small files, this is fine.
But for larger files, this can be very costly and for relatively little benefit.
Often the column types guessed based on a subset of the file are "good enough".

Note also that `guess_max = n`, for finite `n`, works better in the second edition parser.
Due to its different design, vroom is able to sub-sample `n` rows throughout the file and it always includes the last row, whereas earlier versions of readr just consulted the first `n` rows.
In practice, the result is that the default of `guess_max = min(1000, n_max)` produces better guessed column types that it used to.
It should feel less necessary to fiddle with `guess_max` now.

<!-- 
https://github.com/r-lib/vroom/issues/352

In fact vroom generally does this guessing better than readr currently does. Readr always uses the first guess_max number of rows for guessing, whereas vroom uses rows interspersed throughout the file if guess_max is less than the total number of rows.

However as this example shows we should probably always include the last row when guessing, so I have made a change to always include the last row in the guess in the future.
-->

As always, remember that the best strategy is to provide explicit column types as any data analysis project matures past the exploratory phase.

<!-- Future link to a vignette on column specification or, if that content gets co-located here, link to later in this vignette. -->

## readr first edition and readr < 2.0.0

The parsing engine in readr versions prior to 2.0.0 is now called the first edition.
If you're using readr >= 2.0.0, you can still access first edition parsing via the functions `with_edition()` and `local_edition()`.
And, obviously, if you're using readr < 2.0.0, you will get first edition parsing, by definition, because that's all there is.

The first edition parser doesn't have a perfect way to convey "use all of the data to guess the column types".
(This is one of several reasons to prefer readr >= 2.0.0.)

Let's set up a slightly tricky file, so we can demonstrate different approaches.
The column `x` is mostly empty, but has some numeric data at the very end, in row 1001.

```{r}
tricky_dat <- tibble::tibble(
  x = rep(c("", "2"), c(1000, 1)),
  y = "y"
)
tfile <- tempfile("tricky-column-type-guessing-", fileext = ".csv")
write_csv(tricky_dat, tfile)
```

First, note that the second edition parser guesses the right type for `x`, even with the default `guess_max` behaviour.

```{r}
tail(read_csv(tfile))
```

In contrast, the first edition parser doesn't guess the right type for `x` with the `guess_max` default.
`x` is imported as logical and the `2` becomes an `NA`.

```{r}
with_edition(
  1,
  tail(read_csv(tfile))
)
```

There are three ways to proceed, each of which has some downside:

* Specify `guess_max = Inf`, just like we do for the second edition parser.

  Since readr does not know how much data it will be processing, the first
  edition engine pre-allocates a large amount of memory in the face of this
  uncertainty.
  This means that reading with `guess_max = Inf` can be extremely slow and
  might even crash your R session.
  
    ```{r}
    with_edition(
      1,
      tail(read_csv(tfile, guess_max = Inf))
    )
    ```

* Specify an actual, non-infinite value for `guess_max`.

  This is an awkward suggestion, because if you knew how many rows there were,
  we wouldn't be having this conversation in the first place.
  But sometimes you have a decent estimate and can choose a value of `guess_max`
  that is "big enough".
  This usually results in much better performance than `guess_max = Inf`.
  
    ```{r}
    with_edition(
      1,
      tail(read_csv(tfile, guess_max = 1200))
    )
    ```

* Read all columns as character, then use `type_convert()`.
  This is a bit clunky, since this obligates you to post-processing once you've
  brought you data into R.

    ```{r}
    dat_chr <- with_edition(
      1,
      read_csv(tfile, col_types = cols(.default = col_character()))
    )
    tail(dat_chr)
    
    dat <- type_convert(dat_chr)
    tail(dat)
    ```

<!-- 
https://github.com/tidyverse/readr/issues/982
https://github.com/tidyverse/readr/issues/588
  
Using type_convert() is another approach however this requires an extra step (or a few extra steps) after reading in the data.
-->

Clean up the temporary tricky csv file.
```{r}
file.remove(tfile)
```