`slice_sample()` errors if `n` bigger than number of rows

[swannyy14]

In the man page for slice functions, the description for the argument n states:

If n is greater than the number of rows in the group (or prop > 1), the result will be silently truncated to the group size. If the proportion of a group size does not yield an integer number of rows, the absolute value of prop*nrow(.data) is rounded down.

The output of slice_sample used to be the same data.frame (with different ordering) if n is higher than the number of rows, but it is now returning an error.

library(dplyr)
#> 
#> Attaching package: 'dplyr'
#> The following objects are masked from 'package:stats':
#> 
#>     filter, lag
#> The following objects are masked from 'package:base':
#> 
#>     intersect, setdiff, setequal, union

tmp <- data.frame(x = c(rep('a',6), rep('b',4)), y = 1:10)
print(tmp)
#>    x  y
#> 1  a  1
#> 2  a  2
#> 3  a  3
#> 4  a  4
#> 5  a  5
#> 6  a  6
#> 7  b  7
#> 8  b  8
#> 9  b  9
#> 10 b 10

tmp %>% slice_sample(n = 15)
#> Error in `slice_sample()`:
#> ! Problem while computing indices.
#> Caused by error in `sample.int()`:
#> ! cannot take a sample larger than the population when 'replace = FALSE'

tmp %>% group_by(x) %>% slice_sample(n = 15)
#> Error in `slice_sample()`:
#> ! Problem while computing indices.
#> ℹ The error occurred in group 1: x = "a".
#> Caused by error in `sample.int()`:
#> ! cannot take a sample larger than the population when 'replace = FALSE'

^{Created on 2022-02-11 by the reprex package (v2.0.1)}

[romanzenka]

I had to write my own wrapper around slice_sample. I use it to show 5 example values to user (or less), and this update broke the functionality. I had to add extra logic to handle sampling of fewer samples than present, and also handling sampling 0 samples from data set of 0 rows...

One note: if you go from 1.0.7 -> 1.0.8, I would not expect breaking changes in the API. This is a breaking change, as my code that previously worked okay now ceased working.

[swannyy14]

I have the same problem. I have a package that randomly samples n rows from each grouping of a data.frame, and right now the package is failing to build do to this change. I used this function due to its handling of n being greater than the number of rows from each group in the data.frame. I hope they fix this soon.

[fegue]

This seems to be introduced on purpose because there is a test specifically for this error. The documentation however was not updated.
So my question would be: should we wait for this regression to be fixed or is it time to re-write the code?

[swannyy14]

That's interesting. I also see the test change... In that case, it seems like the change in behavior is on purpose and I'm strongly considering re-writing my own code. I hope that the description for the argument is updated so that there is no ambiguity.

[huftis]

slice_sample() now also results in an error when n = Inf`:

> slice_sample(iris, n = Inf)
Error in `slice_sample()`:
! Problem while computing indices.
Caused by error in `sample.int()`:
! vector size cannot be infinite
Run `rlang::last_error()` to see where the error occurred.

This used to work, and would permute the rows, which was very useful. But we can use prop = 1 instead.

BTW, values of prop > 1 now also results in an error message (this is not consistent with the documentation).

[willschulz]

Definitely need to retain the option to sample "n, or however many rows exist, whichever is smaller". Potentially good to change it from silently doing this behavior without warning the user, but a lot of us have code that depends on the function not throwing an error in this situation

[hadley]

Regression introduced in #6172 — the PR doesn't explicitly mention this change and there's no NEWS bullet so I suspect it's an omission, not a deliberate change.

[dtatarak]

Has this been addressed? I did not have this issue until recently, but now it is happening for me as well.

[hadley]

Reading this thread through more carefully, I think this is mainly a documentation issue — if you want to sample more rows than exist, you do need to set replace = TRUE as the error suggests.

Minimal reprex:

library(dplyr, warn.conflicts = FALSE)

df <- data.frame(x = 1:3)

df %>% slice_sample(n = 4)
#> Error in `slice_sample()`:
#> ! Problem while computing indices.
#> Caused by error:
#> ! Can't sample without replacement using a size that is larger than the
#>   number of rows in the data.
#> ℹ 4 rows were requested in the sample.
#> ℹ 3 rows are present in the data.
#> ℹ Set `replace = TRUE` to sample with replacement.
df %>% slice_sample(n = 4, replace = TRUE)
#>   x
#> 1 1
#> 2 1
#> 3 3
#> 4 1

^{Created on 2022-07-21 by the reprex package (v2.0.1)}