Is slice() meant to give the same results as filter with row_number() on grouped data frames?

[alexfun]

From:

http://stackoverflow.com/questions/40187530/in-dplyr-0-5-0-why-does-slice1-not-give-the-same-results-as-filterrow-number?noredirect=1#comment67643041_40187530

Using example data

tmp_df2 <- data.frame(a = c(1, 3, 2, 4), b = c(1, 2, 3, 4))

The operations

tmp_df2 %>%
    group_by(a) %>%
    slice(1)

produces a different row order to

tmp_df2 %>%
    group_by(a) %>%
    filter(row_number() == 1)

[krlmlr]

Confirmed. @hadley: Do we guarantee any particular row order after group_by()?

[hadley]

I think we should preserve the existing order of the rows (i.e. it's a "stable" split)

[krlmlr]

@alexfun: Would you like to contribute a testthat test?

[alexfun]

@krlmlr I would be happy to. Let me know if you have any specific requests.

[krlmlr]

@hadley @alexfun: I was confused.

The following two give identical results (ordered by the grouping variable):

tmp_df2 <- data.frame(a = c(1, 3, 2, 4), b = c(1, 2, 3, 4))

tmp_df2 %>%
  group_by(a) %>%
  slice(1) %>% 
  ungroup

tmp_df2 %>%
  group_by(a) %>%
  summarize(b = b[[1]]) %>% 
  ungroup

The filter example returns the data in order of appearance:

tmp_df2 %>%
    group_by(a) %>%
    filter(row_number() == 1)

The question is: Is the slice() a summarize-like or a filter-like operation? I tend to think it's more like a filter operation, in this case the observed behavior is a bug.

[alexfun]

I would argue that summarise should also return the grouping variables in the original order presented. I had no idea that summarise exhibits this behaviour as well. This in particular is not good for my workflow, as I tend to like to cbind summary data with data frames with rows sorted in a particular order. To get around the fact that summarise rearranges rows, I will need to have key columns in both summarised data frames and other data frames, and do a join instead of a straight cbind.

[hadley]

Minimal reprex:

library(dplyr, warn.conflicts = FALSE)
df <- tibble(a = c(2, 1), b = c("x", "y")) %>% group_by(a)

df %>% slice(1)
#> Source: local data frame [2 x 2]
#> Groups: a [2]
#> 
#>       a     b
#>   <dbl> <chr>
#> 1     1     y
#> 2     2     x
df %>% filter(row_number() == 1)
#> Source: local data frame [2 x 2]
#> Groups: a [2]
#> 
#>       a     b
#>   <dbl> <chr>
#> 1     2     x
#> 2     1     y

Thinking on it more, I don't think there's any guarantee that the row order should be the same. Relying on row order (instead of doing a join) is extremely dangerous because it will silently fail. So I don't think this is like to get high enough up on my priority list to be fixed.