summarise problems when using just-created summary columns

[andrewblim]

Not sure if this is a problem per se or whether we should just not use just-created summary columns for new columns (but maybe in this case nicer to produce an error?). I get slightly different values each time I run summarise, as shown below. It does work properly when I avoid reusing the just-created columns. So far I've only seen this when I have more than one summary column reusing a previously created summary column, which is why in all the examples below I've got columns diff1 and diff2.

I'm running dplyr 0.1.1 and R 3.0.2 on OS X Mavericks.

> require('dplyr')
> df <- tbl_df(data.frame(id=c(1,1,2,2,3,3), a=1:6))
> df %.% group_by(id) %.% summarise(biggest=max(a), smallest=min(a), diff1=biggest-smallest, diff2=smallest-biggest)
Source: local data frame [3 x 5]

  id biggest smallest diff1 diff2
1  3       6        5     1     0
2  2       4        3     1     0
3  1       2        1     1     0
> # produces some randomly different values when rerun, not always the same ones, and sometimes segfaulted on big tbl_dfs I was working with
> df %.% group_by(id) %.% summarise(biggest=max(a), smallest=min(a), diff1=biggest-smallest, diff2=smallest-biggest) 
Source: local data frame [3 x 5]

  id biggest smallest diff1 diff2
1  3       6        5     1 32643
2  2       4        3     1     0
3  1       2        1     1     0
> df %.% group_by(id) %.% summarise(biggest=max(a), smallest=min(a), diff1=biggest-smallest, diff2=smallest-biggest)
Source: local data frame [3 x 5]

  id biggest smallest diff1  diff2
1  3       6        5     1 -32641
2  2       4        3     1      0
3  1       2        1     1      0
> # but this seems to work consistently
> df %.% group_by(id) %.% summarise(biggest=max(a), smallest=min(a), diff1=max(a)-min(a), diff2=min(a)-max(a))  # seems to work
Source: local data frame [3 x 5]

  id biggest smallest diff1 diff2
1  3       6        5     1    -1
2  2       4        3     1    -1
3  1       2        1     1    -1

[hadley]

Reproducible for me with the latest dev version.

[romainfrancois]

Thanks. That was a tricky one. The c++ class I used to handle newly created variables (SummarisedSubsetTemplate) was using a field to keep track of the group the indices were referring to, assuming that groups were being handled in sequence.

When making diff1, it considered it was processing groups 0, 1 and 2, and when making diff2, it kept going and considered it was processing groups 3, 4, 5.

I augmented the SlicingIndex class so that it has both the actual indices but also the index of the group.

[andrewblim]

Thank you! Impressed by the turnaround.