group_by: equal R float values not grouped together (machine dependent)

[sebschub]

This report is based on a discussion on so.
Here the minimal example which produces different output for different machines but all with R 3.1.0 und dplyr 0.2 (more details below) :

library(dplyr)

df <- data.frame(value=seq(1,10), height=c(rep(1,5),rep(2,5)))

# height is no factor
dfs <- df %>% group_by(height) %>% summarize(m=mean(value))
dfs$height==dfs$height[1]

# height is factor
df$height <- as.factor(df$height)
dfs <- df %>% group_by(height) %>% summarize(m=mean(value))

The faulty output from one machine: The data frame

   value height
1      1      1
2      2      1
3      3      1
4      4      1
5      5      1
6      6      2
7      7      2
8      8      2
9      9      2
10    10      2

Is summarized by height. When df$height is not a factor, the results looks like this:

  height        m
1      1 4.500000
2      1 3.000000
3      1 2.000000
4      1 1.000000
5      2 9.000000
6      2 6.000000
7      2 8.333333

while if df$height is a factor, the result is what I want:

  height m
1      1 3
2      2 8

Why does something like this happen? My first guess was that it is a numerical issue and non-factor heights differ very slightly. However, the test above shows that all height==1 are equal:

> dfs$height==dfs$height[1]
[1]  TRUE  TRUE  TRUE  TRUE FALSE FALSE FALSE

The faulty output is from an older system with with e.g. gcc 4.3.2: SUSE Linux Enterprise Server 11 (x86_64). On all more recent systems I checked, the issue did not occur.

Is this a bug, or is grouping by float just stupid and proper results cannot be guaranteed?

[hadley]

I can't reproduce this problem. Any thoughts @romainfrancois ?