Inconsistent results when overwriting and reusing variable in summarise()

[krlmlr]

library(dplyr)
library(tibble)
a <- as_tibble(seq(1:10))
for (i in 1:50) print( a %>% summarise(value = mean(value), sd = sd(value)) )

On my machine the sd result is sometimes 1.73, sometimes Inf. Something very strange is going on here.

#2169 (comment), CC @nuskcalb.

[krlmlr]

With R --debugger=valgrind I'm getting, among others:

==26363== Conditional jump or move depends on uninitialised value(s)
==26363==    at 0x4EACFA8: R_IsNA (arithmetic.c:120)
==26363==    by 0x4F4D2E4: Rf_formatReal (format.c:329)
==26363==    by 0x4F8179C: do_format (paste.c:482)
==26363==    by 0x4F39DC0: bcEval (eval.c:5658)
==26363==    by 0x4F4499F: Rf_eval (eval.c:616)
==26363==    by 0x4F4526A: forcePromise (eval.c:515)
==26363==    by 0x4F45757: FORCE_PROMISE (eval.c:4258)
==26363==    by 0x4F45757: getvar (eval.c:4300)
==26363==    by 0x4F3D3B1: bcEval (eval.c:5425)
==26363==    by 0x4F4499F: Rf_eval (eval.c:616)
==26363==    by 0x4F4661C: Rf_applyClosure (eval.c:1135)
==26363==    by 0x4F4060C: bcEval (eval.c:5630)
==26363==    by 0x4F4499F: Rf_eval (eval.c:616)
==26363== 

[krlmlr]

Reprex output, note the different results after some iterations:

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
library(tibble)
a <- as_tibble(seq(1:10))
for (i in 1:50) print( a %>% summarise(value = mean(value), sd = sd(value)) )
#> # A tibble: 1 × 2
#>   value       sd
#>   <dbl>    <dbl>
#> 1   5.5 1.739253
#> # A tibble: 1 × 2
#>   value       sd
#>   <dbl>    <dbl>
#> 1   5.5 1.739253
#> # A tibble: 1 × 2
#>   value       sd
#>   <dbl>    <dbl>
#> 1   5.5 1.739253
#> # A tibble: 1 × 2
#>   value       sd
#>   <dbl>    <dbl>
#> 1   5.5 1.739253
#> # A tibble: 1 × 2
#>   value       sd
#>   <dbl>    <dbl>
#> 1   5.5 1.739253
#> # A tibble: 1 × 2
#>   value       sd
#>   <dbl>    <dbl>
#> 1   5.5 1.739253
#> # A tibble: 1 × 2
#>   value       sd
#>   <dbl>    <dbl>
#> 1   5.5 1.739253
#> # A tibble: 1 × 2
#>   value       sd
#>   <dbl>    <dbl>
#> 1   5.5 1.739253
#> # A tibble: 1 × 2
#>   value       sd
#>   <dbl>    <dbl>
#> 1   5.5 1.739253
#> # A tibble: 1 × 2
#>   value       sd
#>   <dbl>    <dbl>
#> 1   5.5 1.739253
#> # A tibble: 1 × 2
#>   value       sd
#>   <dbl>    <dbl>
#> 1   5.5 1.739253
#> # A tibble: 1 × 2
#>   value           sd
#>   <dbl>        <dbl>
#> 1   5.5 5.469883e+96
#> # A tibble: 1 × 2
#>   value       sd
#>   <dbl>    <dbl>
#> 1   5.5 1.739253
#> # A tibble: 1 × 2
#>   value       sd
#>   <dbl>    <dbl>
#> 1   5.5 1.739253
#> # A tibble: 1 × 2
#>   value       sd
#>   <dbl>    <dbl>
#> 1   5.5 1.739253
#> # A tibble: 1 × 2
#>   value       sd
#>   <dbl>    <dbl>
#> 1   5.5 1.739253
#> # A tibble: 1 × 2
#>   value       sd
#>   <dbl>    <dbl>
#> 1   5.5 1.739253
#> # A tibble: 1 × 2
#>   value       sd
#>   <dbl>    <dbl>
#> 1   5.5 1.739253
#> # A tibble: 1 × 2
#>   value       sd
#>   <dbl>    <dbl>
#> 1   5.5 1.739253
#> # A tibble: 1 × 2
#>   value       sd
#>   <dbl>    <dbl>
#> 1   5.5 1.739253
#> # A tibble: 1 × 2
#>   value       sd
#>   <dbl>    <dbl>
#> 1   5.5 1.739253
#> # A tibble: 1 × 2
#>   value    sd
#>   <dbl> <dbl>
#> 1   5.5   Inf
#> # A tibble: 1 × 2
#>   value       sd
#>   <dbl>    <dbl>
#> 1   5.5 1.739253
#> # A tibble: 1 × 2
#>   value       sd
#>   <dbl>    <dbl>
#> 1   5.5 1.739253
#> # A tibble: 1 × 2
#>   value       sd
#>   <dbl>    <dbl>
#> 1   5.5 1.739253
#> # A tibble: 1 × 2
#>   value       sd
#>   <dbl>    <dbl>
#> 1   5.5 1.739253
#> # A tibble: 1 × 2
#>   value       sd
#>   <dbl>    <dbl>
#> 1   5.5 1.739253
#> # A tibble: 1 × 2
#>   value       sd
#>   <dbl>    <dbl>
#> 1   5.5 1.739253
#> # A tibble: 1 × 2
#>   value       sd
#>   <dbl>    <dbl>
#> 1   5.5 1.739253
#> # A tibble: 1 × 2
#>   value       sd
#>   <dbl>    <dbl>
#> 1   5.5 1.739253
#> # A tibble: 1 × 2
#>   value       sd
#>   <dbl>    <dbl>
#> 1   5.5 1.739253
#> # A tibble: 1 × 2
#>   value       sd
#>   <dbl>    <dbl>
#> 1   5.5 1.739253
#> # A tibble: 1 × 2
#>   value       sd
#>   <dbl>    <dbl>
#> 1   5.5 1.739253
#> # A tibble: 1 × 2
#>   value       sd
#>   <dbl>    <dbl>
#> 1   5.5 1.739253
#> # A tibble: 1 × 2
#>   value       sd
#>   <dbl>    <dbl>
#> 1   5.5 1.739253
#> # A tibble: 1 × 2
#>   value       sd
#>   <dbl>    <dbl>
#> 1   5.5 1.739253
#> # A tibble: 1 × 2
#>   value       sd
#>   <dbl>    <dbl>
#> 1   5.5 1.739253
#> # A tibble: 1 × 2
#>   value    sd
#>   <dbl> <dbl>
#> 1   5.5   Inf
#> # A tibble: 1 × 2
#>   value       sd
#>   <dbl>    <dbl>
#> 1   5.5 1.739253
#> # A tibble: 1 × 2
#>   value       sd
#>   <dbl>    <dbl>
#> 1   5.5 1.739253
#> # A tibble: 1 × 2
#>   value       sd
#>   <dbl>    <dbl>
#> 1   5.5 1.739253
#> # A tibble: 1 × 2
#>   value       sd
#>   <dbl>    <dbl>
#> 1   5.5 1.739253
#> # A tibble: 1 × 2
#>   value       sd
#>   <dbl>    <dbl>
#> 1   5.5 1.739253
#> # A tibble: 1 × 2
#>   value    sd
#>   <dbl> <dbl>
#> 1   5.5   Inf
#> # A tibble: 1 × 2
#>   value       sd
#>   <dbl>    <dbl>
#> 1   5.5 1.739253
#> # A tibble: 1 × 2
#>   value           sd
#>   <dbl>        <dbl>
#> 1   5.5 6.999517e+20
#> # A tibble: 1 × 2
#>   value       sd
#>   <dbl>    <dbl>
#> 1   5.5 1.739253
#> # A tibble: 1 × 2
#>   value       sd
#>   <dbl>    <dbl>
#> 1   5.5 1.739253
#> # A tibble: 1 × 2
#>   value       sd
#>   <dbl>    <dbl>
#> 1   5.5 1.739253
#> # A tibble: 1 × 2
#>   value       sd
#>   <dbl>    <dbl>
#> 1   5.5 1.739253

[krlmlr]

Similar results on Windows.

I think the hybrid handlers don't handle new summary variables very well. This looks similar to #2312.