Expressions using `<-` can overwrite active bindings in the data mask

[DavisVaughan]

Seen in metacore

library(dplyr, warn.conflicts = FALSE)

foo <- tibble(
   x = c("CHG", "CHG", "CHG", "ABLFL", "ABLFL"),
   y = c("float", "integer", "float", "text", "text")
)
foo <- group_by(foo, x)

mutate(foo, y = {
   attr(y, "label") <- NULL
   y
})
#> Error in `mutate()`:
#> ℹ In argument: `y = { ... }`.
#> ℹ In group 2: `x = "CHG"`.
#> Caused by error:
#> ! `y` must be size 3 or 1, not 2.

#> Backtrace:
#>      ▆
#>   1. ├─dplyr::mutate(...)
#>   2. ├─dplyr:::mutate.data.frame(...) at dplyr/R/mutate.R:146:2
#>   3. │ └─dplyr:::mutate_cols(.data, dplyr_quosures(...), by) at dplyr/R/mutate.R:181:2
#>   4. │   ├─base::withCallingHandlers(...) at dplyr/R/mutate.R:268:2
#>   5. │   └─dplyr:::mutate_col(dots[[i]], data, mask, new_columns) at dplyr/R/mutate.R:273:6
#>   6. │     └─mask$eval_all_mutate(quo) at dplyr/R/mutate.R:380:8
#>   7. │       └─dplyr (local) eval() at dplyr/R/data-mask.R:69:6
#>   8. ├─dplyr:::dplyr_internal_error(...)
#>   9. │ └─rlang::abort(class = c(class, "dplyr:::internal_error"), dplyr_error_data = data) at dplyr/R/conditions.R:194:2
#>  10. │   └─rlang:::signal_abort(cnd, .file)
#>  11. │     └─base::signalCondition(cnd)
#>  12. └─dplyr (local) `<fn>`(`<dpl:::__>`)
#>  13.   └─rlang::abort(message, class = error_class, parent = parent, call = error_call) at dplyr/R/conditions.R:235:4

^{Created on 2023-01-26 with reprex v2.0.2.9000}

[DavisVaughan]

Simpler example, still weird. Seems like <- is messing up the active binding or something

library(dplyr, warn.conflicts = FALSE)

foo <- tibble(
  x = c("a", "b"),
  y = c("c", "d")
)
foo <- group_by(foo, x)

mutate(foo, y = {
  y <- paste0(y, "x")
  y
})
#> # A tibble: 2 × 2
#> # Groups:   x [2]
#>   x     y    
#>   <chr> <chr>
#> 1 a     cx   
#> 2 b     cxx

^{Created on 2023-01-26 with reprex v2.0.2.9000}

[DavisVaughan]

Both of these examples actually fail in CRAN dplyr too. They are appearing now because we now more aggressively inline anonymous functions in across().

i.e. these two are now equivalent in 1.1.0 and both fail

library(dplyr, warn.conflicts = FALSE)

foo <- tibble(
  x = c("a", "b"),
  y = c("c", "d")
)
foo <- group_by(foo, x)

# this:
mutate(foo, across(everything(), function(col) {
  col <- paste0(col, "x")
  col
}))

# now gets inlined now as this in 1.1.0:
mutate(foo, y = {
  y <- paste0(y, "x")
  y
})

The across() version doesn't fail in CRAN dplyr because the anonymous function provides a function environment that acts as a buffer between the <- assignment and the data mask. So the <- assigns into the function env, not the mask.

---

Ultimately this is a side effect of this behavior, which is documented in ?eval_tidy

Lexical side effects, such as assignment with ⁠<-⁠, occur in the mask rather than env.

library(rlang)

df <- data.frame(x = 1:2)

data <- as_data_mask(df)

# nothing here
data$x
#> NULL

# it is here instead, as it should be
data$.top_env$x
#> [1] 1 2

eval_tidy(data = data, expr({
  x <- paste0(x, "x")
  x
}))
#> [1] "1x" "2x"

# that evaluated directly in the mask, so now this exists
data$x
#> [1] "1x" "2x"

# So `data$x` masks `data$.top_env$x` now
eval_tidy(data = data, expr({
  x
}))
#> [1] "1x" "2x"

[DavisVaughan]

@lionel- I'm definitely going to need some advice on how to resolve this correctly.

[DavisVaughan]

• We need to change how anonymous functions are inlined. i.e. we don't inline if we detect <-, =, <<- anywhere in the body (this would fix this exact issue)
• We need to change add_mask_binding() to work off a function that takes advantage of the fact that we can tell when an active binding is being assigned to, and we can throw an error in that case (this would ensure an error is thrown in the case where a user uses { but not an anonymous function)

---

We can also think about changing dplyr_data_masks_setup() so that a fresh data mask is instead created before each group evaluation so that any lexical scoping changes don't bleed across groups