findGlobals() incorrectly detects 'x' as a global in subset(df, x < 3)

[HenrikBengtsson]

Problem

findGlobals() incorrectly detects x as a global in subset(df, x < 3), e.g.

> library(globals)
> df <- data.frame(x=1:5, y=(1:5)^2)
> findGlobals(substitute({ subset(df, x < 3) }))
[1] "{"      "<"      "df"     "subset" "x"
}

This is a problem with codetools::findGlobals() that globals uses internally.

Workaround

The workaround is to not use subset(), e.g.

> library(globals)
> df <- data.frame(x=1:5, y=(1:5)^2)
> findGlobals(substitute({ keep <- (df$x < 3); df[keep,] }))
[1] "$"  "("  "["  "{"  "<"  "<-" "df"

or to create fake local variables, e.g.

> library(globals)
> df <- data.frame(x=1:5, y=(1:5)^2)
> findGlobals(substitute({ x <- NULL; subset(df, x < 3) }))
[1] "{"      "<"      "<-"     "df"     "subset"
}

Possible approach

Maybe it is possible to add some type of metadata to subset() that give enough hints for findGlobals() to conclude that x could be part of the df object and/or the subset argument is passed without being evaluated. For instance,

argument_hint(subset, "subset", as="substitute")

such that findGlobals() will known that argument subset of subset() is passed as a "substituted" expression. This could then give findGlobals() the option to ignore whatever is passed to argument subset. With even more information, we might also be able to guide findGlobals() to - at least at run time - validate that x is available from either first argument (df) or the calling environment, e.g.

argument_hint(subset, "subset", where=list(argument=1L, "parent.frame"))

Now, could this be inferred from code inspection of subset() itself?

[HenrikBengtsson]

The codetools package uses so called usage handlers internally to handle this for common cases, e.g. base::library(tools), cf. Issue #12.

[HenrikBengtsson]

This works for futures since future (>= 1.0.0) because it now takes an optimistic approach to globals, i.e. it accepts false-positive globals (including non-existing ones as here).

Closing this since the globals package mostly exists such that the future package works.