Internal: slightly inaccurate as_join_by() implementation for character

[krlmlr]

waldo::compare(dplyr:::as_join_by("a"), dplyr::join_by(a == a))
#> `old$exprs[[1]]`: `"a" == "a"`
#> `new$exprs[[1]]`: `a == a`

^{Created on 2024-03-08 with reprex v2.1.0}

I can work around, but I wonder if this may cause some downstream gotchas.

[DavisVaughan]

We do allow this though:

unclass(dplyr::join_by("a" == "a"))
#> $exprs
#> $exprs[[1]]
#> "a" == "a"
#> 
#> 
#> $condition
#> [1] "=="
#> 
#> $filter
#> [1] "none"
#> 
#> $x
#> [1] "a"
#> 
#> $y
#> [1] "a"

But maybe that should be the "special" case, and coming from character should try to generate the "typical" case of a == a

[krlmlr]

What are the semantics of dplyr::join_by("a" == "a") ? Shouldn't the result be identical to dplyr::join_by(a == a) ?

[psychelzh]

Out of curiosity, will as_join_by() be exported? Currently, I found join_by() is kind of hard to program with, for example, if I have stored the column names in a vector. For example:

suppressPackageStartupMessages(library(dplyr))
x <- data.frame(a = 1:5, b = 2:6)
y <- data.frame(a = 1:3, b = c(2, 3, 5), x = 2:4)
cols_by <- c("a", "b")
inner_join(x, y, by = cols_by)
#>   a b x
#> 1 1 2 2
#> 2 2 3 3
# oops
inner_join(x, y, by = join_by(!!cols_by))
#> Error in `join_by()`:
#> ! Each element of `...` must be a single column name or a join by
#>   expression.
#> ✖ Element 1 is not a name and not an expression.
inner_join(x, y, by = dplyr:::as_join_by(cols_by))
#>   a b x
#> 1 1 2 2
#> 2 2 3 3

^{Created on 2024-03-20 with reprex v2.1.0}

[krlmlr]

Do you need join_by(!!!syms(cols_by)) ?

[psychelzh]

Ah, thanks! !!! works as expected. And without syms works too (the docs of join_by() tell us these can be quoted or not quoted).

suppressPackageStartupMessages(library(dplyr))
x <- data.frame(a = 1:5, b = 2:6)
y <- data.frame(a = 1:3, b = c(2, 3, 5), x = 2:4)
cols_by <- c("a", "b")
inner_join(x, y, by = join_by(!!!cols_by))
#>   a b x
#> 1 1 2 2
#> 2 2 3 3

^{Created on 2024-03-20 with reprex v2.1.0}

But I do think we do not recommend using !! or !!! a lot, do we? So, I am wondering if there will be some helpers without knowing quasiquotation or unqutation.

[DavisVaughan]

if I have stored the column names in a vector

The old method is not going away, so if you have this case then I'd recommend just supplying the character vector as you did before join_by()

But using !!! is also totally fine here, the !!enquo() pattern is replaced by {{ }} but using !!! is still somewhat common https://rlang.r-lib.org/reference/topic-metaprogramming.html

---

as_join_by() is only for creating equi joins, and is basically a nicely named internal function. I don't intend for us to export it since it has such a limited use case

[DavisVaughan]

What are the semantics of dplyr::join_by("a" == "a") ? Shouldn't the result be identical to dplyr::join_by(a == a) ?

They are identical yes

[krlmlr]

Should we support the invariant that two joins are identical iff their join_by() arguments are identical? I'm still advocating for

identical(dplyr::join_by("a" == "a"), dplyr::join_by(a == a))

because it will simplify code that analyzes join_by objects.