revisit: transpose() doesn't work as expected when inner keys are disjoint

[jennybc]

Deliberately repeating name of #205. I'm not sure #205 really was a duplicate of #164.

transpose() seems to work just fine with names, so this is not about position. But it only seems to consult the first object to determine inner keys.

library(purrr)

ll <- list(
  list(x = 1, y = "one"),
  list(z = "deux", x = 2)
)

## only x and y appear, but what about z?
transpose(ll)
#> $x
#> $x[[1]]
#> [1] 1
#> 
#> $x[[2]]
#> [1] 2
#> 
#> 
#> $y
#> $y[[1]]
#> [1] "one"
#> 
#> $y[[2]]
#> NULL

## I expected this
list(
  x = map(ll, "x"),
  y = map(ll, "y"),
  z = map(ll, "z")
)
#> $x
#> $x[[1]]
#> [1] 1
#> 
#> $x[[2]]
#> [1] 2
#> 
#> 
#> $y
#> $y[[1]]
#> [1] "one"
#> 
#> $y[[2]]
#> NULL
#> 
#> 
#> $z
#> $z[[1]]
#> NULL
#> 
#> $z[[2]]
#> [1] "deux"

[hadley]

Have a look at the new .names argument 😉

[jennybc]

OK great!

The lists I'm dealing with are large and I don't know in advance how all the children are named. So I would have to pre-map names(), reduce with union, and then provide that to .names. It feels like it would be handy to get some help from a default along those lines.

Here's how I am doing this task by hand right now (not a reprex, sorry ... just copy/paste). I make some effort to get a decent global order for the names:

## determine the names of sub-items and a consensus order
## get vector of names for each endpoint
nms_list <- endpoints %>%
  map(names)
## union of names across endpoints
nms_glop <- nms_list %>%
  reduce(union)
## reach a consensus on name order
nms_rksum <- nms_list %>%
  map_dfc(~ match(nms_glop, .x)) %>% 
  rowSums(na.rm = TRUE)
nms <- nms_glop[order(nms_rksum)]

[hadley]

What's the principle behind the global order?

[jennybc]

It's the order most consistent with how the names appear in the children of the input list, based on rank sum. If they always appear in same order, it will be that. In my current example, there is variation across the children in which names appear at all and in their order.

A simpler approach would be to just take the unique values from the union of the names. I had to reorder mine manually anyway, so maybe it's not worth trying to be clever.

[hadley]

Yeah, I'd think x %>% map(names) %>% reduce(union) would usually be sufficient. Maybe that's worth including as an example?

[jennybc]

Maybe that's worth including as an example?

If it's not destined to be the default, then yes I think that would be great. Do you want me to do in a PR?

[hadley]

Yes please!