toJSON(auto_unbox=TRUE, force=TRUE) should not unbox objects wrapped in I()

[wch]

For compatibility with RJSONIO, objects wrapped in I() shouldn't be unboxed, even when auto_unbox=TRUE.

For example:

RJSONIO:::toJSON(list(x = I("text")))
# [1] "{\n \"x\": [ \"text\" ] \n}"

jsonlite:::toJSON(list(x = I("text")), auto_unbox = TRUE, force = TRUE)
# {"x":"text"} 

@jeroenooms would this just involve writing a new method for AsIs objects?

This is related to timelyportfolio/parcoords#6.

[yihui]

+1

I like the I() notation to box scalars in RJSONIO. It is very easy to opt-in, since I() is a one-letter function in base R. By comparison, it is relatively expensive to opt-out boxing in jsonlite when auto_unbox = FALSE: one need to type jsonlite::unbox(), and this can be cumbersome when it has to be done multiple times, e.g.

datatable(iris, options = list(order = list(
  list(jsonlite::unbox(1), jsonlite::unbox('asc')),
  list(jsonlite::unbox(2), jsonlite::unbox('desc'))
)))

On the other hand, when auto_unbox = TRUE, currently there is no way to opt-in boxing in jsonlite if we do want to box certain scalars but not all of them. It will be nice if the class AsIs is supported.

[jeroen]

The only thing that makes me a little nervous is that we introduce undefined behavior for I() objects that are not scalars.

[wch]

Can you give an example? In my mind it seems pretty clear: I() only affects the behavior if the content is a length-1 scalar (atomic vector).

For things like I(list(3)) or I(1:3), the content of I() is not a length-1 atomic vector, and therefore the I() has no effect. I tried to get at that in the tests in my PR.

[wch]

Also, if you'd prefer an option for this behavior, I'm amenable to that as well -- although to me it doesn't seem necessary since auto_unbox, I(), and a scalar would be a rare combination to see by chance.

[wch]

Oh, I just realized that you have a class named scalar in jsonlite. Is that what you're referring to?

[jeroen]

Sorry I was confused. I meant the case of I(1:3) where the user can use I() when it is not appropriate. But it is harmless.

[yihui]

Exactly. It is harmless.

[vspinu]

Big thanks for this, but it should be documented. Maybe in unbox help page.