Make ...j the standard disambiguating suffix in name repair

[jennybc]

I think we should go ahead and use ...j as the standard suffix when .name_repair = "unique". These names don't have to be syntactic but it's so easy to make them so (they are already ugly and have lots of dots) and it will reduce the downstream burden of dealing with non-syntactic names (example: tidyverse/dplyr#4094).

This would be good to do ASAP I think in a tiny release. What do you think @krlmlr?

[krlmlr]

Can we instead recommend to use .name_repair = "universal" as a general rule, or .name_repair = "unique" and a mandatory cleaning of column names?

• I'd rather think carefully about another change of the behavior
• It won't help if the column name is a keyword
• Non-syntactic names prompt the user to do something about it

[jennybc]

I'm motivated by the case of empty names and a phenomenon I think will be fairly prevalent: the nonsyntactic names are short-lived or beside the point, but they prevent someone from doing something that seems totally unrelated. Like the dplyr issue above where, surprisingly, nonsyntactic names prevent mutate_all() from working.

[jennybc]

I agree it feels icky to make another change but we're still in early days of the tibble 2.x series and I suspect it's a net positive, long-term.

[krlmlr]

I still believe we can educate users to default to .name_repair = "universal", and use .name_repair = "unique" only if they are comfortable with non-syntactic names. The invariant here is that "unique" doesn't do more than necessary, and both "unique" and "universal" use ..# as disambiguator. (This holds no matter what the disambiguator is.) I don't think that another release is a good use of our time, even in the long run.

dplyr is being fixed and now also works for names like if, $ and o o p s .

[krlmlr]

tidyverse/dplyr#4094 seems difficult to fix for names like ..1. If we avoid creating them, that issue becomes much less of a problem.

How about a renaming strategy that sits between "unique" and "universal" and creates ...1 instead of ..1, and .... instead of ..., but does not touch names that contain keywords? The features of each repair strategy would be:

• Minimal: Proper names
• Unique: Unique names
• Quotable: Names can be used with backticks around them
• Universal: Names can be used without any backticks

What would quotable repair do to backticks?

library(tibble)
library(magrittr)

data <- tibble(`\`` = 1)
data %>% names()
#> [1] "`"
as_tibble(data, .name_repair = "unique")
#> # A tibble: 1 x 1
#>    `\``
#>   <dbl>
#> 1     1
as_tibble(data, .name_repair = "universal")
#> New names:
#> * `\`` -> .
#> # A tibble: 1 x 1
#>       .
#>   <dbl>
#> 1     1

^{Created on 2019-02-15 by the reprex package (v0.2.1.9000)}

[jennybc]

I still feel like there's time to just make three dots the standard suffix. I really think we should.

[nathancday]

I'm surprised to see this issue open after almost a month and 100% agree with @jennybc about fixing this ASAP. It is breaking things, I just opened an issue in tidycensus because of it walkerke/tidycensus#148.

The stated goal of tibble is in "keeping what time has proven to be effective, and throwing out what is not" to make a better data frame experience. This behavior seems counter to that goal and in general forcing people into invalid syntax feels anti-pattern for the tidyverse.

Why was the character prefix convention, eg X1, X2, abandoned in favor of dots? Could the default value be adjusted to 'universal', just to dodge the error messages?

[krlmlr]

@jennybc: Why three dots? Why not just one?

[krlmlr]

@nathancday: The issue you linked to seems unrelated to the problem discussed here. The particular prefix used is immaterial to the problem. The current default is safe, because it nudges the user to explicitly specify a name repair.

[jennybc]

Well, .1 is not syntactic either, so we gain nothing.

[jennybc]

Just to get it all written out, the reason to reconsider the "disambiguating numeric suffix" is that, in cases where the name was missing, it becomes the entire name. And even though we don't promise to make names syntactic when .name_repair = "unique", I think the world would probably be a better place if we did, for these cases.

[jennybc]

Recording a bit more motivation ...

We are happy for our automatically repaired names to be ugly, because that helps create awareness that the user is working with names created by someone else.

But in many cases, part of the reason people are oblivious -- and can safely remain so -- is that they were never going to use those weird variable names anyway. But the mere presence of non-syntactic names can cause friction in unexpected places, such as mutate_all(). So that's why I'd like to make the form of the suffix one that produces syntactic names (albeit ugly).

[krlmlr]

.1 is syntactic, but doesn't mean what I thought it means :-\

I hear you, and I agree now that changing the default suffix is the best way to handle that problem in the short-medium term.

The problem with ..1 and ... (and also with ``) is that these names not only not syntactic, they're invalid even as a quoted name. That's worse -- mutate_all() doesn't have a problem with unquoted names, but does with non-quotable names. Maybe we can even support a naming strategy that's conceptually between "minimal" and "unique", and make that naming strategy the default. (This means that all empty names become a dot -- that's not too bad for tibbles IMO.)

I'll implement the changes and see how many revdep failures we have. If we can't fix all downstream failures in the very short term, I'd rather add a compatibility mode.

[jennybc]

.1 is syntactic, but doesn't mean what I thought it means :-\

Can you spell this out more? It's so easy to overlook weird stuff and edge cases in this naming domain, that I want to make sure I'm holding the important facts in my head.

[krlmlr]

It's a number. That's why .1x is forbidden, too.

.1 <- 5
#> Error in 0.1 <- 5: invalid (do_set) left-hand side to assignment
a <- .1
a
#> [1] 0.1

^{Created on 2019-02-18 by the reprex package (v0.2.1.9000)}