Could we use one set of 6 like this, and then pre-process the strings with a regex like was suggested in https://issues.apache.org/jira/browse/ARROW-14471?focusedCommentId=17446011&page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel#comment-17446011 ?

Something kind of like:

format_map <- list(
        ymd_hyphen1 = "%Y-%m-%d",
        ymd_hyphen2 = "%y-%m-%d",
        ymd_hyphen3 = "%Y-%B-%d",
        ymd_hyphen4 = "%y-%B-%d",
        ymd_hyphen5 = "%Y-%b-%d",
        ymd_hyphen6 = "%y-%b-%d"
)

x <- gsub("[^A-Za-z0-9.]", "-", x)
call_binding(
      "coalesce",
      call_binding("strptime", x, format = format_map[[1]], unit = "s"),
      call_binding("strptime", x, format = format_map[[2]], unit = "s"),
      call_binding("strptime", x, format = format_map[[3]], unit = "s"),
      call_binding("strptime", x, format = format_map[[4]], unit = "s"),
      call_binding("strptime", x, format = format_map[[5]], unit = "s"),
      call_binding("strptime", x, format = format_map[[6]], unit = "s")
    )

You might need to use call_binding("gsub", ...) instead of being able to use it directly

This would at least cut down on the number of formats we need to try and mostly get the right answers

Reducing the formats works (I might have to implement it in several steps to handle " " and multiple instances of the same separator). There are a couple of outstanding issues:

• coalesce() needs the first call to call_binding("strptime") to return NA in order to move on to the second one and so on. Instead the first call errors and the second one never get evaluated. tryCatch() doesn't seem to work in this context.
• short ("%y") vs long ("%Y") years. If the string contains a short year (e.g. "19") libarrrow will not error for the first format, but instead return 0019.

I think I will move the focus on parse_date_time() of which most other date/time parsing functions are a particular case of. ymd(x) becomes parse_date_time(x, orders = "ymd"), but the problems highlighted here still remain.

coalesce() needs the first call to call_binding("strptime") to return NA in order to move on to the second one and so on. Instead the first call errors and the second one never get evaluated. tryCatch() doesn't seem to work in this context.

Good point! Is there a jira to make that a possibility for the strptime binding (or in the C++ itself, which is the ultimate place it'll really need to be...)?

https://issues.apache.org/jira/browse/ARROW-15659, but this is for the R side only.

We should also have a check that if the format fails, we get NAs out — so long as that's possible)

It looks like right now, it throws an error for the whole conversion: https://github.com/apache/arrow/runs/5094263843?check_suite_focus=true#step:6:22119