statquant statquant - 11 months ago 50
R Question

cast string directly to IDateTime

I am using the new version of

and especially the AWESOME
function. My files contain dates that are loaded as strings (cause I don't know to do it otherwise) looking like

I need to sort the data.table on those datetimes and then for the sort to be efficient to cast then in the
format (or anything alse I would not know yet).

> strptime("01APR2008:09:00:00","%d%b%Y:%H:%M:%S")
[1] "2008-04-01 09:00:00"

> IDateTime(strptime("01APR2008:09:00:00","%d%b%Y:%H:%M:%S"))
idate itime
1: 2008-04-01 09:00:00

> IDateTime("01APR2008:09:00:00","%d%b%Y:%H:%M:%S")
Error in charToDate(x) :
character string is not in a standard unambiguous format

It looks like I cannot do
DT[ , newType := IDateTime(strptime(oldType, "%d%b%Y:%H:%M:%S"))]

My questions are then:

  1. Is there a way to cast directly to
    , such that I can sort afterward efficiently?

  2. If not, what is the most efficient way to go knowing that I would like to be able to sort DT by this datetime column

Answer Source

Unfortunately (for efficiency) strptime produces a POSIXlt type, which is unsupported by data.table and always will be due its size (40 bytes per date!) and structure. Although strftime produces the much better POSIXct, it still does it via POSIXlt. More info here :

Looking to base functions such as as.Date, it uses strptime too, creating an integer offset from epoch (oddly) stored as double. The IDate (and friends) class in data.table aims to achieve integer epoch offsets stored as, um, integer. Suitable for fast sorting by base::sort.list(method = "radix") (which is really a counting sort). IDate doesn't really aim to be fast at (usually one off) conversion.

So to convert string dates/times, rightly or wrongly, I tend to roll my own helper function.

If the string date is "2012-12-24" I'd lean towards: as.integer(gsub("-", "", col)) and proceed with YYYYMMDD integer dates. Similarly times can be HHMMDD as an integer. Two columns: date and time separately can be useful if you generally want to roll = TRUE within a day, but not to the previous day. Grouping by month is simple and fast: by = date%/%100L. Adding and subtracting days is troublesome, but it is anyway because rarely do you want to add calendar days, rather weekdays or business days. So that's a lookup to your business day vector anyway.

In your case the character month would need a conversion to 1:12. There isn't a separator in your dates "01APR2008", so a substring would be one way followed by a match or fmatch on the month name. Are you in control of the file format? If so, numbers are better in an unambiguous format that sorts naturally such as %Y-%m-%d, or %Y%m%d.

I haven't yet got to how best do this in fread, so date/times are left as character currently because I'm not yet sure how to detect the date format or which type to output. It does need to output either integer or double dates though, rather than inefficient character. I suspect that my use of YYYYMMDD integers are seen as unconventional, so I'm a little hesitant to make that the default. They have their place, and there are pros and cons of epoch based dates too. Dates don't have to be always epoch based is all I'm suggesting.

What do you think? Btw, thanks for encouragement on fread; was nice to see.