[R-Forge #2660] Improve fread na.strings handling

[arunsrinivasan]

Submitted by: Matt Dowle; Assigned to: Nobody; R-Forge link

As raised here and here on SO.

[arunsrinivasan]

:bump:

[ghost]

I was going to post this on SO but I found this issue so I'll instead provide a screenshot for the fread vs. read.table results - it's the main reason I've put off learning to use data.table. It's a pretty common data format with a few missing values in numeric columns. I wonder if people have been making sure to check their column classes after using fread:

[dselivanov]

@arunsrinivasan, @mattdowle today I had a look into this.
Seems Strto*() functions family from fread.c should care about na.strings parameter. As of now the only one case is handled - "NA" strings. And the way this case handled is not very agile.

if (lch==start && lch<eof-1 && *lch++=='N' && *lch++=='A' && (lch==eof || *lch==sep || *lch==eol)) {
        ch = lch;
        u.d = NA_REAL;
        return(TRUE);
    }

If you agree that this is work for these functions (not just post-processing at the end readfile() function ) , I have ideas how to fix this.

[arunsrinivasan]

Thanks again for the excellent fix @dselivanov. Opened another issue, #1314, for na.strings="-999" like cases.