Hex floating-point numbers no longer parse in 1.10.5

[scottstanfield]

We know data.table can read floating point numbers in a CSV file just fine. What surprised me is that data.table 1.10.4 could read the Java specific hexadecimal floating point numbers.

If you're not familiar with them, they look like 0x1.21d6353p-1 which is roughly 0.56608. The JDK section 3.10.2 has more details.

Data.table 1.10.4 reads them just fine. But 1.10.5 reads them as a string.

My source CSV file is a simple POJO file from H2O.ai. Three columns: the prediction target (true/false), and the probability of either result (columns FALSE and TRUE). Sample file attached.

Steps to reproduce are trivial, as you can see in the attached screenshots. I'm expecting 1.10.5 to produce the same results as 1.10.4. I don't have a workaround yet.

data.table 1.10.5 (current top-of-tree as of 2017-08-21)

data.table 1.10.4

The CSV file (had to upload it as a .TXT file)
pojo.txt

[scottstanfield]

As I'm poking around the source to see when this regressed, I noticed a comment in foverlap.r

# Conclusion: floating point manipulations are hell!

[scottstanfield]

My work-around in 1.10.5. Since my column names are also reserved-words in R (yeah, I know), I use column positions. Always happy to play code-golf and find a shorter or faster way to do this.

# Works in 1.10.5 (and in 1.10.4, but not necessary)

cols <- c(2, 3)  # My FALSE and TRUE columns number 2 and 3
d[, (cols) := lapply(.SD, as.numeric), .SDcols=cols]

[st-pasha]

The workaround is no longer needed: fread is now able to parse these numbers directly