Fast CSV writer

As write.csv but much faster (e.g. 2 seconds versus 1 minute) and just as flexible. Modern machines almost surely have more than one CPU so fwrite uses them; on all operating systems including Linux, Mac and Windows.

fwrite(x, file = "", append = FALSE, quote = "auto",
  sep = ",", sep2 = c("","|",""),
  eol = if (.Platform$OS.type=="windows") "\r\n" else "\n",
  na = "", dec = ".", row.names = FALSE, col.names = TRUE,
  qmethod = c("double","escape"),
  logical01 = getOption("datatable.logical01", FALSE),  # due to change to TRUE; see NEWS
  logicalAsInt = logical01,  # deprecated
  scipen = getOption('scipen', 0L),
  dateTimeAs = c("ISO","squash","epoch","write.csv"),
  buffMB = 8L, nThread = getDTthreads(verbose),
  showProgress = getOption("datatable.showProgress", interactive()),
  compress = c("auto", "none", "gzip"),
  yaml = FALSE,
  bom = FALSE,
  verbose = getOption("datatable.verbose", FALSE))

Arguments

x	Any list of same length vectors; e.g. data.frame and data.table. If matrix, it gets internally coerced to data.table preserving col names but not row names
file	Output file name. "" indicates output to the console.
append	If TRUE, the file is opened in append mode and column names (header row) are not written.
quote	When "auto", character fields, factor fields and column names will only be surrounded by double quotes when they need to be; i.e., when the field contains the separator sep, a line ending \n, the double quote itself or (when list columns are present) sep2[2] (see sep2 below). If FALSE the fields are not wrapped with quotes even if this would break the CSV due to the contents of the field. If TRUE double quotes are always included other than around numeric fields, as write.csv.
sep	The separator between columns. Default is ",".
sep2	For columns of type list where each item is an atomic vector, sep2 controls how to separate items within the column. sep2[1] is written at the start of the output field, sep2[2] is placed between each item and sep2[3] is written at the end. sep2[1] and sep2[3] may be any length strings including empty "" (default). sep2[2] must be a single character and (when list columns are present and therefore sep2 is used) different from both sep and dec. The default (|) is chosen to visually distinguish from the default sep. In speaking, writing and in code comments we may refer to sep2[2] as simply "sep2".
eol	Line separator. Default is "\r\n" for Windows and "\n" otherwise.
na	The string to use for missing values in the data. Default is a blank string "".
dec	The decimal separator, by default ".". See link in references. Cannot be the same as sep.
row.names	Should row names be written? For compatibility with data.frame and write.csv since data.table never has row names. Hence default FALSE unlike write.csv.
col.names	Should the column names (header row) be written? The default is TRUE for new files and when overwriting existing files (append=FALSE). Otherwise, the default is FALSE to prevent column names appearing again mid-file when stacking a set of data.tables or appending rows to the end of a file.
qmethod	A character string specifying how to deal with embedded double quote characters when quoting strings. "escape" - the quote character (as well as the backslash character) is escaped in C style by a backslash, or "double" (default, same as write.csv), in which case the double quote is doubled with another one.
logical01	Should logical values be written as 1 and 0 rather than "TRUE" and "FALSE"?
logicalAsInt	Deprecated. Old name for `logical01`. Name change for consistency with `fread` for which `logicalAsInt` would not make sense.
scipen	integer In terms of printing width, how much of a bias should there be towards printing whole numbers rather than scientific notation? See Details.
dateTimeAs	How Date/IDate, ITime and POSIXct items are written. "ISO" (default) - 2016-09-12, 18:12:16 and 2016-09-12T18:12:16.999999Z. 0, 3 or 6 digits of fractional seconds are printed if and when present for convenience, regardless of any R options such as digits.secs. The idea being that if milli and microseconds are present then you most likely want to retain them. R's internal UTC representation is written faithfully to encourage ISO standards, stymie timezone ambiguity and for speed. An option to consider is to start R in the UTC timezone simply with "$ TZ='UTC' R" at the shell (NB: it must be one or more spaces between TZ='UTC' and R, anything else will be silently ignored; this TZ setting applies just to that R process) or Sys.setenv(TZ='UTC') at the R prompt and then continue as if UTC were local time. "squash" - 20160912, 181216 and 20160912181216999. This option allows fast and simple extraction of yyyy, mm, dd and (most commonly to group by) yyyymm parts using integer div and mod operations. In R for example, one line helper functions could use %/%10000, %/%100%%100, %%100 and %/%100 respectively. POSIXct UTC is squashed to 17 digits (including 3 digits of milliseconds always, even if 000) which may be read comfortably as integer64 (automatically by fread()). "epoch" - 17056, 65536 and 1473703936.999999. The underlying number of days or seconds since the relevant epoch (1970-01-01, 00:00:00 and 1970-01-01T00:00:00Z respectively), negative before that (see ?Date). 0, 3 or 6 digits of fractional seconds are printed if and when present. "write.csv" - this currently affects POSIXct only. It is written as write.csv does by using the as.character method which heeds digits.secs and converts from R's internal UTC representation back to local time (or the "tzone" attribute) as of that historical date. Accordingly this can be slow. All other column types (including Date, IDate and ITime which are independent of timezone) are written as the "ISO" option using fast C code which is already consistent with write.csv. The first three options are fast due to new specialized C code. The epoch to date-part conversion uses a fast approach by Howard Hinnant (see references) using a day-of-year starting on 1 March. You should not be able to notice any difference in write speed between those three options. The date range supported for Date and IDate is [0000-03-01, 9999-12-31]. Every one of these 3,652,365 dates have been tested and compared to base R including all 2,790 leap days in this range. This option applies to vectors of date/time in list column cells, too. A fully flexible format string (such as "%m/%d/%Y") is not supported. This is to encourage use of ISO standards and because that flexibility is not known how to make fast at C level. We may be able to support one or two more specific options if required.
buffMB	The buffer size (MB) per thread in the range 1 to 1024, default 8MB. Experiment to see what works best for your data on your hardware.
nThread	The number of threads to use. Experiment to see what works best for your data on your hardware.
showProgress	Display a progress meter on the console? Ignored when file=="".
compress	If compress = "auto" and if file ends in .gz then output format is gzipped csv else csv. If compress = "none", output format is always csv. If compress = "gzip" then format is gzipped csv. Output to the console is never gzipped even if compress = "gzip". By default, compress = "auto".
yaml	If TRUE, fwrite will output a CSVY file, that is, a CSV file with metadata stored as a YAML header, using as.yaml. See Details.
bom	If TRUE a BOM (Byte Order Mark) sequence (EF BB BF) is added at the beginning of the file; format 'UTF-8 with BOM'.
verbose	Be chatty and report timings?

Details

fwrite began as a community contribution with pull request #1613 by Otto Seiskari. This gave Matt Dowle the impetus to specialize the numeric formatting and to parallelize: http://blog.h2o.ai/2016/04/fast-csv-writing-for-r/. Final items were tracked in issue #1664 such as automatic quoting, bit64::integer64 support, decimal/scientific formatting exactly matching write.csv between 2.225074e-308 and 1.797693e+308 to 15 significant figures, row.names, dates (between 0000-03-01 and 9999-12-31), times and sep2 for list columns where each cell can itself be a vector.

To save space, fwrite prefers to write wide numeric values in scientific notation -- e.g. 10000000000 takes up much more space than 1e+10. Most file readers (e.g. fread) understand scientific notation, so there's no fidelity loss. Like in base R, users can control this by specifying the scipen argument, which follows the same rules as options('scipen'). fwrite will see how much space a value will take to write in scientific vs. decimal notation, and will only write in scientific notation if the latter is more than scipen characters wider. For 10000000000, then, 1e+10 will be written whenever scipen<6.

CSVY Support:

The following fields will be written to the header of the file and surrounded by --- on top and bottom:

• source - Contains the R version and data.table version used to write the file

• creation_time_utc - Current timestamp in UTC time just before the header is written

• schema with element fields giving name-type (class) pairs for the table; multi-class objects (e.g. c('POSIXct', 'POSIXt')) will have their first class written.

• header - same as col.names (which is header on input)

• sep

• sep2

• eol

• na.strings - same as na

• dec

• qmethod

• logical01

See also

setDTthreads, fread, write.csv, write.table, bit64::integer64

References

http://howardhinnant.github.io/date_algorithms.html
https://en.wikipedia.org/wiki/Decimal_mark

Examples

DF = data.frame(A=1:3, B=c("foo","A,Name","baz"))
fwrite(DF)
#> A,B
#> 1,foo
#> 2,"A,Name"
#> 3,baz
write.csv(DF, row.names=FALSE, quote=FALSE)  # same
#> A,B
#> 1,foo
#> 2,A,Name
#> 3,baz

fwrite(DF, row.names=TRUE, quote=TRUE)
#> "","A","B"
#> "1",1,"foo"
#> "2",2,"A,Name"
#> "3",3,"baz"
write.csv(DF)                                # same
#> "","A","B"
#> "1",1,"foo"
#> "2",2,"A,Name"
#> "3",3,"baz"

DF = data.frame(A=c(2.1,-1.234e-307,pi), B=c("foo","A,Name","bar"))
fwrite(DF, quote='auto')        # Just DF[2,2] is auto quoted
#> A,B
#> 2.1,foo
#> -1.234e-307,"A,Name"
#> 3.14159265358979,bar
write.csv(DF, row.names=FALSE)  # same numeric formatting
#> "A","B"
#> 2.1,"foo"
#> -1.234e-307,"A,Name"
#> 3.14159265358979,"bar"

DT = data.table(A=c(2,5.6,-3),B=list(1:3,c("foo","A,Name","bar"),round(pi*1:3,2)))
fwrite(DT)
#> A,B
#> 2,1|2|3
#> 5.6,foo|"A,Name"|bar
#> -3,3.14|6.28|9.42
fwrite(DT, sep="|", sep2=c("{",",","}"))
#> A|B
#> 2|{1,2,3}
#> 5.6|{foo,"A,Name",bar}
#> -3|{3.14,6.28,9.42}

if (FALSE) {

set.seed(1)
DT = as.data.table( lapply(1:10, sample,
         x=as.numeric(1:5e7), size=5e6))                            #     382MB
system.time(fwrite(DT, "/dev/shm/tmp1.csv"))                        #      0.8s
system.time(write.csv(DT, "/dev/shm/tmp2.csv",                      #     60.6s
                      quote=FALSE, row.names=FALSE))
system("diff /dev/shm/tmp1.csv /dev/shm/tmp2.csv")                  # identical

set.seed(1)
N = 1e7
DT = data.table(
  str1=sample(sprintf("%010d",sample(N,1e5,replace=TRUE)), N, replace=TRUE),
  str2=sample(sprintf("%09d",sample(N,1e5,replace=TRUE)), N, replace=TRUE),
  str3=sample(sapply(sample(2:30, 100, TRUE), function(n)
     paste0(sample(LETTERS, n, TRUE), collapse="")), N, TRUE),
  str4=sprintf("%05d",sample(sample(1e5,50),N,TRUE)),
  num1=sample(round(rnorm(1e6,mean=6.5,sd=15),2), N, replace=TRUE),
  num2=sample(round(rnorm(1e6,mean=6.5,sd=15),10), N, replace=TRUE),
  str5=sample(c("Y","N"),N,TRUE),
  str6=sample(c("M","F"),N,TRUE),
  int1=sample(ceiling(rexp(1e6)), N, replace=TRUE),
  int2=sample(N,N,replace=TRUE)-N/2
)                                                                   #     774MB
system.time(fwrite(DT,"/dev/shm/tmp1.csv"))                         #      1.1s
system.time(write.csv(DT,"/dev/shm/tmp2.csv",                       #     63.2s
                      row.names=FALSE, quote=FALSE))
system("diff /dev/shm/tmp1.csv /dev/shm/tmp2.csv")                  # identical

unlink("/dev/shm/tmp1.csv")
unlink("/dev/shm/tmp2.csv")
}