Faster way to read fixed-width files

Question

I work with a lot of fixed width files (i.e., no separating character) that I need to read into R. So, there is usually a definition of the column width to parse the string into variables. I can use read.fwf to read in the data without a problem. However, for large files, this can take a long time. For a recent dataset, this took 800 seconds to read in a dataset with ~500,000 rows and 143 variables.

seer9 <- read.fwf("~/data/rawdata.txt", 
  widths = cols,
  header = FALSE,
  buffersize = 250000,
  colClasses = "character",
  stringsAsFactors = FALSE))

fread in the data.table package in R is awesome for solving most data read problems, except it doesn't parse fixed width files. However, I can read each line in as a single character string (~500,000 rows, 1 column). This takes 3-5 seconds. (I love data.table.)

seer9 <- fread("~/data/rawdata.txt", colClasses = "character",
               sep = "\n", header = FALSE, verbose = TRUE)

There are a number of good posts on SO on how to parse text files. See JHoward's suggestion here, to create a matrix of start and end columns, and substr to parse the data. See GSee's suggestion here to use strsplit. I couldn't figure out how to make that work with this data. (Also, Michael Smith made some suggestions on the data.table mailing list involving sed that were beyond my ability to implement.) Now, using fread and substr() I can do the whole thing in about 25-30 seconds. Note that coercing to a data.table at end takes a chunk of time (5 sec?).

end_col <- cumsum(cols)
start_col <- end_col - cols + 1
start_end <- cbind(start_col, end_col) # matrix of start and end positions
text <- lapply(seer9, function(x) {
        apply(start_end, 1, function(y) substr(x, y[1], y[2])) 
        })
dt <- data.table(text$V1)
setnames(dt, old = 1:ncol(dt), new = seervars)

What I am wondering is whether this can be improved any further? I know I am not the only one who has to read fixed width files, so if this could be made faster, it would make loading even larger files (with millions of rows) more tolerable. I tried using parallel with mclapply and data.tableinstead of lapply, but those didn't change anything. (Likely due to my inexperience in R.) I imagine that an Rcpp function could be written to do this really fast, but that is beyond my skill set. Also, I may not be using lapply and apply appropriately.

My data.table implementation (with magrittr chaining) takes the same time:

text <- seer9[ , apply(start_end, 1, function(y) substr(V1, y[1], y[2]))] %>% 
  data.table(.)

Can anyone make suggestions to improve the speed of this? Or is this about as good as it gets?

Here is code to create a similar data.table within R (rather than linking to actual data). It should have 331 characters, and 500,000 rows. There are spaces to simulate missing fields in the data, but this is NOT space delimited data. (I am reading raw SEER data, in case anyone is interested.) Also including column widths (cols) and variable names (seervars) in case this helps someone else. These are the actual column and variable definitions for SEER data.

seer9 <-
  data.table(rep((paste0(paste0(letters, 1000:1054, " ", collapse = ""), " ")),
                 500000))

cols = c(8,10,1,2,1,1,1,3,4,3,2,2,4,4,1,4,1,4,1,1,1,1,3,2,2,1,2,2,13,2,4,1,1,1,1,3,3,3,2,3,3,3,3,3,3,3,2,2,2,2,1,1,1,1,1,6,6,6,2,1,1,2,1,1,1,1,1,2,2,1,1,2,1,1,1,1,1,1,1,1,1,1,1,1,1,1,7,5,4,10,3,3,2,2,2,3,1,1,1,1,2,2,1,1,2,1,9,5,5,1,1,1,2,2,1,1,1,1,1,1,1,1,2,3,3,3,3,3,3,1,4,1,4,1,1,3,3,3,3,2,2,2,2)
seervars <- c("CASENUM", "REG", "MAR_STAT", "RACE", "ORIGIN", "NHIA", "SEX", "AGE_DX", "YR_BRTH", "PLC_BRTH", "SEQ_NUM", "DATE_mo", "DATE_yr", "SITEO2V", "LATERAL", "HISTO2V", "BEHO2V", "HISTO3V", "BEHO3V", "GRADE", "DX_CONF", "REPT_SRC", "EOD10_SZ", "EOD10_EX", "EOD10_PE", "EOD10_ND", "EOD10_PN", "EOD10_NE", "EOD13", "EOD2", "EOD4", "EODCODE", "TUMOR_1V", "TUMOR_2V", "TUMOR_3V", "CS_SIZE", "CS_EXT", "CS_NODE", "CS_METS", "CS_SSF1", "CS_SSF2", "CS_SSF3", "CS_SSF4", "CS_SSF5", "CS_SSF6", "CS_SSF25", "D_AJCC_T", "D_AJCC_N", "D_AJCC_M", "D_AJCC_S", "D_SSG77", "D_SSG00", "D_AJCC_F", "D_SSG77F", "D_SSG00F", "CSV_ORG", "CSV_DER", "CSV_CUR", "SURGPRIM", "SCOPE", "SURGOTH", "SURGNODE", "RECONST", "NO_SURG", "RADIATN", "RAD_BRN", "RAD_SURG", "SS_SURG", "SRPRIM02", "SCOPE02", "SRGOTH02", "REC_NO", "O_SITAGE", "O_SEQCON", "O_SEQLAT", "O_SURCON", "O_SITTYP", "H_BENIGN", "O_RPTSRC", "O_DFSITE", "O_LEUKDX", "O_SITBEH", "O_EODDT", "O_SITEOD", "O_SITMOR", "TYPEFUP", "AGE_REC", "SITERWHO", "ICDOTO9V", "ICDOT10V", "ICCC3WHO", "ICCC3XWHO", "BEHANAL", "HISTREC", "BRAINREC", "CS0204SCHEMA", "RAC_RECA", "RAC_RECY", "NHIAREC", "HST_STGA", "AJCC_STG", "AJ_3SEER", "SSG77", "SSG2000", "NUMPRIMS", "FIRSTPRM", "STCOUNTY", "ICD_5DIG", "CODKM", "STAT_REC", "IHS", "HIST_SSG_2000", "AYA_RECODE", "LYMPHOMA_RECODE", "DTH_CLASS", "O_DTH_CLASS", "EXTEVAL", "NODEEVAL", "METSEVAL", "INTPRIM", "ERSTATUS", "PRSTATUS", "CSSCHEMA", "CS_SSF8", "CS_SSF10", "CS_SSF11", "CS_SSF13", "CS_SSF15", "CS_SSF16", "VASINV", "SRV_TIME_MON", "SRV_TIME_MON_FLAG", "SRV_TIME_MON_PA", "SRV_TIME_MON_FLAG_PA", "INSREC_PUB", "DAJCC7T", "DAJCC7N", "DAJCC7M", "DAJCC7STG", "ADJTM_6VALUE", "ADJNM_6VALUE", "ADJM_6VALUE", "ADJAJCCSTG")

UPDATE: LaF did the entire read in just under 7 seconds from the raw .txt file. Maybe there is an even faster way, but I doubt anything could do appreciably better. Amazing package.

27 July 2015 Update Just wanted to provide a small update to this. I used the new readr package, and I was able to read in the entire file in 5 seconds using readr::read_fwf.

seer9_readr <- read_fwf("path_to_data/COLRECT.TXT",
  col_positions = fwf_widths(cols))

Also, the updated stringi::stri_sub function is at least twice as fast as base::substr(). So, in the code above that uses fread to read the file (about 4 seconds), followed by apply to parse each line, the extraction of 143 variables took about 8 seconds with stringi::stri_sub compared to 19 for base::substr. So, fread plus stri_sub is still only about 12 seconds to run. Not bad.

seer9 <-  fread("path_to_data/COLRECT.TXT",     
  colClasses = "character", 
  sep = "\n", 
  header = FALSE)
text <- seer9[ , apply(start_end, 1, function(y) substr(V1, y[1], y[2]))] %>% 
  data.table(.)

10 Dec 2015 update:

Please also see the answer below by @MichaelChirico who has added some great benchmarks and the iotools package.

Solution 1

Now that there are (between this and the other major question about effective reading of fixed-width files) a fair amount of options on the offer for reading in such files, I think some benchmarking is appropriate.

I'll use the following on-the-large-side (400 MB) file for comparison. It's just a bunch of random characters with randomly defined fields and widths:

set.seed(21394)
wwidth = 400L
rrows = 1000000
    
#creating the contents at random
contents = write.table(
  replicate(
    rrows,
    paste0(sample(letters, wwidth, replace = TRUE), collapse = "")
  ),
  file = "testfwf.txt",
  quote = FALSE, row.names = FALSE, col.names = FALSE
)
    
#defining the fields & writing a dictionary
n_fields = 40L
endpoints = unique(
  c(1L, sort(sample(wwidth, n_fields - 1L)), wwidth + 1L)
)
cols = list(
  beg = endpoints[-(n_fields + 1L)], 
  end = endpoints[-1L] - 1L
)
    
dict = data.frame(
  column = paste0("V", seq_len(length(endpoints)) - 1L)),
  start = endpoints[-length(endpoints)] - 1,
  length = diff(endpoints)
)
    
write.csv(dict, file = "testdic.csv", quote = FALSE, row.names = FALSE)

I'll compare five methods mentioned between these two threads (I'll add some others if the authors would like): the base version (read.fwf), piping the result of in2csv to fread (@AnandaMahto's suggestion), Hadley's new readr (read_fwf), that using LaF/ffbase (@jwijffls' suggestion), and an improved (streamlined) version of that suggested by the question author (@MarkDanese) combining fread with stri_sub from stringi.

Here is the benchmarking code:

library(data.table)
library(stringi)
library(readr)
library(LaF)
library(ffbase)
library(microbenchmark)
    
microbenchmark(
  times = 5L,
  utils = read.fwf("testfwf.txt", diff(endpoints), header = FALSE),
  in2csv = fread(cmd = sprintf(
    "in2csv -f fixed -s %s %s",
    "testdic.csv", "testfwf.txt"
  )),
  readr = read_fwf("testfwf.txt", fwf_widths(diff(endpoints))),
  LaF = {
    my.data.laf = laf_open_fwf(
      'testfwf.txt', 
      column_widths = diff(endpoints),
      column_types = rep("character", length(endpoints) - 1L)
    )
    my.data = laf_to_ffdf(my.data.laf, nrows = rrows)
    as.data.frame(my.data)
  },
  fread = {
    DT = fread("testfwf.txt", header = FALSE, sep = "\n")
    DT[ , lapply(seq_len(length(cols$beg)), function(ii) {
      stri_sub(V1, cols$beg[ii], cols$end[ii])
    })]
  }
)

And the output:

# Unit: seconds
#    expr       min        lq      mean    median        uq       max neval cld
#   utils 423.76786 465.39212 499.00109 501.87568 543.12382 560.84598     5   c
#  in2csv  67.74065  68.56549  69.60069  70.11774  70.18746  71.39210     5 a  
#   readr  10.57945  11.32205  15.70224  14.89057  19.54617  22.17298     5 a  
#     LaF 207.56267 236.39389 239.45985 237.96155 238.28316 277.09798     5  b 
#   fread  14.42617  15.44693  26.09877  15.76016  20.45481  64.40581     5 a  

So it seems readr and fread + stri_sub are pretty competitive as the fastest; built-in read.fwf is the clear loser.

Note that the real advantage of readr here is that you can pre-specify column types; with fread you'll have to type convert afterwards.

EDIT: Adding some alternatives

At @AnandaMahto's suggestion I am including some more options, including one that appears to be a new winner! To save time I excluded the slowest options above in the new comparison. Here's the new code:

library(iotools)
    
microbenchmark(
  times = 5L,
  readr = read_fwf("testfwf.txt", fwf_widths(diff(endpoints))),
  fread = {
    DT = fread("testfwf.txt", header = FALSE, sep = "\n")
    DT[ , lapply(seq_len(length(cols$beg)), function(ii) {
      stri_sub(V1, cols$beg[ii], cols$end[ii])
    })]
  },
  iotools = input.file(
    "testfwf.txt", formatter = dstrfw, 
    col_types = rep("character", length(endpoints) - 1L), 
    widths = diff(endpoints)
  ),
  awk = fread(header = FALSE, cmd = sprintf(
    "awk -v FIELDWIDTHS='%s' -v OFS=', ' '{$1=$1 \"\"; print}' < testfwf.txt",
    paste(diff(endpoints), collapse = " ")
  ))
)

And the new output:

# Unit: seconds
#     expr       min        lq      mean    median        uq       max neval cld
#    readr  7.892527  8.016857 10.293371  9.527409  9.807145 16.222916     5  a 
#    fread  9.652377  9.696135  9.796438  9.712686  9.807830 10.113160     5  a 
#  iotools  5.900362  7.591847  7.438049  7.799729  7.845727  8.052579     5  a 
#      awk 14.440489 14.457329 14.637879 14.472836 14.666587 15.152156     5   b

So it appears iotools is both very fast and very consistent.

Solution 2

You can use the LaF package, which was written to handle large fixed width files (also too large to fit into memory). To use it you first need to open the file using laf_open_fwf. You can then index the resulting object as you would a normal data frame to read the data you need. In the example below, I read the entire file, but you can also read specific columns and/or lines:

library(LaF)
laf <- laf_open_fwf("foo.dat", column_widths = cols, 
  column_types=rep("character", length(cols)),
  column_names = seervars)
seer9 <- laf[,]

Your example using 5000 lines (instead of your 500,000) took 28 seconds using read.fwf and 1.6 seconds using LaF.

Addition Your example using 50,000 lines (instead of your 500,000) took 258 seconds using read.fwf and 7 seconds using LaF on my machine.

Solution 3

I'm not sure what OS you are using, but this worked pretty straightforwardly for me in Linux:

Step 1: Create a command for awk to convert the file to a csv

You can have it stored to an actual csv file if you plan to use the data in other software too.

myCommand <- paste(
  "awk -v FIELDWIDTHS='", 
  paste(cols, collapse = " "), 
  "' -v OFS=',' '{$1=$1 \"\"; print}' < ~/rawdata.txt", 
  collapse = " ")

Step 2: Use fread directly on that command that you just created

seer9 <- fread(myCommand)

---

I haven't timed this because I'm obviously using a slower system than you and Jan :-)

Solution 4

I wrote a parser for this kind of thing yesterday, but it was for a very specific kind of input to the header file, so I will show you how to format your column widths to be able to use it.

Converting your flat file to csv

First download the tool in question.

You can download the binary from the bin directory if you are on OS X Mavericks (where I compiled it on) or compile it by going to src and using clang++ csv_iterator.cpp parse.cpp main.cpp -o flatfileparser.

The flat file parser needs two files, a CSV header file in which every fifth element specifies the variable width (again, this is due to my extremely specific application), which you can generate using:

cols = c(8,10,1,2,1,1,1,3,4,3,2,2,4,4,1,4,1,4,1,1,1,1,3,2,2,1,2,2,13,2,4,1,1,1,1,3,3,3,2,3,3,3,3,3,3,3,2,2,2,2,1,1,1,1,1,6,6,6,2,1,1,2,1,1,1,1,1,2,2,1,1,2,1,1,1,1,1,1,1,1,1,1,1,1,1,1,7,5,4,10,3,3,2,2,2,3,1,1,1,1,2,2,1,1,2,1,9,5,5,1,1,1,2,2,1,1,1,1,1,1,1,1,2,3,3,3,3,3,3,1,4,1,4,1,1,3,3,3,3,2,2,2,2)
writeLines(sapply(c(-1, cols), function(x) paste0(',,,,', x)), '~/tmp/header.csv')

and copying the resulting ~/tmp/header.csv to the same directory as your flatfileparser. Move the flat file to the same directory as well, and you can run it on your flat file:

./flatfileparser header.csv yourflatfile

which will produce yourflatfile.csv. Add the header you have above in manually using piping (>> from Bash).

Reading in your CSV file quickly

Use Hadley's experimental fastread package by passing the filename to fastread::read_csv, which yields a data.frame. I don't believe he supports fwf files yet although it is on the way.

View more solutions

Comments

[Jan van der Laan]

Parallel reading your file isn't going to help. The bottleneck is the file IO. (Except of course when the data is spread across multiple machines/hard drives.)

[bdemarest]

@JanvanderLaan, He is able to read all the data into ram in 5 seconds with fread(). Parsing the 500k strings in parallel is the question I think.

[Jan van der Laan]

@bdemarest Yes, you are right. For the code using fread and substr, the parsing of the substrings is indeed the bottleneck and this can be done in parallel.

[Mark Danese]

I can't seem to get it to work. I am not a command line person, so it may just be me doing something wrong. mark-mbp-osx:bin mark$ flatfileparser header.csv COLRECT.TXT gives me -bash: flatfileparser: command not found on Mavericks. This is the listing of the directory: mark-mbp-osx:bin mark$ ls COLRECT.TXT flatfileparser header.csv

[Robert Krzyzanowski]

Try chmod +x flatfileparser; ./flatfileparser header.csv COLRECT.TXT

[Mark Danese]

It seems to have worked even though it gave errors: mark-mbp-osx:bin mark$ chmod +x flatfileparserchmod +x flatfileparser; ./flatfileparser header.csv COLRECT.TXT chmod: flatfileparserchmod: No such file or directory chmod: +x: No such file or directory mark-mbp-osx:bin mark$

[Robert Krzyzanowski]

I think you pasted the string "chmod +x flatfileparser" twice. Try two separate commands: first chmod +x flatfileparser and then ./flatfileparser header.csv COLRECT.TXT

[Mark Danese]

My fault, I pasted it into SO twice. I ended up with 144 columns instead of 143. It seems to work fine, so thanks. I am not sure I could use this regularly or on our Windows server. It would be great if it were easy to access from within R. I am just not a real programmer.

[Mark Danese]

I did not know about this package. Wow. 6 seconds. Excellent. About as fast as fread for a CSV file, which is very impressive. Will look into this more, since we have some large datasets. Thanks.

[Mark Danese]

Thanks a lot. I was hoping someone might suggest something like this. I tried it and it returned an error. Error in fread(myCommand) : ' ends field 14 on line 26 when detecting types: 428135680000001527 . . . I couldn't paste the entire 331 char string. Not sure what the issue is. This is OSX (Mavericks). I should probably force all to char for now.

[Mark Danese]

I tried forcing all to character. But the issue is that freed is only detecting 15 columns, not 143. Here is an edited version of my Command dropping many col values to fit in this comment: "awk -v FIELDWIDTHS=' 8 10 1 2 1 1 1 3 4 3 2 2 4 4 1 4 1 4 1 1 1 1 3 2 2 1 2 2 13 2 4 1 1 ' -v OFS=',' '{$1=$1 \"\"; print}' < ~/file.TXT"

[A5C1D2H2I1M1N2O1R2T1]

The benchmarks are useful. In the comments at the other question, I suggested trying the "iotools" package. Can you include that in the benchmarks, as well as the "awk" solution? I'm guessing the "awk" approach would be faster than "in2csv", but slower than "fread"/"readr", and, based on my experience with "iotools", I wouldn't be surprised if that's faster than the options available so far. Not tested, but the approach should be something like: library(iotools); input.file("testfwf.txt", formatter = dstrfw, col_types = rep("character", length(col_ends)-1), widths = diff(col_ends)). (+1)

[A5C1D2H2I1M1N2O1R2T1]

Oh, and for the error with "sqldf" (which I wouldn't bother testing for speed comparison), it's probably because we need to specify whatever the equivalent of header = FALSE would be. Don't really have the time to explore at this moment....

[Mark Danese]

Thanks to both of you. This is great information. I will edit the original question to guide readers to look down here.

[Martin Schmelzer]

It bothers me though that there is no option to set the encoding of the input file in input.file.