r - Reading in chunks at a time using fread in package data.table

Question

I'm trying to input a large tab-delimited file (around 2GB) using the fread function in package data.table. However, because it's so large, it doesn't fit completely in memory. I tried to input it in chunks by using the skip and nrow arguments such as:

chunk.size = 1e6
done = FALSE
chunk = 1
while(!done)
{
    temp = fread("myfile.txt",skip=(chunk-1)*chunk.size,nrow=chunk.size-1)
    #do something to temp
    chunk = chunk + 1
    if(nrow(temp)<2) done = TRUE
}

In the case above, I'm reading in 1 million rows at a time, performing a calculation on them, and then getting the next million, etc. The problem with this code is that after every chunk is retrieved, fread needs to start scanning the file from the very beginning since after every loop iteration, skip increases by a million. As a result, after every chunk, fread takes longer and longer to actually get to the next chunk making this very inefficient.

Is there a way to tell fread to pause every say 1 million lines, and then continue reading from that point on without having to restart at the beginning? Any solutions, or should this be a new feature request?

Answer 1

You should use the LaF package. This introduces a sort of pointer on your data, thus avoiding the - for very large data - annoying behaviour of reading the whole file. As far as I get it fread() in data.table pckg need to know total number of rows, which takes time for GB data. Using pointer in LaF you can go to every line(s) you want; and read in chunks of data that you can apply your function on, then move on to next chunk of data. On my small PC I ran trough a 25 GB csv-file in steps of 10e6 lines and extracted the totally ~5e6 observations needed - each 10e6 chunk took 30 seconds.

UPDATE:

library('LaF')
huge_file <- 'C:/datasets/protein.links.v9.1.txt'

#First detect a data model for your file:
model <- detect_dm_csv(huge_file, sep=" ", header=TRUE)

Then create a connection to your file using the model:

df.laf <- laf_open(model)

Once done you can do all sort of things without needing to know the size of the file as in data.table pckgs. For instance place the pointer to line no 100e6 and read 1e6 lines of data from here:

goto(df.laf, 100e6)
data <- next_block(df.laf,nrows=1e6)

Now data contains 1e6 lines of your CSV file (starting from line 100e6).

You can read in chunks of data (size depending on your memory) and only keep what you need. e.g. the huge_file in my example points to a file with all known protein sequences and has a size of >27 GB - way to big for my PC. To get only human sequence I filtered using organism id which is 9606 for human, and this should appear in start of the variable protein1. A dirty way is to put it into a simple for-loop and just go read one data chunk at a time:

library('dplyr')
library('stringr')

res <- df.laf[1,][0,]
for(i in 1:10){
  raw <-
    next_block(df.laf,nrows=100e6) %>% 
    filter(str_detect(protein1,"^9606\."))
  res <- rbind(res, raw)

    }

Now res contains the filtered human data. But better - and for more complex operations, e.g. calculation on data on-the-fly - the function process_blocks() takes as argument a function. Hence in the function you do what ever you want at each piece of data. Read the documentation.