How to read only lines that fulfil a condition from a csv into R?

By far the easiest (in my book) is to use pre-processing.

R> DF <- data.frame(n=1:26, l=LETTERS)
R> write.csv(DF, file="/tmp/data.csv", row.names=FALSE)
R> read.csv(pipe("awk 'BEGIN {FS=\",\"} {if ($1 > 20) print $0}' /tmp/data.csv"),
+           header=FALSE)
  V1 V2
1 21  U
2 22  V
3 23  W
4 24  X
5 25  Y
6 26  Z
R> 

Here we use awk. We tell awk to use a comma as a field separator, and then use the conditon 'if first field greater than 20' to decide if we print (the whole line via $0).

The output from that command can be read by R via pipe().

This is going to be faster and more memory-efficient than reading everythinb into R.

You could use the read.csv.sql function in the sqldf package and filter using SQL select. From the help page of read.csv.sql:

library(sqldf)
write.csv(iris, "iris.csv", quote = FALSE, row.names = FALSE)
iris2 <- read.csv.sql("iris.csv", 
    sql = "select * from file where `Sepal.Length` > 5", eol = "\n")

I was looking into readr::read_csv_chunked when I saw this question and thought I would do some benchmarking. For this example, read_csv_chunked does well and increasing the chunk size was beneficial. sqldf was only marginally faster than awk.

library(tidyverse)
library(sqldf)
library(data.table)
library(microbenchmark)

# Generate an example dataset with two numeric columns and 5 million rows
tibble(
  norm = rnorm(5e6, mean = 5000, sd = 1000),
  unif = runif(5e6, min = 0, max = 10000)
) %>%
  write_csv('medium.csv')

microbenchmark(
  readr  = read_csv_chunked('medium.csv', callback = DataFrameCallback$new(function(x, pos) subset(x, unif > 9000)), col_types = 'dd', progress = F),
  readr2 = read_csv_chunked('medium.csv', callback = DataFrameCallback$new(function(x, pos) subset(x, unif > 9000)), col_types = 'dd', progress = F, chunk_size = 1000000),
  sqldf  = read.csv.sql('medium.csv', sql = 'select * from file where unif > 9000', eol = '\n'),
  awk    = read.csv(pipe("awk 'BEGIN {FS=\",\"} {if ($2 > 9000) print $0}' medium.csv")),
  awk2   = read_csv(pipe("awk 'BEGIN {FS=\",\"} {if ($2 > 9000) print $0}' medium.csv"), col_types = 'dd', progress = F),
  fread  = fread(cmd = "awk 'BEGIN {FS=\",\"} {if ($2 > 9000) print $0}' medium.csv"),
  check  = function(values) all(sapply(values[-1], function(x) all.equal(values[[1]], x))),
  times  = 10L
)

# Updated 2020-05-29

# Unit: seconds
#   expr   min    lq  mean  median    uq   max neval
#  readr   2.6   2.7   3.1     3.1   3.5   4.0    10
# readr2   2.3   2.3   2.4     2.4   2.6   2.7    10
#  sqldf  14.1  14.1  14.7    14.3  15.2  16.0    10
#    awk  18.2  18.3  18.7    18.5  19.3  19.6    10
#   awk2  18.1  18.2  18.6    18.4  19.1  19.4    10
#  fread  17.9  18.0  18.2    18.1  18.2  18.8    10

# R version 3.6.2 (2019-12-12)
# macOS Mojave 10.14.6        

# data.table 1.12.8
# readr      1.3.1 
# sqldf      0.4-11

You can read the file in chunks, process each chunk, and then stitch only the subsets together.

Here is a minimal example assuming the file has 1001 (incl. the header) lines and only 100 will fit into memory. The data has 3 columns, and we expect at most 150 rows to meet the condition (this is needed to pre-allocate the space for the final data:

# initialize empty data.frame (150 x 3)
max.rows <- 150
final.df <- data.frame(Variable1=rep(NA, max.rows=150), 
                       Variable2=NA,  
                       Variable3=NA)

# read the first chunk outside the loop
temp <- read.csv('big_file.csv', nrows=100, stringsAsFactors=FALSE)
temp <- temp[temp$Variable2 >= 3, ]  ## subset to useful columns
final.df[1:nrow(temp), ] <- temp     ## add to the data
last.row = nrow(temp)                ## keep track of row index, incl. header

for (i in 1:9){    ## nine chunks remaining to be read
  temp <- read.csv('big_file.csv', skip=i*100+1, nrow=100, header=FALSE,
                   stringsAsFactors=FALSE)
  temp <- temp[temp$Variable2 >= 3, ]
  final.df[(last.row+1):(last.row+nrow(temp)), ] <- temp
  last.row <- last.row + nrow(temp)    ## increment the current count
}

final.df <- final.df[1:last.row, ]   ## only keep filled rows
rm(temp)    ## remove last chunk to free memory

Edit: Added stringsAsFactors=FALSE option on @lucacerone's suggestion in the comments.