I would like to use the same pipe for different applications, like in:

cat my_file | {

Cmd1 should consume part of the input. Cmd2 should consume another part and so on.

However, each cmd eats more of the input then it read really needs due buffering.

For example:

yes | nl | { 
  head -n 10 > /dev/null
} | head -n 10

Outputs from line 912 instead of line 11.

Tee is not a good option, because each command is supposed to consume part of the stdin.

Is there a simple way to get this working?

  • 1
    Unfortunately I don't think there is any general solution to this. Ideally, commands should put back anything they read that they don't use but with pipes that is not always possible. Even if they did you would have to trust the commands to do so - there's no way for the shell to enforce any such agreement. For specific solutions you can come up with your own custom script to replace everything inside the { }. – jw013 Jul 22 '13 at 20:31
  • Thanks, I think I'll write some more functions avoiding buffering to replace head, tail and other commands. – arnaldocan Jul 22 '13 at 21:06
  • Use cat in each script to flush STDIN and pass stream to STDOUT. – F. Hauri Jul 22 '13 at 21:10
  • 1
    @jw013 There is no way for an application to put back a character. The input buffer is internal to the application: as far as the OS is concerned, the application has read the byte, whether it's processed it usefully or not. – Gilles 'SO- stop being evil' Jul 22 '13 at 22:06
  • @Gilles Not even reading one character at a time and using ungetc()? – jw013 Jul 23 '13 at 13:15

You may use tee to duplicate command for processing whole stream by many command:

( ( seq 1 10 | tee /dev/fd/5 | sed s/^/line..\ / >&4 ) 5>&1 | wc -l ) 4>&1 
line.. 1
line.. 2
line.. 3
line.. 4
line.. 5
line.. 6
line.. 7
line.. 8
line.. 9
line.. 10

or split line by line, using bash:

while read line ;do
    echo cmd1 $line
    read line && echo cmd2 $line
    read line && echo cmd3 $line
  done < <(seq 1 10)
cmd1 1
cmd2 2
cmd3 3
cmd1 4
cmd2 5
cmd3 6
cmd1 7
cmd2 8
cmd3 9
cmd1 10

Finaly there is a way for running cmd1, cmd2 and cmd3 only once with 1/3 of stream as STDIN:

( ( ( seq 1 10 |
         tee /dev/fd/5 /dev/fd/6 |
           sed -ne '1{:a;p;N;N;N;s/^.*\n//;ta;}' |
           cmd1 >&4
     ) 5>&1 |
       sed -ne '2{:a;p;N;N;N;s/^.*\n//;ta;}' |
       cmd2 >&4
  ) 6>&1 |
    sed -ne '3{:a;p;N;N;N;s/^.*\n//;ta;}' |
    cmd3 >&4
) 4>&1 
command_1: 1
command_1: 4
command_1: 7
command_1: 10
Command-2: 2
Command-2: 5
Command-2: 8
command 3: 3
command 3: 6
command 3: 9

For trying this, you could use:

alias cmd1='sed -e "s/^/command_1: /"' \
    cmd2='sed -e "s/^/Command_2: /"' \
    cmd3='sed -e "s/^/Command_3: /"'

For using one stream on different process if on same script, you could do:

    for ((i=(RANDOM&7);i--;));do
        read line;
        echo CMD1 $line
    for ((i=RANDOM&7;i--;));do
        read line
        echo CMD2 $line
    while read line ;do
        echo CMD3 $line
CMD1 1
CMD1 2
CMD1 3
CMD2 4
CMD2 5
CMD2 6
CMD2 7
CMD2 8
CMD2 9
CMD3 10

For this, you may have to transform your separated scripts into bash function to be able to build one overall script.

Another way could be to ensure each script won't output anything to STDOUT, than add a cat at end of each script to be able to chain them:


for ((i=1;1<n;i++));do
   read line
   pRoCeSS the $line
   echo >output_log


Final command could look like:

seq 1 10 | cmd1 | cmd2 | cmd2
  • I don't think this answers what the OP is trying to do. The idea is for sequential commands to each consume a portion of the same pipe stream, each picking up exactly where the previous one left off. – jw013 Jul 22 '13 at 20:29
  • What I'm trying to achieve is a bit different of your script. I would like, for example, cmd1 processing the first 10 lines, cmd2 processing the next 15 lines, and so on. Cmd1, cmd2, etc are bash functions, and these functions decide the number of lines to be processed according to the input. – arnaldocan Jul 22 '13 at 20:42
  • @arnaldocan Ok, I think rightly understand the need, look the sample script and last command sample. – F. Hauri Jul 22 '13 at 21:09
  • I think the fourth example solves the problem. Is there any way to avoid using "read" because of its speed limitations on large files? – arnaldocan Jul 22 '13 at 21:12
  • You may use mapfile -n $numLines arraOfLines – F. Hauri Jul 22 '13 at 21:46

For head -n 10 to be able to read 10 lines and not a single more character from the pipe on stdin, it would have to read one character at a time to be not read anything after the last newline character. That would be inefficient.

That's what the read shell builtin does when stdin is not seekable.

  head -n 10 > /dev/null
} < myfile

Works because head reads a chunk of data and lseeks back to just after the end of the 10th line. That obviously can't be done with pipes.

On recent GNU or FreeBSD systems, some applications that use stdio can be told to read their input one character at a time by using stdbuf -i1 or stdbuf -i0.

However, that doesn't work with GNU head. It works with GNU sed though, so you could do:

seq 20 | {
  stdbuf -i0 sed -n 10q

Alternatively, what you could do is control what goes on a pipe so that there's ever at most only one line in it at a time.

For instance, on Linux, you could do:

one_line_at_a_time() {
  perl -MTime::HiRes=usleep -pe '
    BEGIN{$|=1;open F, "<", "/dev/fd/1"; $r=""; vec($r,fileno(F),1) = 1}
    usleep(1000) while select($ro=$r,undef,undef,0)'
seq 20 | one_line_at_a_time | { head -n 10 > /dev/null; cat; }

That perl script opens "/dev/fd/1" in read mode which on Linux causes the other end of the pipe connected to fd 1 (stdout) to be opened. That way, with select, it can check if there's something in the pipe (and sleep until it's been emptied) before sending the next line.

Of course, that's also terribly inefficient.

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.