I have had the situation where I have data on a pipe that I want to give to `foo a` and `foo b` one line at a time. They should both receive all input.

I currently do:

  foo() { echo "$@"; cat; }
  foorun() { parallel --pipe -N1 foo "$@"; }
  export -f foorun foo

  seq 3 | parallel --pipe --tee foorun {} ::: a b
  seq 3 | parallel --pipe --tee foorun {} ::: a b :::+ c d
  seq 3 | parallel --pipe --tee foorun {} ::: a b ::: c d

I feel a bit silly having to build foorun(). So I have the feeling this ought do the same (it does not):

  seq 3 | parallel --pipe -N1 foo ::: a b
  seq 3 | parallel --pipe -N1 foo ::: a b
  seq 3 | parallel --pipe -N1 foo ::: a b :::+ c d
  seq 3 | parallel --pipe -N1 foo ::: a b ::: c d

The current behaviour works more like --xapply, so it should only work if you use :::+ or --xapply (this already works):

  seq 3 | parallel --pipe -N1 foo :::+ a b c

As a first version I intend to simply let GNU Parallel call parallel (very much as above). That means that sequence number and slot number will not give the correct values (and probably some other limitations). Similar to how --(n)onall works right now.

If this works, a later version will read block n from stdin. Add a(n) .. z(n) to the queue. When the queue is empty, read block n+1. Repeat.

So even if, say, c(x) is always fast, GNU Parallel will at most keep one block in memory per jobslot.