Your race condition is: one side of fork gets to the (buffered) read before the other side gets to its seek().

I find it easier to see what is going on by using non buffered reads (sysread). Half-offsetting (timewise) the two forks leads to consistent results.

Removing the (unnecessary) seeks()'s should also lead to consistent output size (although it could be from either the parent or child).

#!/usr/bin/perl

# https://perlmonks.org/?node_id=1230099

use strict;
use warnings;
use IO::Handle;
$| = 1;

my $pos = tell DATA;
my $pid = fork();

my $x= IO::Handle->new();
$x->fdopen(fileno(DATA), "r");

$pid and select undef, undef, undef, 0.05;
seek $x, $pos, 0;

while( sysread $x, $_, 1 )
  {
  print $pid ? uc : lc;
  select undef, undef, undef, 0.1;
  }

1 while wait > 0;

__DATA__
aa
bb
cc
dd
ee
[download]

Outputs:

aAa
bB
Cc
dD
Ee
[download]

Note the extra 'a' as a result of the seek().