This is how I did it:

    # Creation of a COMPUTE data source is not contemplated in the
    # documentation, nor it is in the code. That's why we use an
    # object in dry-run mode: we'll get a reference to the function
    # and arguments that were about to be called, and add our
    # resource. Finally, we run the call.
    # Pretty bad it didn't have the function already...
    my $dsdef =
      q{DS:aggregate:COMPUTE:}.
        join(",",@dsources).",ADDNAN"x$#dsources ;

    $rrdbuilder->create(@rrd_parms) ;

    # Get references to the RRD function and arguments
    # Push the COMPUTE data source in it
    # Run the create command, see what happens
    my ($subref,$args) = $rrdbuilder->get_exec_env();
    push @$args,$dsdef ;

    eval {
      $subref->(@$args) ;
    } ;
[download]