Re: [PATCH v3 2/6] run-command: add hide_output to run_processes_parallel_opts

[Calvin Wan <calvinwan@google.com>]

> > Setting "no_stdout", "no_stderr", etc. in a
> > "child_process" shouldn't imply that we still pass the stdout and stderr to
> >  "parallel_processes" and then we send the output to "/dev/null".
>
> Sure, but if they're not producing any output because it's being piped
> to /dev/null how worthwhile is it to optimize that?
>
> We still can optimize it, but I still think the interface should just be
> the equivalent of:
>
>         parallel -k -j100% 'sleep 0.0$RANDOM && echo {} >/dev/null' ::: {1..100}
>
> Whereas what you seem to be trying to implement is the equivalent of a:
>
>         parallel -u -j100% 'sleep 0.0$RANDOM && echo {} ::: {1..100} >/dev/null
>
> Except as an option to the parallel API, but the end result seems to be
> equivalent.
>
> > That being said, I can understand the aversion to adding an option like
> > this that doesn't also add support for stdout and stderr. I can remove this
> > patch and instead reset the buffer inside of pipe_output and task_finished
> > in a later patch
>
> I'm not necessarily opposed to it, just puzzled about it, maybe I don't
> have the full picture.
>
> In general I highly recomend looking at whatever GNU parallel is doing,
> and seeing if new features in run-command.[ch] can map to that mental
> model.
>
> Our API is basically a small subset of its featureset, and I've found it
> useful both to steal ideas from there, and to test
> assumptions. E.g. "ungroup" is just a straight rip-off of the
> "--ungroup" option, it's also had to think about combining various
> options we don't have yet (but might want).
>
> In that case the supervisor API/parallel(1) needs to do something
> special, but for "I don't want output" it seems best to just do that at
> the worker level, i.e. equivalent to piping to /dev/null.

Well I want the output to be able to parse it, but not to print it. Piping
to /dev/null at the worker level denies me the ability to parse it in the
parent process.

Am I understanding correctly that what you're suggesting is if a child
process has "no_stderr" and "no_stdout" set to true, then
parallel_processes would temporarily set them to false before
start_command, and then honor it later after the output is read?
This would allow me to call pipe_output and parse it before sending
the output to /dev/null without the need for "hide_output"