Re: [PATCH/RFC] recv_sideband: Band #2 always goes to stderr

[Johannes Sixt <j.sixt@viscovery.net>]

Johannes Schindelin schrieb:
> FWIW GitTorrent may be implemented as part of git-daemon, if Sam's ideas 
> become reality.  And then, sideband transport is _the_ means to do 
> asyncrounous communication while pushing bytes.

I do not see how recv_sideband() in its current form could be helpful here
(assuming that you really are thinking of sending binary data over band #2).

> On Tue, 10 Mar 2009, Johannes Sixt wrote:
>> Johannes Sixt schrieb:
>>> For use-cases that you have in mind in GitTorrent, the *protocol* may 
>>> be a good choice, but the current implementation is definitely a 
>>> special case.
>> And it really is: Did you notice that stuff that recv_sideband sends over
>> the channel named 'err' (before my patch) has "remote: " prepended on
>> every line? That's certainly not an implementation that you want if you
>> send binary data over that band!
> 
> Yes, that is unfortunate, but can be fixed easily.

I don't believe this. Every treatment of "remote: " that you take away
from recv_sideband() you must insert somewhere else. Perhaps easy, but
certainly not as trivial as my patch.


Just a reminder: You proposed to override write() on Windows in a
non-trivial way, and we are discussing the topic above because I think
that is not a good idea. The reasons are:

- write() is a fundamental operation, and we should not mess with it out
of caution.

- Your proposal is not a catch-all. For example, combine-diff.c uses
puts() in dump_quoted_path(). If your goal was to not touch code outside
of compat/ then you need to override at least puts(), too.

- All code that writes ANSI escapes should use fprintf() anyway.
(Currently that is not the case, but all cases I'm aware of can be fixed
trivially.)

-- Hannes