[Libvir] Re: virsh outputs to stderr(?!) why not stdout

[Daniel P. Berrange]

On Thu, Apr 06, 2006 at 04:18:40PM +0200, Karel Zak wrote:
> On Thu, Apr 06, 2006 at 02:18:10PM +0100, Daniel P. Berrange wrote:
> > On Thu, Apr 06, 2006 at 02:44:41PM +0200, Karel Zak wrote:
> > > On Thu, Apr 06, 2006 at 08:40:17AM -0400, Daniel Veillard wrote:
> > > > On Thu, Apr 06, 2006 at 02:37:37PM +0200, Karel Zak wrote:
> > > > > On Thu, Apr 06, 2006 at 11:46:57AM +0200, Jim Meyering wrote:
> > > > > > So, now I can use something like this to get a list of domain names:
> > > > > > 
> > > > > >   virsh list | tail -n +4 | awk '{print $2}'
> > > > > > 
> > > > > > Is there some easier way?
> > > > > 
> > > > >  I think we can add some options to the list command:
> > > > > 
> > > > >     --noheader
> > > > >     --nodom0
> > > > >     --cols=name,id,state,maxmem,usemem,....
> > > > > 
> > > > >  without these options it will use default (current) format. I've
> > > > >  already thought about this solution for the others commands (e.g.
> > > > >  dominfo). I think this is a way how make it useful in scripts.
> > 
> > This raises the question of what the primary intended use case of virsh 
> > is / should be. Currently I see it as being an interactive administrator
> 
>  It's already possible use it as interactive tool ("virsh" without
>  any option) and non-interactive (e.g "virsh list").
> 
> > Now, I don't disagree that there is a need for something that is easy to 
> > script in the shell, but if we are going to make virsh easier to script
> > we should be careful not to compromise the primary use case of interactive
> > administrator control. Conversly we must be careful not be end up with the
> 
>  Agree. I think the example above with the list command doesn't discredit
>  interactive mode.
> 
> > presentation focused output becoming a defacto API, because there is much
> > pain & suffering that way - simple changes to improve user experiance would
> > end up breaking scripting. Another example - gettext translations of the
> > output fields could break scripts which are grokking for English strings.
> 
>  This is pretty common for all commands, the solution is use "LANG=C" in
>  shell scripts. 

And add several more environment variables to get predicatable date formatting,
timezone settings, etc. I think it would make it simpler & more reliable for
people scripting to have a single --batch script which gave a formatting which
we can guarentee to be consitent. 

> > So, I'd say, rather than add a whole range of extra command line options
> > for tweaking the existing administrator presentation-focused output, we 
> > should instead add a single '--batch' option. This would activate a completely 
> > separate presentation mode tailored explicitly to the needs of scripts. It
> > would have a formal structure for outputting data, perhaps a series of rows
> > expressed as 'key:value' pairs, or a single line in CSV format, or some other
> > easily shell-parsable format.
> 
>  Is there any example where any UNIX command use this way? (It doesn't
>  mean that I disagree with you, but the virsh shouldn't be like from
>  foreign planet:-).

I don't think it'll make it look like its from a foreign planet - simple
lines of key:value pairs are pretty command for text based values, eg
/proc/cpu_info. I also don't think some of the comparable existing UNIX 
tools are a particularly good example. For example, I know of 3rd party
monitoring software which used sar, iostat & friends to collect system
performance metrics. Which worked fine, until a point release of sysstat
changed the presentation format. sar did actually used to have a formal
batch mode command line switch for doing sensible formatting, but they
removed it in latest release !

> 
>  IMHO it will be better use --fieldsep=<foo> and --recordsep=<bar>
>  global options rather than hardcode there CSV or key:value format.

I'm not convinced that providing customizeable field separators is better,
because it puts a burden on the caller to guess at what field separator 
is unambigous wrt to the data being output. It make look like ':' is a
safe separator on their current machine, but if someone were  to create
a VM with ':' in its name, then the script would go wrong. I think it'll
lead to a more reliable tool if we just designate one separator & guarentee
that it will be safe - either because we know the valid characters for
each field, or by defining some escaping scheme that will be applied.

The real issue being addressed by options such as '--noheader', '--fieldsep',
'--cols', etc is to try and change the output into something predictable.
While options for customiziing various aspects of output may or may not be
useful features own right, they don't really satisfactorily address the real
goal here which is predictable output formatting, to enable reliable isimple
and reliable scripting. A single --batch flag with formally defined sematnics
would satisfy the goal of making scripting reliable.

Dan.
-- 
|=- Red Hat, Engineering, Emerging Technologies, Boston.  +1 978 392 2496 -=|
|=-           Perl modules: http://search.cpan.org/~danberr/              -=|
|=-               Projects: http://freshmeat.net/~danielpb/               -=|
|=-  GnuPG: 7D3B9505   F3C9 553F A1DA 4AC2 5648 23C1 B3DF F742 7D3B 9505  -=|