Open Stack

Excerpts from Daniel P. Berrange's message of 2015-07-24 09:48:15 +0100:
> On Thu, Jul 23, 2015 at 05:55:36PM +0300, mhorban wrote:
> > Hi all,
> > 
> > During development process in nova I faced with an issue related with config
> > options. Now we have lists of config options and registering options mixed
> > with source code in regular files.
> > From one side it can be convenient: to have module-encapsulated config
> > options. But problems appear when we need to use some config option in
> > different modules/packages.
> > 
> > If some option is registered in module X and module X imports module Y for
> > some reasons...
> > and in one day we need to import this option in module Y we will get
> > exception
> > NoSuchOptError on import_opt in module Y.
> > Because of circular dependency.
> > To resolve it we can move registering of this option in Y module(in the
> > inappropriate place) or use other tricks.
> > 
> > I offer to create file options.py in each package and move all package's
> > config options and registration code there.
> > Such approach allows us to import any option in any place of nova without
> > problems.
> > 
> > Implementations of this refactoring can be done piece by piece where piece
> > is
> > one package.
> > 
> > What is your opinion about this idea?
> 
> I tend to think that focusing on problems with dependancy ordering when
> modules import each others config options is merely attacking a symptom
> of the real root cause problem.
> 
> The way we use config options is really entirely wrong. We have gone
> to the trouble of creating (or trying to create) structured code with
> isolated functional areas, files and object classes, and then we throw
> in these config options which are essentially global variables which are
> allowed to be accessed by any code anywhere. This destroys the isolation
> of the various classes we've created, and means their behaviour often
> based on side effects of config options from unrelated pieces of code.
> It is total madness in terms of good design practices to have such use
> of global variables.
> 
> So IMHO, if we want to fix the real big problem with config options, we
> need to be looking to solution where we stop using config options as
> global variables. We should change our various classes so that the
> neccessary configurable options as passed into object constructors
> and/or methods as parameters.

We've tried to do this in a lot of places throughout Oslo. It mostly
works, until you hit a driver that uses options that the other drivers
don't (oslo.messaging has this problem especially).

We had a ConfigFilter class in oslo.config to enforce the isolation (an
option registered on a filter object is only visible to the code that
owns the filter). However, that causes challenges for value
interpolation. A deployer doesn't know which options are visible to each
other, and has a flat file where they can clearly see all of the
values together, so they expect %(foo)s to work no matter where "foo" is
defined. Until we solve those issues, the ConfigFilter isn't really
supported.

The best solution we've come up with so far is to isolate the options
into groups based on the driver name, and then not allow code outside of
the driver to access those options for any reason.

To deal with import ordering issues, we pass ConfigObj instance
around, and register options at runtime instead of import time,
usually in an __init__ method.  Registering an option more than
once is fine.

> 
> As an example in the libvirt driver.
> 
> I would set it up so that /only/ the LibvirtDriver class in driver.py
> was allowed to access the CONF config options. In its constructor it
> would load all the various config options it needs, and either set
> class attributes for them, or pass them into other methods it calls.
> So in the driver.py, instead of calling CONF.libvirt.libvirt_migration_uri
> everywhere in the code,  in the constructor we'd save that config param
> value to an attribute 'self.mig_uri = CONF.libvirt.libvirt_migration_uri'
> and then where needed, we'd just call "self.mig_uri".

There are, from time to time, specs proposed to provide ways for
applications to reload their configuration settings without restarting.
So far we've said that was an application issue, because oslo.config
can reload the files, but it doesn't know when (that's an app signal
handling thing) and it doesn't know where values might have been
saved like you describe.  I don't know if that's something the nova
team wants to support.

> Now in the various other libvirt files, imagebackend.py, volume.py
> vif.py, etc. None of those files would /ever/ access CONF.*. Any time
> they needed a config parameter, it would be passed into their constructor
> or method, by the LibvirtDriver or whatever invoked them.
> 
> Getting rid of the global CONF object usage in all these files trivially
> now solves the circular dependancy import problem, as well as improving
> the overall structure and isolation of our code, freeing all these methods
> from unexpected side-effects from global variables.
> 
> Regards,
> Daniel