[openstack-dev] [openstack][nova] Streamlining of config options in nova

Sean Dague sean at dague.net
Mon Jul 27 14:35:56 UTC 2015


On 07/27/2015 10:05 AM, Daniel P. Berrange wrote:
> On Fri, Jul 24, 2015 at 09:48:15AM +0100, Daniel P. Berrange wrote:
>> 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.
>>
>> 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".
>>
>> 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.

How does that address config reload on SIGHUP? It seems like that
approach would break that feature.

> Another significant downside of using CONF objects as global variables
> is that it is largely impossible to say which nova.conf setting is
> used by which service. Figuring out whether a setting affects nova-compute
> or nova-api or nova-conductor, or ... largely comes down to guesswork or
> reliance on tribal knowledge. It would make life significantly easier for
> both developers and administrators if we could clear this up and in fact
> have separate configuration files for each service, holding only the
> options that are relevant for that service.  Such a cleanup is not going
> to be practical though as long as we're using global variables for config
> as it requires control-flow analysis find out what affects what :-(

Part of the idea that came up in the room is to annotate variables with
the service they were used in, and deny access to in services they are
not for.

	-Sean

-- 
Sean Dague
http://dague.net



More information about the OpenStack-dev mailing list