← Back to team overview

nova team mailing list archive

Re: Gflags, Settings, Dependency Injection


On 07/28/2010 02:28 PM, Eric Day wrote:
> ++
> I'm all for using an existing solution if one exists. I've not looked
> enough to make calls either way though. I want to figure out *what*
> we are looking for in features to make those decisions.


> On Wed, Jul 28, 2010 at 01:37:18PM -0700, Monty Taylor wrote:
>> So I know I haven't convinced everyone to love bzr yet ... but as they
>> are a large python project with command line and config file options -
>> and plugins - perhaps looking at the infrastructure/design they use
>> might be a good idea?
>> Also, the work derks did with cement might be of help.
>> I believe both are designed to do things similar to how you are
>> discussing them below (although different, of course - we're all python
>> devs, there's no way we're going to actually do things the same. :) )
>> Monty
>> (what Eric is saying makes sense to me - but I don't have a whole bunch
>> of stake either way here- I am a fan of reusing solutions that exist
>> where possible though of course)
>> On 07/28/2010 01:24 PM, Eric Day wrote:
>>> Hi Vish,
>>> If we want to keep things modular and have runtime module selection
>>> like you mention, we probably need to rethink flags. Using gflags
>>> may not be an option unless we can somehow make 'undefok=' a global
>>> option. In other project (that was not in Python, so no code to help),
>>> the flow is:
>>> * Enforce the use of module names in the options. For example, for
>>>   generic queue module options use --queue.*, for libvirt module
>>>   options, use --libvirt.*. If we want to make this seamless, we
>>>   would probably need to use something else instead gflags or create
>>>   a wrapper to enforce the required behavior.
>>> * Import the core option manager, first thing that happens when
>>>   starting a binary.
>>> * Parse all options, separating each out into the modules they belong
>>>   to. We don't know what is valid yet, but we can at least group by module.
>>> * Load any required modules via normal 'import' lines. They can verify
>>>   options for their module space.
>>> * Have some core flags that specify which modules to load, for example,
>>>   use rabbit vs fakerabbit. Then 'import' the selected optional modules.
>>> * As optional modules load, let them verify the module namespace
>>>   options just like the required modules did.
>>> * Any options for modules that were not loaded are just ignored.
>>> Thoughts on this? It has worked out quite well in the other C++ project
>>> for me, and with Python it would be even easier to put together. :)
>>> -Eric
>>> On Wed, Jul 28, 2010 at 11:13:40AM -0700, Vishvananda Ishaya wrote:
>>>>    I'm having some annoyances with gflags which I'd like to air out here.
>>>>     Maybe we can come to a consensus about how to move forward with them.  I
>>>>    find gflags annoying in the following ways:
>>>>    a) flags are irritating for global settings.  Settings that apply to the
>>>>    project as a whole have to be set in multiple places so that the binaries
>>>>    all get them properly.  This can be fixed somewhat by a shared flagfile.
>>>>     For example:
>>>>    /etc/nova/nova-manage.conf:
>>>>      --flagfile=/etc/nova/nova-common.conf # shared settings
>>>>      --otherflag=true #manage specific settings
>>>>    The problem here is that the shared settings can only include settings
>>>>    that are imported by EVERY binary, or one of the binaries will choke.  So
>>>>    if you have a flag that 4 of 5 binaries use, you either have to set it in
>>>>    four flagfiles or put it in common with an ugly undefok= line.  This all
>>>>    seems nasty to me.  Other possibilities include moving truly
>>>>    common/settings related flags into the common flags.py so that they are
>>>>    available to all binaries.  It all seems a bit hackish.
>>>>    b) including files for flags only
>>>>    There are places where we need access to a flag, but we aren't actually
>>>>    making calls in the file.  Pyflakes and pylint complain about unused
>>>>    imports.  Perhaps we fix this by moving these flags into common flagfile?
>>>>    c) dependency injection
>>>>    This is related to the issue above.  If we are dynamically loading
>>>>    specific drivers (for example the auth driver or a datastore backend) as
>>>>    specified by a flag, the import is often done later than the parent file
>>>>    is imported.  Therefore using flags to configure settings for the driver
>>>>    will fail, because the binary recognizing the flags is dependent on the
>>>>    file that contains the flags being imported.  Workarounds here include
>>>>    finding a different method for dependency injection, hacking flags to
>>>>    search for flags in injected dependencies somehow, or configuring drivers
>>>>    differently than the rest of the system.
>>>>    So I see 3 options for moving forward
>>>>    1) ditch gflags completely and use a different method for specifying
>>>>    settings
>>>>    2) use a combination of some kind of settings file for general
>>>>    configuration, and flags for specific runtime settings/hacks
>>>>    3) find good standard practices/workarounds for the above issues
>>>>    Thoughts?
>>>>    Vish
>>>> _______________________________________________
>>>> Mailing list: https://launchpad.net/~nova
>>>> Post to     : nova@xxxxxxxxxxxxxxxxxxx
>>>> Unsubscribe : https://launchpad.net/~nova
>>>> More help   : https://help.launchpad.net/ListHelp
>>> _______________________________________________
>>> Mailing list: https://launchpad.net/~nova
>>> Post to     : nova@xxxxxxxxxxxxxxxxxxx
>>> Unsubscribe : https://launchpad.net/~nova
>>> More help   : https://help.launchpad.net/ListHelp

Follow ups