Open Stack

----- Original Message -----
> From: "Mehdi Abaakouk" <sileht at sileht.net>
> To: openstack-dev at lists.openstack.org
> Sent: Friday, August 7, 2015 1:57:54 AM
> Subject: [openstack-dev] [oslo][keystone] oslo_config and wsgi middlewares
> 
> Hi,
> 
> I want to share with you some problems I have recently encountered with
> openstack middlewares and oslo.config.
> 
> The issues
> ----------
> 
> In project Gnocchi, I would use oslo.middleware.cors, I have expected to
> just put the name of the middleware to the wsgi pipeline, but I can't.
> The middlewares only works if you pass the oslo_config.cfg.ConfigOpts()
> object or via 'paste-deploy'... Gnocchi doesn't use paste-deploy, so
> I have to modify the code to load it...
> (For the keystonemiddleware, Gnocchi already have a special
> handling/hack to load it [1] and [2]).
> I don't want to write the same hack for each openstack middlewares.
> 
> 
> In project Aodh (ceilometer-alarm), we recently got an issue with
> keystonemiddleware since we remove the usage of the global object
> oslo_config.cfg.CONF. The middleware doesn't load its options from the
> config file of aodh anymore. Our authentication is broken.
> We can still pass them through paste-deploy configuration but this looks
> a method of the past. I still don't want to write a hack for each
> openstack middlewares.
> 
> 
> Then I have digged into other middlewares and applications to see how
> they handle their conf.
> 
> oslo_middlewarre.sizelimit and oslo_middlewarre.ssl take options only
> via the global oslo_config.cfg.CONF. So they are unusable for application
> that doesn't use this global object.
> 
> oslo_middleware.healthcheck take options as dict like any other python
> middleware. This is suitable for 'paste-deploy'. But doesn't allow
> configuration via oslo.config, doesn't have a strong config options
> type checking and co.
> 
> Zaqar seems got same kind of issue about keystonemiddleware, and just
> write a hack to workaround the issue (monkeypatch the cfg.CONF of
> keystonemiddleware with their local version of the object [3] and then
> transform the loaded options into a dict to pass them via the legacy
> middleware dict options [4]) .
> 
> Most applications, just still use the global object for the
> configuration and don't, yet, see those issues.
> 
> 
> All of that is really not consistent.
> 
> This is confusing for developer to have some middlewares that need pre-setup,
> enforce them to rely on global python object, and some others not.
> This is confusing for deployer their can't do the configuration of
> middlewares in the same way for each middlewares and each projects.
> 
> But keystonemiddleware, oslo.middleware.cors,... are supposed to be wsgi
> middlewares, something that is independant of the app.
> And this is not really the case.
> 
> From my point of view and what wsgi looks like generally in python, the
> middleware object should be just MyMiddleware(app, options_as_dict),
> if the middleware want to rely to another configuration system it should
> do the setup/initialisation itself.
> 
> 
> 
> So, how to solve that ?
> ------------------------
> 
> Do you agree:
> 
> * all openstack middlewares should load their options with oslo.config ?
>   this permits type checking and all other features it provides, it's cool :)
>   configuration in paste-deploy conf is thing of past
> 
> * we must support local AND global oslo.config object ?
>   This is an application choice not something enforced by middleware.
>   The deployer experience should be the same in both case.
> 
> * the middleware must be responsible of the section name in the oslo.config ?
>   Gnocchi/Zaqar hack have to hardcode the section name in their code,
>   this doesn't looks good.
> 
> * we must support legacy python signature for WSGI object,
>   MyMiddleware(app, options_as_dict) ? To be able to use paste for
>   application/deployer that want it and not break already deployed things.
> 
> 
> I really think all our middlewares should be consistent:
> 
> * to be usable by all applications without enforcing them to write crap
> around them.
> * and to made the deployer life easier.
> 
> 
> Possible solution:
> ------------------
> 
> I have already started to work on something that do all of that for all
> middlewares [5], [6]
> 
> The idea is, the middleware should create a oslo_config.cfg.ConfigOpts()
> (instead of rely on the global one) and load the configuration file of the
> application in. oslo.config will discover the file location just with the
> name of application as usual.
> 
> So the middleware can now be loaded like this:
> 
> code example:
> 
>    app = MyMiddleware(app, {"oslo_config_project": "aodh"})
> 
> paste-deploy example:
> 
>    [filter:foobar]
>    paste.filter_factory = foobar:MyMiddleware.filter_factory
>    oslo_config_project = aodh
> 
> oslo_config.cfg.ConfigOpts() will easly find the /etc/aodh/aodh.conf,
> This cut the hidden links between middleware and the application
> (through the global object).
> 
> And of course if oslo_config_project is not provided, the middleware
> fallback the global oslo.config object.
> The middleware options can still be passed via the legacy dict.
> The backward compatibility is conserved.
> 
> I have already tested that in Gnocchi and Aodh, and that solves all of
> my issues. Remove all hacks, the application doesn't need special pre
> setup. All our middleware become normal middleware but still can use
> oslo.config.
> 
> 
> WDYT ?

So i agree on most of your points, the middlewares are certainly inconsistent, i would like to make things easier for deployers and I don't like global CONF objects (but i've now had that argument so many times i'm done). 

Also to support some of the newer services that don't use paste i think we should absolutely make it so that the CONF object is passed to middleware rather than sourced globally. I think gnochhi and zaqar both fit into this case. 

The problem i see with what you are saying is that it is mixing deployment methodologies in a way that is unintended. Paste is designed to allow deployers to add and remove middleware independent of configuring the service. This means that at paste time there is no CONF object unless it's globally registered and this is why most middlewares allow taking options from paste.ini because if you don't have global CONF then it's the only way to actually get options into the middleware. 

The proposed solution is then to provide enough options (via paste) to load a CONF object local to the middleware that will be an exact copy of what the service will have already loaded. In your test so far this has only required project_name which might be enough because i don't know how often services specify their own default_config_files[1] and it looks like CONF will use sys.argv[1:] by default [2] for parsing CLI arguments. This conf loading code would then have to be replicated into all middleware that wants to use oslo.config for its options.  

Fixing this problem is always where i loose enthusiasm for removing global CONF files.

>From a developer perspective I feel the solution is for us to reconsider how we deploy and configure middleware. If you are using paste those pieces of middleware should each be able to be configured without any integration with the service underneath it. Otherwise if your service needs a piece of middleware like auth_token middleware to operate or relies on oslo.config options like cors then that is not something that should be controlled by paste.

>From a deployer perspective there is no great answer i just want everything to be in oslo.config. 

Equally from a deployer perspective this wasn't an issue until aodh (et al) decided to remove the global CONF object which developers from all the projects hate but live with. I don't mean to imply we shouldn't look at better ways to do things, but if you're main concern is deployers the easiest thing we can do for consistency is add back the global CONF object or remove paste from aodh. :)

Jamie 


[1] https://git.openstack.org/cgit/openstack/oslo.config/tree/oslo_config/cfg.py#n1832
[2] https://git.openstack.org/cgit/openstack/oslo.config/tree/oslo_config/cfg.py#n1880
 
> 
> Cheers,
> 
> [1] https://github.com/openstack/gnocchi/blob/master/gnocchi/rest/app.py#L140
> [2]
> https://github.com/openstack/gnocchi/blob/master/gnocchi/service.py#L64-L73
> [3]
> https://github.com/openstack/zaqar/blob/87fd1aa93dafb64097f731dbd416c2eeb697d403/zaqar/transport/auth.py#L63
> [4]
> https://github.com/openstack/zaqar/blob/87fd1aa93dafb64097f731dbd416c2eeb697d403/zaqar/transport/auth.py#L70
> [5] https://review.openstack.org/#/c/208965/
> [6] https://review.openstack.org/#/c/209817/
> 
> 
> --
> Mehdi Abaakouk
> mail: sileht at sileht.net
> irc: sileht
> 
> __________________________________________________________________________
> OpenStack Development Mailing List (not for usage questions)
> Unsubscribe: OpenStack-dev-request at lists.openstack.org?subject:unsubscribe
> http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev
>