Open Stack

On 12/06/2013 01:38 PM, Clint Byrum wrote:
> Excerpts from Jay Pipes's message of 2013-12-05 21:32:54 -0800:
>> On 12/05/2013 04:25 PM, Clint Byrum wrote:
>>> Excerpts from Andrew Plunk's message of 2013-12-05 12:42:49
>>> -0800:
>>>>> Excerpts from Randall Burt's message of 2013-12-05 09:05:44
>>>>> -0800:
>>>>>> On Dec 5, 2013, at 10:10 AM, Clint Byrum <clint at
>>>>>> fewbar.com> wrote:
>>>>>>
>>>>>>> Excerpts from Monty Taylor's message of 2013-12-04
>>>>>>> 17:54:45 -0800:
>>>>>>>> Why not just use glance?
>>>>>>>>
>>>>>>>
>>>>>>> I've asked that question a few times, and I think I can
>>>>>>> collate the responses I've received below. I think
>>>>>>> enhancing glance to do these things is on the table:
>>>>>>>
>>>>>>> 1. Glance is for big blobs of data not tiny templates. 2.
>>>>>>> Versioning of a single resource is desired. 3.
>>>>>>> Tagging/classifying/listing/sorting 4. Glance is designed
>>>>>>> to expose the uploaded blobs to nova, not users
>>>>>>>
>>>>>>> My responses:
>>>>>>>
>>>>>>> 1: Irrelevant. Smaller things will fit in it just fine.
>>>>>>
>>>>>> Fitting is one thing, optimizations around particular
>>>>>> assumptions about the size of data and the frequency of
>>>>>> reads/writes might be an issue, but I admit to ignorance
>>>>>> about those details in Glance.
>>>>>>
>>>>>
>>>>> Optimizations can be improved for various use cases. The
>>>>> design, however, has no assumptions that I know about that
>>>>> would invalidate storing blobs of yaml/json vs. blobs of
>>>>> kernel/qcow2/raw image.
>>>>
>>>> I think we are getting out into the weeds a little bit here. It
>>>> is important to think about these apis in terms of what they
>>>> actually do, before the decision of combining them or not can
>>>> be made.
>>>>
>>>> I think of HeatR as a template storage service, it provides
>>>> extra data and operations on templates. HeatR should not care
>>>> about how those templates are stored. Glance is an image
>>>> storage service, it provides extra data and operations on
>>>> images (not blobs), and it happens to use swift as a backend.
>>>>
>>>> If HeatR and Glance were combined, it would result in taking
>>>> two very different types of data (template metadata vs image
>>>> metadata) and mashing them into one service. How would adding
>>>> the complexity of HeatR benefit Glance, when they are dealing
>>>> with conceptually two very different types of data? For
>>>> instance, should a template ever care about the field "minRam"
>>>> that is stored with an image? Combining them adds a huge
>>>> development complexity with a very small operations payoff, and
>>>> so Openstack is already so operationally complex that HeatR as
>>>> a separate service would be knowledgeable. Only clients of Heat
>>>> will ever care about data and operations on templates, so I
>>>> move that HeatR becomes it's own service, or becomes part of
>>>> Heat.
>>>>
>>>
>>> I spoke at length via G+ with Randall and Tim about this earlier
>>> today. I think I understand the impetus for all of this a little
>>> better now.
>>>
>>> Basically what I'm suggesting is that Glance is only narrow in
>>> scope because that was the only object that OpenStack needed a
>>> catalog for before now.
>>>
>>> However, the overlap between a catalog of images and a catalog
>>> of templates is quite comprehensive. The individual fields that
>>> matter to images are different than the ones that matter to
>>> templates, but that is a really minor detail isn't it?
>>>
>>> I would suggest that Glance be slightly expanded in scope to be
>>> an object catalog. Each object type can have its own set of
>>> fields that matter to it.
>>>
>>> This doesn't have to be a minor change to glance to still have
>>> many advantages over writing something from scratch and asking
>>> people to deploy another service that is 99% the same as Glance.
>>
>> My suggestion for long-term architecture would be to use Murano
>> for catalog/metadata information (for images/templates/whatever)
>> and move the block-streaming drivers into Cinder, and get rid of
>> the Glance project entirely. Murano would then become the
>> catalog/registry of objects in the OpenStack world, Cinder would be
>> the thing that manages and streams blocks of data or block devices,
>> and Glance could go away. Imagine it... OpenStack actually
>> *reducing* the number of projects instead of expanding! :)
>
> Have we not learned our lesson with Nova-Net/Neutron yet? Rewrites
> of existing functionality are painful.

I said *long-term* above, Clint.

> The Murano-concerned people have already stated they are starting
> over on that catalog.
>
> I suggest they start over by expanding Glance's catalog. If the
> block streaming bits of Glance need to move somewhere else, that
> sounds like a completely separate concern that distracts from this
> point.

Yes, the block streaming bits of Glance are separate -- sorry for
distracting from your point. However, Glance's architecture actually
went from having the registry and streaming pieces separated to having
the two pieces more closely interwoven in the last two release cycles.
It's harder now, IMO, to separate out the "registry" aspects of Glance 
from the block management pieces. Which is why I suggested using Murano 
as the catalog since there is already a separate catalog sub-project in 
Murano and the Murano devs are familiar with this space.

Please try not to shoot the messenger here -- I'm only bringing up my 
thoughts on where the various related projects seem to be moving.

> And to be clear, (I think I will just stop talking as I think I've
> made this point), my point is, we have a catalog, let's make it
> better.

Well, we have multiple catalogs... which one shall we make better? 
Sounds like contributors are coalescing around using Glance as the 
catalog -- which is fine I suppose -- but I think it would have been 
easier to use the old Glance Registry server (which now does not exist) 
as this catalog service, than the newer Glance API, which has gotten 
sidetracked with things like image tasks (which IMO should never have 
been added to the Images API; the taskflow or Mistral projects should 
have handled this functionality).

Anyway, just figured I'd offer some commentary. Sorry if you think I was 
causing trouble.

Best,
-jay