On Thu, Oct 15, 2015 at 3:04 AM, Alexander Tivelkov
<ativel...@mirantis.com> wrote:
> Hi folks,
> I’ve noticed that the Community Application Catalog has begun to implement
> its own API, and it seems to me that we are going to have some significant
> duplication of efforts with the code which has already been implemented in
> Glance as Artifact Repository initiative (also known as Glance V3).
> From the very beginning of the App Catalog project (and I’ve been involved
> in it since February) I’ve been proposing to use Glance as its backend,
> because from my point of view it covers like 90% of the needed
> functionality. But it looks like we have some kind of miscommunication here,
> as I am getting some confusing questions and assumptions, like the vision of
> Glance V3 being dedicated backend for Murano (which is definitely
> incorrect).
> So, I am writing the email to clarify my vision of what Glance V3 is and how
> its features may be used to provide the REST API for Community App Catalog.
> 1.  Versioned schema
> First of all, Glance V3 operates on entities called “artifacts”, and these
> ones perfectly map to the Data Assets of community app catalog. The
> artifacts are strongly typed: there are artifact types for murano packages,
> glance images, heat templates - and there may be (and will be) more. Each
> artifact type is represented by a plugin, defining the schema of objects’
> data and metadata and - optionally - custom logic. So, this thing is
> extensible: when a new type of asset needs to be added to a catalog it can
> be done really quickly by just defining the schema and putting that schema
> into a plugin. Also, these plugins are versioned, so the possible changes in
> the schema are handled properly.
> 2. Generic properties
> Next, all the artifact types in Glance V3 have some generic metadata
> properties (i.e. part of the schema which is common for all the types),
> including the name, the version, description, authorship information and so
> on. This also corresponds to the data schema of community app catalog. The
> mapping is not 1:1, but we can work together on this to make sure that these
> generic properties match the expectations of the catalog.
> 3. Versioning
> Versions are very important for catalogs of objects, so Glance V3 was
> initially designed keeping versioning questions in mind: each artifact has a
> semver-based version assigned, so the artifacts having the same name but
> different versions are grouped into the proper sequences. API is able to
> query artifacts based on their version spec, e.g. it is possible to fetch
> the latest artifact with the name “foo” having the version greater than 2.1
> and less than 3.5.7 - or any other version spec, similar to pip or any other
> similar tool. As far as I know, community app catalog does not have such
> capabilities right now - and I strongly believe that it is really a must
> have feature for a catalog to be successful. At least it is absolutely
> mandatory for Murano packages, which are the only “real apps” among the
> asset types right now.
> 4. Cross artifact dependencies
> Glance V3 also has the dependency relations from the very beginning, so they
> may be defined as part of artifact type schema. As a result, an artifact may
> “reference” any number of other artifacts with various semantic. For
> example, murano package may define a set of references to other murano
> packages and call it “requires” - and this will act similar to the
> requirements of a python package. Similar properties may be defined for heat
> templates and glance images - they may reference each other with various
> semantics.
> Of course, the definitions of such dependencies may be done internally
> inside the packages, so they may be resolved locally by the service which is
> going to use it, but letting the catalog know about them will allow us to do
> the import-export operations for any given artifacts and its dependencies
> automatically, only by the means of the catalog itself.
> 5. Search and filtering API
> Right now Glance V3 API is in experimental state (we plan to stabilize it
> during the Mitaka cycle), but it already provides quite good capabilities to
> discover things. It can search artifacts by their type, name and
> (optionally) aforementioned version specs, by tag or even by arbitrary set
> of metadata properties. We have plans to integrate Glance V3 with the
> Searchlight project to have even more index and search capabilities using
> its elastic search engine.
> 6. Data storage
> As you probably know, Glance does not own the binary data of its images.
> Instead, it provides an abstraction of the backend storage, which may be
> swift, ceph, s3 or something else. The same approach is used in Glance V3
> for artifacts data, but with more per-type control: particular artifact
> types may be configured independently to store their blobs in different
> backends. This may be of use for Community App Catalog which operates on
> different storages for its assets.
> 7. Sharing and access control.
> Glance V3 inherits the same access mechanics present in Glance V2: an
> artifact may be visible to its owner tenant only, be public (i.e. visible to
> all the tenants) or directly shared by the owner to a specific tenant. Also,
> Glance can act in the anonymous mode (i.e. without an access token), thus
> providing access to public artifacts even to unauthenticated users.
> This can be easily applied to a public web service, such as community app
> catalog: regular unauthenticated users use anonymous mode to access only the
> public assets (this is the current behavior of apps.o.o), while registered
> users will have their own private spaces (“tenants”) with various access
> restrictions.
> 8. The federation.
> The ultimate goal for Glance Artifact Repository is ability to build trees
> of artifact repos in different clouds. The top node of that tree is some
> Global Application Catalog (and the apps.openstack.org may be this global
> catalog - if it is glance-based or at least supports glance v3 federation),
> then there are repositories of particular openstack vendors or distributors,
> then - the catalogs of enterprises operating different openstack clouds. The
> particular glance deployments in that clouds are the leafs of that tree,
> being able to search for data assets in all the upstream repositories,
> download them from there or - if permitted - submit their local assets back
> upstream. This will be the ultimate network for application delivery and
> exchange in openstack world - and this is one of the main reasons we’ve
> began the Artifacts initiative in Glance.
> Unlike other aforementioned features this one is not implemented yet, but we
> are planning to add it as soon as we are done with API stabilization goal.
> There are many other features which are present in V3’s roadmap and may be
> useful for the app catalog, such as ability to sign artifacts with their
> developers’ keys and verify that keys on usage to ensure the authenticity of
> the artifact.
> What we don’t have right now is the ability to associate ratings (“stars”)
> and comments to the artifact, as well as aggregating different usage and
> download statistics: such features are really needed only for the public
> website such as apps.o.o but are not required for Glance’s in particular
> clouds. But we may find some way to solve this, either by wrapping glance
> API with additional middleware which would add appropriate info from a
> different data source, or by having custom plugins which are able to do
> that, or in some other way: I am sure we may find a solution for this.
> So, this was just a brief description of what Glance v3 has to offer as a
> backend for App Catalog API.
> It also worths to mention that this API is in “EXPERIMENTAL” state right
> now, which means that it is not fixed and we may modify it significantly if
> there is a need to. So we may work closer together to adopt it for the needs
> of Community App Catalog.
> I would really prefer to not create any overlaps between Glance v3 and the
> community app catalog: if the app catalog builds its own incompatible
> implementation of assets discovery and distribution API then we’ll have a
> huge duplication of efforts for developers and lots of confusion to the
> end-users who will get two entirely different ways to do the same task.
> So, I’d propose to discuss these potential overlaps, look at the features
> need by App Catalog and see how Glance V3 may be of use here. I’ll be more
> than happy to help with that. We can dive deeper into the details here in
> the mailing list or meet in person in Tokyo. I'll try to have a
> demonstratable prototype by that time.

Alexander, you make some great points here about the feature overlaps
between the Glance v3 API and where we are heading with the App
Catalog.  There's good potential to re-use some of this work for sure,
and I'm really looking forward to what comes of the PoC you talked

Two important considerations against adopting Glance v3 wholesale
(aside from the experimental nature and the possibility that key bits
won't be ultimately adopted in the finished stable product) are
development velocity and deployment goals.  As glance is a key
component in any OpenStack cloud, and is expected to be deployed in
hundreds (or thousands!) of environments, the dependencies and default
requirements are tightly coupled to most other OpenStack projects.
Obviously any changes to glance are weighed against how those will
impact all these clouds, how upgrades will work, backwards and
forwards compatibility, etc.  For all the benefits that come with
working on a distributed open source project like this, it also tends
to mean we can't always move very fast because of the potential impact
across the landscape of the whole project.

The App Catalog is in a different position - to put it really simply,
we're developing a web site intended to benefit user of OpenStack
clouds, and including some integration points to make it much easier
for users to find and share things that run on OpenStack clouds.  But
we're not tightly coupled to any of the other projects, and we're not
expecting this to be something that's deployed on it's own with every
OpenStack cloud.  In fact I'd argue deploying a local/private App
Catalog dilutes the value and harms the ecosystem as a whole.

Ideally we'll find a balanced way to benefit from the great work
coming from the Glance team without limiting our design considerations
or encumbering the App Catalog with requirements or limitations that
shouldn't impact us.

I'm really looking forward to speaking more about this in person in
Tokyo, hopefully you can make it to the App Catalog sessions!


OpenStack Development Mailing List (not for usage questions)
Unsubscribe: openstack-dev-requ...@lists.openstack.org?subject:unsubscribe

Reply via email to