← Back to team overview

openstack team mailing list archive

Re: [Keystone] What exactly are we modeling with endpoints?

 

A common use-case for a "service" to have more than one "endpoints" is when there are different regions for different endpoints serving the same service.  This is potentially a place where the endpoints may be discriminated, if the consumer of the catalogue is "geo"-aware.

Liem

From: antiver@xxxxxxxxx [mailto:antiver@xxxxxxxxx] On Behalf Of Dolph Mathews
Sent: Wednesday, April 25, 2012 11:00 AM
To: Joseph Heck
Cc: Nguyen, Liem Manh; Adam Gandelman; openstack@xxxxxxxxxxxxxxxxxxx (openstack@xxxxxxxxxxxxxxxxxxx)
Subject: Re: [Openstack] [Keystone] What exactly are we modeling with endpoints?

I'm not sure it makes sense for a "service" to have more than one "endpoint." Two different endpoints might as well be two separate services IMO, even if they directly affect each other.

For example, creating a tenant on the identity admin endpoint makes that tenant just as available on the identity public endpoint as it does on the compute public endpoint, but of course it doesn't make sense to say all of nova is suddenly part of keystone.

There are (at least) two attributes that must be provided to distinguish between two endpoints: the API "type" and the endpoint/url "type." ("admin + identity" vs "public + compute" vs "internal + volumes"). Trying to group endpoints at a high level by API type alone strikes me as very arbitrary, at best.

If I'm going to "collect" anything, it would be which endpoints work (or work *best*) with which *other* endpoints. See: https://blueprints.launchpad.net/keystone/+spec/service-relationships

-Dolph

On Wed, Apr 25, 2012 at 11:47 AM, Joseph Heck <heckj@xxxxxx<mailto:heckj@xxxxxx>> wrote:
This isn't about parsing the data structure - it's about what a "service" represents so that we can model it appropriately. So what does the "service" here represent? the collection of all possible services of that type? That's what your example would seem to indicate.

In your example, the service is a pretty simple structure - just having a type (driven by convention and API spec) and human readable name, and each service is expected to have 1 to N endpoints.

Is it reasonable to have a service without any endpoints? Regardless of reasonable, is it allowable?

What does an endpoint represent? The API's URI point, clearly. Is there a uniqueness constraint of any kind on endpoints? Is it allowable (if strange) to list 3 duplicate endpoints with exactly the same metadata on it?

-joe

On Apr 25, 2012, at 9:37 AM, Nguyen, Liem Manh wrote:
I would like to keep the service type  and name under the service and not the endpoint, too.  Make it easier to parse for a given service.

One thing is that I am not sure if we need the metadata tag... In the Keystone XSD, we have the construct <anyAttribute namespace="##other" processContents="lax"/>, which allows any additional, implementation-specific attribute to be added.  Those that do not support the specific attribute can simply ignore it.   A couple of benefits I can see with not using the metadata tag, and just use the custom element directly like this:  http://paste.openstack.org/show/13832/, which the anyAttribute supports, are:

*         Simplier parsing, one level less.
*         If that attribute becomes a core attribute later, no need to change the parser.

Liem

From: openstack-bounces+liem_m_nguyen=hp.com@xxxxxxxxxxxxxxxxxxx<mailto:openstack-bounces+liem_m_nguyen=hp.com@xxxxxxxxxxxxxxxxxxx> [mailto:openstack-bounces+liem_m_nguyen<mailto:openstack-bounces%2Bliem_m_nguyen>=hp.com@xxxxxxxxxxxxxxxxxxx<mailto:hp.com@xxxxxxxxxxxxxxxxxxx>] On Behalf Of Joe Savak
Sent: Tuesday, April 24, 2012 1:04 PM
To: Joseph Heck; openstack@xxxxxxxxxxxxxxxxxxx<mailto:openstack@xxxxxxxxxxxxxxxxxxx> (openstack@xxxxxxxxxxxxxxxxxxx<mailto:openstack@xxxxxxxxxxxxxxxxxxx>)
Cc: Adam Gandelman
Subject: Re: [Openstack] [Keystone] What exactly are we modeling with endpoints?

Having endpoints under the service construct is supposed to make it easier to programmatically find the endpoint(s) you are interested in.

For example - as nova client I can parse the service catalog and identity nova by service-type "compute" in order to get the public, internal, and admin endpoints for nova.

By having service type & name as attributes under the endpoint, I'll have a harder time doing that (having to dive into each endpoint construct to identify the ones with service-type "compute").
Maybe it would be better to have each endpoint have its own construct inside of a service.

So instead of http://paste.openstack.org/show/13678/
Maybe http://paste.openstack.org/show/13682/


From: openstack-bounces+joe.savak=rackspace.com@xxxxxxxxxxxxxxxxxxx<mailto:openstack-bounces+joe.savak=rackspace.com@xxxxxxxxxxxxxxxxxxx> [mailto:openstack-bounces+joe.savak<mailto:openstack-bounces%2Bjoe.savak>=rackspace.com@xxxxxxxxxxxxxxxxxxx<mailto:rackspace.com@xxxxxxxxxxxxxxxxxxx>] On Behalf Of Joseph Heck
Sent: Friday, April 20, 2012 4:16 PM
To: openstack@xxxxxxxxxxxxxxxxxxx<mailto:openstack@xxxxxxxxxxxxxxxxxxx> (openstack@xxxxxxxxxxxxxxxxxxx<mailto:openstack@xxxxxxxxxxxxxxxxxxx>)
Cc: Adam Gandelman
Subject: [Openstack] [Keystone] What exactly are we modeling with endpoints?

While I've been roaming about the summit and conference, I've been trying to figure out exactly what we're modeling with the current "service" and "endpoints" that are in the API today. After talking with a number of folks, it's getting clearer that how it's being used is very installation specific.

I'd like to simplify this aspect of the API if at all possible, especially with a lot of the good ideas around describing the relationships between endpoints and and their installation.

The use cases I'm hearing actively in use are:

* (Horizon/UI/client) To indicate to a user where they can go to access their data
* (Glance, Nova, Keystone client) to find the endpoint relevant to uploading images (current client implementations appear to assume there is only one image endpoint)

The use case to indicate a geographic location for a datacenter or "cloud" is not consistent - some implementations I've learned of have that feature (and use "Region" for that sort of information), and others are load balancing a single endpoint to deploy to multiple datacenters and geographic regions from a single endpoint.

At the summit and conference, I heard a desire to expose geographic information with the endpoints, but that is clearly an operator specific implementation/deployment detail. Likewise I heard a lot of "We could really..." if additional metadata was easily available on endpoints, again in fairly implementation/deployment specific detail.

So looking forward towards a v.next API, what do you all think about having just "endpoints", with everything else being attributes on those endpoints (including what "service" and "type" it is), with some expected conventions (that there are a few well defined types - such as PublicURL and InternalURL, and relevant names for the rest API endpoints (ec2, compute, volume, image, identity...)

Additional metadata can then float on the endpoints in deployment/implementation specific ways that don't lock in other systems to be deployed and implemented in the same fashion.

-joe


On Apr 20, 2012, at 1:47 PM, Lorin Hochstein wrote:
On Apr 13, 2012, at 12:34 PM, Adam Gandelman wrote:
On 04/13/2012 10:50 AM, Dolph Mathews wrote:
While $(tenant_id)s is certainly the documented syntax, it appears that the SQL catalog backend (and *only* the SQL catalog backend, as far as I can tell) explicitly supports both $(tenant_id)s and %(tenant_id)s:

https://github.com/openstack/keystone/blob/master/keystone/catalog/backends/sql.py#L163

Perhaps Adam Gandelman has some insight?

-Dolph

Dolph-

No, the same is supported in the case of templated catalog as well, which is what the SQL catalog was largely based off:

https://github.com/openstack/keystone/blob/master/keystone/catalog/backends/templated.py#L115

Just tested that "sed -i 's/\$/%/g' /etc/keystone/default_catalog.templates" still produces a functional service catalog when configured to use the templated backend.

Seeing as both are supported, perhaps it would be better for docs to be updated to refer to the use of % instead of $ to avoid people running into problems with the $() sub-shell?

The OpenStack Install and Deploy manual has some language about this (see last paragraph): http://docs.openstack.org/trunk/openstack-compute/install/content/elements-of-keystone-service-catalog-entry.html

This hasn't made its way into the admin docs yet, though.


Take care,

Lorin
--
Lorin Hochstein
Lead Architect - Cloud Services
Nimbis Services, Inc.
www.nimbisservices.com<https://www.nimbisservices.com/>



_______________________________________________
Mailing list: https://launchpad.net/~openstack
Post to     : openstack@xxxxxxxxxxxxxxxxxxx<mailto:openstack@xxxxxxxxxxxxxxxxxxx>
Unsubscribe : https://launchpad.net/~openstack
More help   : https://help.launchpad.net/ListHelp


_______________________________________________
Mailing list: https://launchpad.net/~openstack
Post to     : openstack@xxxxxxxxxxxxxxxxxxx<mailto:openstack@xxxxxxxxxxxxxxxxxxx>
Unsubscribe : https://launchpad.net/~openstack
More help   : https://help.launchpad.net/ListHelp


References