Open Stack

-----Original Message-----
From: Tim Bell <tim.bell at cern.ch>
Reply: OpenStack Development Mailing List (not for usage questions) <openstack-dev at lists.openstack.org>
Date: August 4, 2016 at 13:19:02
To: OpenStack Development Mailing List (not for usage questions) <openstack-dev at lists.openstack.org>
Subject:  Re: [openstack-dev] [Glance][TC][Heat][App-Catalog][Murano][Tacker] Glare as a new Project

>  
> > On 04 Aug 2016, at 19:34, Erno Kuvaja wrote:
> >
> > On Thu, Aug 4, 2016 at 5:20 PM, Clint Byrum wrote:
> >> Excerpts from Tim Bell's message of 2016-08-04 15:55:48 +0000:
> >>>
> >>> On 04 Aug 2016, at 17:27, Mikhail Fedosin >  
> wrote:
> >>>>
> >>>> Hi all,
> >>>>>> after 6 months of Glare v1 API development we have decided to continue
> >>>> our work in a separate project in the "openstack" namespace with its own
> >>>> core team (me, Kairat Kushaev, Darja Shkhray and the original creator -
> >>>> Alexander Tivelkov). We want to thank Glance community for their support
> >>>> during the incubation period, valuable advice and suggestions - this time
> >>>> was really productive for us. I believe that this step will allow the
> >>>> Glare project to concentrate on feature development and move forward
> >>>> faster. Having the independent service also removes inconsistencies
> >>>> in understanding what Glance project is: it seems that a single project
> >>>> cannot own two different APIs with partially overlapping functionality. So
> >>>> with the separation of Glare into a new project, Glance may continue its
> >>>> work on the OpenStack Images API, while Glare will become the reference
> >>>> implementation of the new OpenStack Artifacts API.
> >>>>
> >>>
> >>> I would suggest looking at more than just the development process when
> >>> reflecting on this choice.
> >>> While it may allow more rapid development, doing on your own will increase
> >>> costs for end users and operators in areas like packaging, configuration,
> >>> monitoring, quota … gaining critical mass in production for Glare will
> >>> be much more difficult if you are not building on the Glance install base.
> >>
> >> I have to agree with Tim here. I respect that it's difficult to build on
> >> top of Glance's API, rather than just start fresh. But, for operators,
> >> it's more services, more API's to audit, and more complexity. For users,
> >> they'll now have two ways to upload software to their clouds, which is
> >> likely to result in a large portion just ignoring Glare even when it
> >> would be useful for them.
> >>
> >> What I'd hoped when Glare and Glance combined, was that there would be
> >> a single API that could be used for any software upload and listing. Is
> >> there any kind of retrospective or documentation somewhere that explains
> >> why that wasn't possible?
> >>
> >
> > I was planning to leave this branch on it's own, but I have to correct
> > something here. This split is not introducing new API, it's moving the
> > new Artifact API under it's own project, there was no shared API in
> > first place. Glare was to be it's own service already within Glance
> > project. Also the Artifacts API turned out to be fundamentally
> > incompatible with the Images APIs v1 & v2 due to the totally different
> > requirements. And even the option was discussed in the community I
> > personally think replicating Images API and carrying the cost it being
> > in two services that are fundamentally different would have been huge
> > mistake we would have paid for long time. I'm not saying that it would
> > have been impossible, but there is lots of burden in Images APIs that
> > Glare really does not need to carry, we just can't get rid of it and
> > likely no-one would have been happy to see Images API v3 around the
> > time when we are working super hard to get the v1 users moving to v2.
> >
> > Packaging glance-api, glance-registry and glare-api from glance repo
> > would not change the effort too much compared from 2 repos either.
> > Likely it just makes it easier when the logical split it clear from
> > the beginning.
> >
> > What comes to Tim's statement, I do not see how Glare in it's own
> > service with it's own API could ride on the Glance install base apart
> > from the quite false mental image these two thing being the same and
> > based on the same code.
> >
>  
> To give a concrete use case, CERN have Glance deployed for images. We are interested in  
> the ecosystem
> around Murano and are actively using Heat. We deploy using RDO with RPM packages, Puppet-OpenStack  
> for configuration, a set of machines serving Glance in an HA set up across multiple data  
> centres and various open source monitoring tools.
>  
> The multitude of projects and the day two maintenance scenarios with 11 independent  
> projects is a cost and adding further to this cost for the production deployments of OpenStack  
> should not be ignored.
>  
> By Glare choosing to go their own way, does this mean that
>  
> - Can the existing RPM packaging for Glance be used to deploy Glare ? If there needs to be  
> new packages defined, this is additional cost for the RDO team (and the equivalent .deb  
> teams) or will the Glare team provide this ?
> - Can we use our existing templates for Glance for configuration management ? If there  
> need to be new ones defined, this is additional work for the Chef and Ansible teams or will  
> the Glare team provide this ?
> - Log consolidation and parsing using the various OsOps tools for Glance is in place …  
> a new project would need maintenance
> - If new endpoints need to be defined, this is additional work for the operators to allocate  
> appropriate endpoints and HAProxy tweaks
> - Additional database endpoints need to be defined, backed up, configured etc.
> - Install / Config guides are an effort, adding plugins to the OpenStack CLI, Horizon  
> panels …
>  
> I do not feel that sufficient consideration of the cost of the split for the consumers  
> of OpenStack has been considered. It may be the right decision but baseing purely on the  
> development speed and effort is ignoring a significant set of stakeholders.

I was not part of the discussion, but I've been part of Glance/Glare conversations in the past. I think Mike's initial message gives you the main reason why Glare is splitting from Glance: speed of merges.

Most of the vendors employing Glance cores do not see it as a priority in any sense of the word. That means that most of us on the core team have very little to do reviews. I for one tend to focus on stability and bug fix reviews when I can find the time (as well as stable branch reviews as a stable reviewer). On top of most vendors not being very interested in Glance, there's also the problem that the vendor most interested in Glare has always been Mirantis. They've contributed all of the specs, all of the code, etc. Very few other vendors can be bothered with Glare and so most of the cores have tried to give reviews but there just isn't enough time.

Since an Artifacts API and Murano are both priorities for Mirantis, they're deciding to split this off to a project of their own. Make no mistake, if any other vendors had dedicated core reviewer resources to Glare, I doubt this would be happening.

It's also worth noting that I don't blame Mirantis for this. This is a problem that happened with Searchlight and Glance. There was a single vendor that was sufficiently interested in Searchlight to invest resources. They suffered the same fate and decided to split it off into its own project for similar reasons.

--  
Ian Cordasco