On Mar 3, 2020, at 12:55 PM, Tim Bell <tim.bell@cern.ch> wrote:
On 3 Mar 2020, at 19:20, Albert Braden <Albert.Braden@synopsys.com> wrote:
Sean, thank you for clarifying that.
Was my understanding that the community decided to focus on the unified client incorrect? Is the unified/individual client debate still a matter of controversy? Is it possible that the unified client will be deprecated in favor of individual clients after more discussion? I haven’t looked at any of the individual clients since 2018 (except for osc-placement which is kind of a special case), because I thought they were all going away and could be safely ignored until they did, and I haven’t included any information about the individual clients in the documentation that I write for our users, and if they ask I have been telling them to not use the individual clients. Do I need to start looking at individual clients again, and telling our users to use them in some cases?
I remember a forum discussion where a community goal was proposed to focus on OSC rather than individual project CLIs (I think Matt and I were proposers). There were concerns on the effort to do this and that it would potentially be multi-cycle.
My experience in discussion with the CERN user community and other OpenStack operators is that OSC is felt to be the right solution for the end user facing parts of the cloud (admin commands could be another discussion if necessary). Experienced admin operators can remember that glance looks after images and nova looks after instances. Our average user can get very confused, especially given that OSC supports additional options for authentication (such as Kerberos and Certificates along with clouds.yaml) so users need to re-authenticate with a different openrc to work on their project.
While I understand there are limited resources all round, I would prefer that we focus on adding new project functions to OSC which will eventually lead to feature parity.
Attracting ‘drive-by’ contributions from operations staff for OSC work (it's more likely to be achieved if it makes the operations work less e.g. save on special end user documentation by contributing code). This is demonstrated from the CERN team contribution to the OSC ‘coe' and ‘share' functionality along with lots of random OSC updates as listed hat https://www.stackalytics.com/?company=cern&metric=commits&module=python-openstackclient)
We’ve been working in SDK also to empower more people directly by being a bit more liberal with core. I think it’s time to start applying this approach to OSC as well. It’s never going to work to require the OSC team to implement everything, but neither is it super awesome to completely decentralize as the plugin/entrypoints issues have shown. I think SDK has been happy with blessing service humans rather quickly.
BTW, I also would vote for =auto as the default
This is what the case will be as we move towards replacing more and more of OSC’s guts with SDK. But let me describe it slightly differently: The way this works in SDK is that there is ONE user interface, which wants to track the latest as best as it can. But we can’t just do “auto” - because microversions can introduce breaking changes, so we need to add support to SDK for the most recent microversion we’re aware of. Then SDK negotiates to find the best microversion that is understands, and it always uses that. SDK has the POV that an end-user should almost never need to care about a micro version - if a user cares they are either in nova-core, or we’ve done something wrong. Case in point is this: https://opendev.org/openstack/openstacksdk/src/branch/master/openstack/compu... The nova team rightfully changed the semantics of live migrate because of safety. Mriedem put together the logic to express what the appropriate behavior would be, given a set of inputs, across the range of versions so that a user can do things and they’ll work. The end result is a live_migrate call that works across versions as safely as it can. I mention all of this because getting this work done was one of the key things we wanted to get right before we started transitioning OSC in earnest. It’s there - it works, and it’s being used across nova and ironic. So - I hear what people want from OSC - they want a thing that behaves like auto does. We agree - and the mechanism that makes us able to do that _safely_ is things like the above.
Tim
We are on Rocky now but I expect that we will upgrade as necessary to stay on supported versions.
From: Sean McGinnis <sean.mcginnis@gmx.com> Sent: Tuesday, March 3, 2020 9:50 AM To: openstack-discuss@lists.openstack.org Subject: Re: OSC future (formerly [glance] Different checksum between CLI and curl)
On 3/3/20 11:28 AM, Albert Braden wrote: Am I understanding correctly that the Openstack community decided to focus on the unified client, and to deprecate the individual clients, and that the Glance team did not agree with this decision, and that the Glance team is now having a pissing match with the rest of the community, and is unilaterally deciding to continue developing the Glance client and refusing to work on the unified client, or is something different going on? I would ask everyone involved to remember that we operators are down here, and the yellow rain falling on our heads does not smell very good. I definitely would not characterize it that way.
With trying not to put too much personal bias into it, here's what I would say the situation is:
- Some part of the community has said OSC should be the only CLI and that individual CLIs should go away - Glance is a very small team with very, very limited resources - The OSC team is a very small team with very, very limited resources - CLI capabilities need to be exposed for Glance changes and the easiest way to get them out for the is by updating the Glance CLI - No one from the OSC team has been able to proactively help to make sure these changes make it into the OSC client (see bullet 3) - There exists a sizable functionality gap between per-project CLIs and what OSC provides, and although a few people have done a lot of great work to close that gap, there is still a lot to be done and does not appear the gap will close at any point in the near future based on the current trends