[Edge-computing] Afterlife of the PTG edge discussions
lebre.adrien at free.fr
lebre.adrien at free.fr
Wed Feb 28 15:07:15 UTC 2018
> I can help in organizing the notes after the PTG workshop.
> I think both of these are valuable information and we should somehow
> have the following things:
> * a description about what we would like to achieve. Maybe in
> some other format than a list of etherpads.
First, I think the monday's list should be completed with the additional informations that have been identified through the allan's problem
- being able to define roles per regions/sites (one user may have some specific rights on one site while being not allowed to use specific services on another one.
Moreover, we should probably redistribute the different requirements according to the perspective of an end-user/developer vs an administrator (right now, most of the admin operations belong to Level 5, not sure this is the best way to classify, especially if the goal is to class each operation according to the complexity level).
> * a list of concrete requirements to specific projects (being
> existing or something new)
Yes, maybe this requires a bit of works from our side but I think you're right: it is definitely the right way to go.
> * maybe some prototypes based on the Tuesday afternoon
> discussions for the keystone data and image distribution (by the
> way can someone post the picture of THE PLAN?)
I think Ildiko took a photo of the Jonathan's diagram.
>From my side to be honest, there are still some points that are unclear (i.e. or at least I need more time to be convinced before starting prototyping :-P).
- Using this new NKB means that each edge site will be passive (in the sense that you cannot create a new role/a new VM image/a new... without going through the NKB service? What will happen if someone wants to create a new VM image from the edge but the NKB is not reachable? Don't get me wrong, I'm convinced that being non intrusive is good (so developing a mechanism that enables collaborations between independent edge sites by leveraging only APIs makes sense). The doubts I have are regarding how such mechanisms can be implemented while mitigating as much as possible (i) communications between edge sites and this service and (ii) possible inconsistency issues.
- At the end, we also briefly mentioned that the NKB will ensure some guarantees (i.e. like an autonomous checker in charge of maintaining the presence/availability of specific images across the sites). This also makes sense but looks to be an advanced self-healing feature. Ok, Why not? If we can do it it is great. However, I'm bit scared about what will be present at the end inside this NKB service (and all golden* services)?
- What is the difference between such a new NKB and the tricircle proposal (before splitting) is also something I think it would be valuable to better understand?
- I didn't attend the nova session unfortunately this morning but I get the feeling that we can also learn a lot from the questions they have been discussing for more than two years regarding the cell V2 architecture (see for instance  line 37 the discussion regarding glance for all cells vs one glance per cells vs....). Moving forward on the question of being able to benefit from the locality aspects, which is also an advanced feature from my viewpoint (i.e. instead of pulling my image from the golden glance, I would like to get it from my neighbourhood). I'm wondering how this run nowadays with Glance on a large cluster (are Glance/Nova rack aware: if an image is available on one compute node and you want to provision a new VM with the same image on a server which belongs to the same rack. Is Glance/Nova pulling the image from the official URI or from the compute node in order to mitigate the overhead on the network).
Maybe we can leverage either the edge sessions organised by the foundation and the FEMDC IRC meetings to dig a bit more of this architecture and to better understand what does it mean also for Nova and Neutron (the discussion focused on keystone and glance yesterday).
This can lead to a possible reference architecture/blueprint with well-identified requirements for the different projects.
We can setup an etherpad and try to prepare the different questions we may all have.
Last but not the least, I really appreciated the discussion with Alan yesterday (thanks BTW ;)). It enabled us to highlight/discuss a few problems that ATT have to deal with. I'm wondering whether we cannot have such an exchange with some other folks. I know OVH for instance are operating several regions also. I think it would be valuable to understand how they deal with all the administrator challenges Alan pointed yesterday and whether they are also facing any other problems. Maybe the foundation can help us to organise such an exchange. Jonathan?
My two cents,
PS: thanks for starting the discussion!
> Any opinions?
More information about the Edge-computing