Open Stack

Thank you for a reply, Flavia:

> Hi Bogdan
> sorry for the late reply - yesterday was a Holiday here in Brazil!
> I am afraid I will not be able to engage in this collaboration with
> such a short time...we had to have started this initiative a little
> earlier...

That's understandable.

I hoped though a position paper is something we (all who reads that, not 
just you and me) could achieve in a couple of days, without a lot of 
research associated. That's a postion paper, which is not expected to 
contain formal prove or implementation details. The vision for tooling 
is the hardest part though, and indeed requires some time.

So let me please [tl;dr] the outcome of that position paper:

* position: given Always Available autonomy support as a starting point,
   define invariants for both operational and data storage consistency
   requirements of control/management plane (I've already drafted some in
   [0])

* vision: show that in the end that data synchronization and conflict
   resolving solution just boils down to having a causally
   consistent KVS (either causal+ or causal-RT, or lazy replication
   based, or anything like that), and cannot be achieved with *only*
   transactional distributed database, like Galera cluster. The way how
   to show that is an open question, we could refer to the existing
   papers (COPS, causal-RT, lazy replication et al) and claim they fit
   the defined invariants nicely, while transactional DB cannot fit it
   by design (it's consensus protocols require majority/quorums to
   operate and being always available for data put/write operations).
   We probably may omit proving that obvious thing formally? At least for
   the postion paper...

* opportunity: that is basically designing and implementing of such a
   causally-consistent KVS solution (see COPS library as example) for
   OpenStack, and ideally, unifying it for PaaS operators
   (OpenShift/Kubernetes) and tenants willing to host their containerized
   workloads on PaaS distributed over a Fog Cloud of Edge clouds and
   leverage its data synchronization and conflict resolving solution
   as-a-service. Like Amazon dynamo DB, for example, except that fitting
   the edge cases of another cloud stack :)

[0] 
https://github.com/bogdando/papers-ieee/blob/master/ICFC-2019/challenges.md

> As for working collaboratively with latex, I would recommend using
> overleaf - it is not that difficult and has lots of edition resources
> as markdown and track changes, for instance.
> Thanks and good luck!
> Flavia



On 11/2/18 5:32 PM, Bogdan Dobrelya wrote:
> Hello folks.
> Here is an update for today. I crated a draft [0], and spend some time 
> with building LaTeX with live-updating for the compiled PDF... The 
> latter is only informational, if someone wants to contribute, please 
> follow the instructions listed by the link (hint: you need no to have 
> any LaTeX experience, only basic markdown knowledge should be enough!)
> 
> [0] 
> https://github.com/bogdando/papers-ieee/#in-the-current-development-looking-for-co-authors 
> 
> 
> On 10/31/18 6:54 PM, Ildiko Vancsa wrote:
>> Hi,
>>
>> Thank you for sharing your proposal.
>>
>> I think this is a very interesting topic with a list of possible 
>> solutions some of which this group is also discussing. It would also 
>> be great to learn more about the IEEE activities and have experience 
>> about the process in this group on the way forward.
>>
>> I personally do not have experience with IEEE conferences, but I’m 
>> happy to help with the paper if I can.
>>
>> Thanks,
>> Ildikó
>>
>>
> 
> (added from the parallel thread)
>>> On 2018. Oct 31., at 19:11, Mike Bayer <mike_mp at zzzcomputing.com> 
>>> wrote:
>>>
>>> On Wed, Oct 31, 2018 at 10:57 AM Bogdan Dobrelya <bdobreli at 
>>> redhat.com> wrote:
>>>>
>>>> (cross-posting openstack-dev)
>>>>
>>>> Hello.
>>>> [tl;dr] I'm looking for co-author(s) to come up with "Edge clouds data
>>>> consistency requirements and challenges" a position paper [0] (papers
>>>> submitting deadline is Nov 8).
>>>>
>>>> The problem scope is synchronizing control plane and/or
>>>> deployments-specific data (not necessary limited to OpenStack) across
>>>> remote Edges and central Edge and management site(s). Including the 
>>>> same
>>>> aspects for overclouds and undercloud(s), in terms of TripleO; and 
>>>> other
>>>> deployment tools of your choice.
>>>>
>>>> Another problem is to not go into different solutions for Edge
>>>> deployments management and control planes of edges. And for tenants as
>>>> well, if we think of tenants also doing Edge deployments based on Edge
>>>> Data Replication as a Service, say for Kubernetes/OpenShift on top of
>>>> OpenStack.
>>>>
>>>> So the paper should name the outstanding problems, define data
>>>> consistency requirements and pose possible solutions for 
>>>> synchronization
>>>> and conflicts resolving. Having maximum autonomy cases supported for
>>>> isolated sites, with a capability to eventually catch up its 
>>>> distributed
>>>> state. Like global database [1], or something different perhaps (see
>>>> causal-real-time consistency model [2],[3]), or even using git. And
>>>> probably more than that?.. (looking for ideas)
>>>
>>>
>>> I can offer detail on whatever aspects of the "shared  / global
>>> database" idea.  The way we're doing it with Galera for now is all
>>> about something simple and modestly effective for the moment, but it
>>> doesn't have any of the hallmarks of a long-term, canonical solution,
>>> because Galera is not well suited towards being present on many
>>> (dozens) of endpoints.     The concept that the StarlingX folks were
>>> talking about, that of independent databases that are synchronized
>>> using some kind of middleware is potentially more scalable, however I
>>> think the best approach would be API-level replication, that is, you
>>> have a bunch of Keystone services and there is a process that is
>>> regularly accessing the APIs of these keystone services and
>>> cross-publishing state amongst all of them.   Clearly the big
>>> challenge with that is how to resolve conflicts, I think the answer
>>> would lie in the fact that the data being replicated would be of
>>> limited scope and potentially consist of mostly or fully
>>> non-overlapping records.
>>>
>>> That is, I think "global database" is a cheap way to get what would be
>>> more effective as asynchronous state synchronization between identity
>>> services.
>>
>> Recently we’ve been also exploring federation with an IdP (Identity 
>> Provider) master: 
>> https://wiki.openstack.org/wiki/Keystone_edge_architectures#Identity_Provider_.28IdP.29_Master_with_shadow_users 
>>
>>
>> One of the pros is that it removes the need for synchronization and 
>> potentially increases scalability.
>>
>> Thanks,
>> Ildikó
> 
> 
> 


-- 
Best regards,
Bogdan Dobrelya,
Irc #bogdando