On Mar 2, 2018, at 4:48 PM, Dan Kohn <dan@...> wrote:

CNCF has three precedents of separate specs and implementations:

+ CNI and the CNI plugins (most prominently Calico, Flannel and Weave Net, none of which are yet CNCF projects)

+ OpenTracing and Jaeger

+ TUF and Notary

So, the example of CSI as the spec and REX-Ray as an implementation seems feasible. Whether it is advisable is, of course, up to the TOC.

--

Dan Kohn <dan@...>

Executive Director, Cloud Native Computing Foundation https://www.cncf.io

+1-415-233-1000 https://www.dankohn.com

On Fri, Mar 2, 2018 at 5:36 PM, Mueller, Garrett <Garrett.Mueller@...> wrote:

I see where you’re coming from Clint, but in this case I agree with Bassam. To follow on to what he said, I’m very concerned that this would become yet another place where the interface to storage would need to be discussed, and I think that’s a really bad move right now.

 

As a community, we already have at least three different regular storage meetings: the CNCF storage WG, the k8s storage-sig and CSI. You have to track at least those to maintain even a basic idea of what’s going on. And if you really want to be involved with k8s, there’s already a lot more than that to deal with. As the other orchestrators become more CSI aware, there will likely be storage meetings for each of them as well.

 

And the line between the orchestrators and the CSI moves all the time. For about a year we’ve been talking about snapshots in k8s, and in just the past week there’s been discussion about moving that into CSI itself. If we make that move it isn’t just done on paper, it has a material change on interfaces and how it needs to be implemented.

 

Adding another thing in the mix in-between makes the lines more blurry than they already are, and an already difficult problem untenable.

 

For this reason, I think the CSI needs to re-consider its “spec only” stance and provide some basic enablement as well as mechanisms that make it easy for different people to experiment in and around it instead. Each orchestrator is going to have its CSI implementation to deal with too. Please, no more! :)

 

..Garrett

Technical Director @ NetApp

https://netapp.io/

 

From: cncf-toc@... <cncf-toc@...> On Behalf Of Kitson, Clinton
Sent: Friday, March 2, 2018 12:28 PM
To: cncf-toc@...

Subject: Re: [cncf-toc] RexRay follow up

 

https://docs.google.com/presentation/d/1BthkP9OftIICEn1h9ML1F0TKXhaUxXILlcDpuz3Sjzg/edit#slide=id.g31eff88140_0_37

 

The best place to start here is to refer to the slide above. CSI by its name is an interface specification. It has always been the intent of the group to keep it CO agnostic and focused on the key primitives that will enable volume storage. CSI tackles the fragmentation problem of a single spec to implement. But there are many other aspects to creating a production grade plugins that up to this point has intentionally been avoided in the CSI spec. So for this I think a implementation framework for the storage eco-system to work together on will be important to the other aspects of our fragmentation problem. REX-Ray is this framework for CSI, abstract of storage platform, that will 1) CO centric deployment and documentation 2) a great user experience from common packaging, docs, configuration which is important to operators and trusting CSI 3) be a placeholder for proving functionality that may or may not eventually end up in the CSI spec. 

 

 

 

Clint Kitson

Technical Director for {code}

CNCF Governing Board Member

--- 

email: Clinton.Kitson@...

mobile: "+1 424 645 4116"

team: theCodeTeam.com

twitter: "@clintkitson"

github: github.com/clintkitson

---

From: cncf-toc@... [cncf-toc@...] on behalf of Gou Rao [grao@...]
Sent: Thursday, March 1, 2018 10:21 AM
To: cncf-toc@...
Cc: CNCF TOC
Subject: Re: [cncf-toc] RexRay follow up

I think Clint should chime in here, but I had seen RexRay as more than just a CSI implementation... as a multi platform storage orchestrator?  Maybe some clarification on what that means could help, but for example, with Mesosphere, we use frameworks for complex stateful applications (take Cassandra for example).  Would RexRay help orchestrate storage provisioning (via CSI) to a framework like that?

 

On Thu, Mar 1, 2018 at 9:59 AM, alexis richardson <alexis@...> wrote:

Yes I think so.  But really I am a storage idiot. Who else could we ask?

 

On Thu, 1 Mar 2018, 17:53 Bassam Tabbara, <bassam@...> wrote:

I’m glad to see the strong alignment between Rex-Ray and CSI.

 

Would it make sense for Rex-Ray to be even more closely aligned with CSI, i.e. as a set of libraries and tools for people wanting to build CSI implementations. For example, CSI has a placeholder repo (see https://github.com/container-storage-interface/libraries) for libraries and tools. Similar to the Kubernetes incubator repo. Could RexRay become part of that or does it need to be its own top-level project?

 

I worry about confusing developers and end-users with another CNCF project that attempt to achieve the same goal — CSI.

On Mar 1, 2018, at 8:48 AM, alexis richardson <alexis@...> wrote:

 

All - questions?

(thanks Clint, this is super helpful)


On Tue, Feb 20, 2018 at 6:00 PM, Kitson, Clinton
<clinton.kitson@...> wrote:

Correct Brian, REX-Ray should be transparent to end users in this space and
provides an important service by helping connect apps to storage. Operators
of clusters are the ones that should be very aware of it as it would provide
trusted and more quality plugins that are built on top of the existing CSI
spec.

REX-Ray stats: Recently REX-Ray went through some refactoring to accommodate
the CSI architecture changes that needed to take place. This meant rolling
in the libStorage functionality which unfortunately skews the numbers a bit.
The {code} team has been primary maintainers on the framework where
collaborators have mainly focused on building drivers. Other storage
companies who understand the complexity involved in building a solid CSI
implementation see the value and commonality that can be addressed by
REX-Ray and are interested in collaborating if supported via a foundation.

Production users: Yes, REX-Ray is being used in production by some of the
users listed in the slides. Up to this point, usage levels have been tied
closely to production deployment of Mesos & Docker.

Sandbox: I believe the numbers and history justify incubation, but we can
discuss it.

Control plane: REX-Ray used to have its own control plane (libStorage API)
prior to CSI. In most recent we have made architectural changes to be adhere
to CSI.  When libStorage was its control-plane, there was integration work
performed to make libStorage a volume plugin and additionally to Cloud
Foundry. Today, anyone who implements CSI on the cluster orchestrator side
can talk with any REX-Ray plugin.

Data plane: REX-Ray is not involved in the data-plane of storage operations.
It is an orchestrator and simply gets two components (local/remote storage &
an OS) connected. It essentially performs the exact same steps that someone
would manually perform to get these two components communicating and the
reverse on tear down.

Persistent state: It is completely stateless today.




Clint Kitson
Technical Director for {code}
CNCF Governing Board Member
---
email: Clinton.Kitson@...
mobile: "+1 424 645 4116"
team: theCodeTeam.com
twitter: "@clintkitson"
github: github.com/clintkitson