Taking this abstraction one more level up, I propose we keep the cluster generic, and not just for container service. More so, because the proposal is to make it core cloudstack pluggable service.
With the proposed container orchestration service, - A notion of cluster is introduced, which is basically group of Vms, network rules etc Container service is not the only use-case which will benefit from this cluster. Any service which requires a bunch of Vms grouped can make use of this cluster. Any service which require the following can be supported by a generic framework, similar to network plugin framework. - Create group of Vms (ServiceCluster) and manage their life cycle - Configure network to make the ServiceCluster functional (e.g. open required ports on all vms to for services to communicate) - Create network rules to make the control pane accessible (e.g. create PortForwarding rule to access k8s/PaaS dashboard) - Driven by cloud-config via userdata (e.g. config to download required binaries and start services on the Vms) - Support a purpose-built template (e.g. CoreOS) This framework can be used by services like Container service, PaaS or DB-as-a-service . I’ve been doing some prototypes on a similar kind of abstraction. -----Original Message----- From: Murali Reddy [mailto:murali.re...@shapeblue.com] Sent: Monday, January 30, 2017 1:24 PM To: dev@cloudstack.apache.org Subject: Re: [PROPOSAL] add native container orchestration service I agree with some good views Will has shared and I also agree to the concerns raised by Wido and Eric. IMO we need balance of staying relevant/add new features vs stability of CloudStack and take corrective action if needed. We have two good examples for both. When SDN was hot technology CloudStack ended up with bunch of SDN controller integrations. Few years later, now CloudStack is carrying baggage with no maintainers for those plugins, with probably not many of CloudStack users needing overlays. On the other hand, other than attempt to liaison with ETSI for NFV no effort was done. OpenStack has become de-facto for NFV. Now for OpenStack, arguably NFV has become larger use case than cloud it self. I concur with Will’s point that with minimal viable solution that does not change the core of CloudStack, and can keep CloudStack relevant is worth to be taken in. Will, To your question of how different is from ShapeBlue’s container service, its design, implementation and API semantics etc remain same. ShapeBlue’s container service was true drop in plug-in to CloudStack, with this proposal I am trying to re-work to make it a core CloudStack pluggable service which is part of CloudStack. Key concepts that this proposal is trying to add - add notion of ‘container cluster’ as a first class entity in CloudStack. Which is bacially collection of other CloudStack resources (like VM’s, networks, public ip, network rules etc) - life cycle operation to manage ‘container cluster’ like create, delete, start, stop, scale-up, scale-down, heal etc - orchestrate container orchestrator control plane on top of provisioned resources At-least for k8s (which is what this proposal is targeting), integration with k8s is bare minimum. There are cloud-config scripts that automatically setup k8s cluster master and node VM’s. All CloudStack is doing in passing the cloud-config to the core OS VM’s as user data. Regards Murali Reddy On 29/01/17, 8:54 AM, "Will Stevens" <williamstev...@gmail.com on behalf of wstev...@cloudops.com> wrote: >I agree that we need to be careful what we take on and own inside >CloudStack. I feel like some of the plugins or integrations which we >have been "maintaining" may serve us better to abandon, but I feel like >that is a whole discussion on its own. > >In this case, I feel like there is a minimum viable solution which puts >CloudStack in a pretty good place to enable container orchestration. >For example, one of the biggest challenges with K8S is the fact that it >is single tenant. CloudStack has good multi tenancy support and has >the ability to orchestrate the underlying infra quite well. We will >have to be very careful not to try to own too deep into the K8S world >though, in my opinion. We only want to be responsible for providing >the infra (and a way to bootstrap K8S ideally) and be able to scale the >infra, everything else should be owned by the K8S on top. That is the >way I see it anyway, but please add your input. > >I think it is a liability to try to go too deep, for the same reasons >Wido and Erik have mentioned. But I also think we need to take it >seriously because that train is moving and this may be a good >opportunity to stay relevant in a rapidly changing market. > >*Will STEVENS* >Lead Developer > ><https://goo.gl/NYZ8KK> > >On Sat, Jan 28, 2017 at 1:13 PM, Wido den Hollander <w...@widodh.nl> wrote: > >> >> > Op 27 januari 2017 om 16:08 schreef Will Stevens >> ><wstev...@cloudops.com >> >: >> > >> > >> > Hey Murali, >> > How different is this proposal than what ShapeBlue already built. >> > It >> looks >> > pretty consistent with the functionality that you guys open sourced >> > in Seville. >> > >> > I have not yet used this functionality, but I have reports that it >> > works quite well. >> > >> > I believe the premise here is to only orchestrate the VM layer and >> > basically expose a "group" of running VMs to the user. The user is >> > responsible for configuring K8S or whatever other container >> > orchestrator >> on >> > top. I saw mention of the "cloud-config" scripts in the FS, how >> > are >> those >> > exposed to the cluster? Maybe the FS can expand on that a bit? >> > >> > I believe the core feature that is being requested to be added is >> > the ability to create a group of VMs which will be kept active as a >> > group if >> at >> > all possible. ACS would be responsible for making sure that the >> > number >> of >> > VMs specified for the group are in running state and it would spin >> > up new VMs as needed in order to satisfy the group settings. In >> > general, it is understood that any application running on this >> > group would have to be fault tolerant enough to be able to >> > rediscover a new VM if one fails and >> is >> > replaced by a fresh copy. Is that fair to say? How is it expected >> > that this service discovery is done, just by VMs being present on the >> > network? >> > >> > As for some of the other people's concerns in this thread. >> > >> > - Regarding Wido's remarks. I understand that there is some added >> > complexity, but I don't feel like the scope of the addition is >> > unrealistic. I think the LXC integration was a lot farther out of >> > the scope of what ACS does then this is. This does not change the "things" >> > which ACS orchestrates, it just adds the concept of a grouping of >> > things which ACS already manages. I think this is the right >> > approach since it >> is >> > not trying to be a container orchestrator. We will never compete >> > with >> K8S, >> > for example, and we should not try, but K8S is here and the market >> > wants it. I do think we should be keeping our head up about that >> > fact because being able to provide a the underlay for K8S is very >> > valuable in the current marketplace. I see this functionality as a >> > way to enable K8S adoption on top of ACS without changing our core values. >> > >> > - Regarding Erik's remarks. The container space is moving fast, >> > but so >> is >> > the industry. If we want to remain relevant, we need to be able to >> adapt a >> > bit. I don't think this is a big shift in what we do, but it is >> > one that enables people to be able to start running with something >> > like K8S on top of their existing ACS. This is something we are >> > interested in doing and >> so >> > are our customers. If we can have a thin layer in ACS which helps >> > enable the use of K8S (or other container orchestrators) by >> > orchestrating infrastructure, as we already do, and making it >> > easier to adopt a >> container >> > orchestrator running on top of ACS, I think that gives us a nice >> > foothold in the market. I don't really feel it is fair to compare >> > containers to IPv6. IPv6 has been out forever and it has taken >> > almost a decade to get anyone to adopt it. Containers have really >> > only been here for like 2 >> years >> > and they are changing the market landscape in a very real way. >> > >> > Kind of on topic and kind of off topic. I think understanding our >> approach >> > to containers is going to be important for the ACS community as a whole. >> > If we don't offer that market anything, then we will not be >> > considered >> and >> > we will lose market share we can't afford to lose. If we try to >> > hitch >> our >> > horse to that cart too much, we will not be able to be agile enough >> > and will fail. I feel like the right approach is for us to know >> > that it is a thriving market and continue to do what we do, but to >> > extend an olive branch to that market. I think this sort of >> > implementation is the right approach because we are not trying to >> > do too much. We are simply giving >> a >> > foundation on which the next big thing in the container >> > orchestration >> world >> > can adopt without us having to compete directly in that space. I >> > think >> we >> > have to focus on what we do best, but at the same time, think about >> > what >> we >> > can do to enable that huge market of users to adopt ACS as their >> > foundation. The ability to offer VMs and containers in the same >> > data >> plane >> > is something we have the ability to do, especially with this >> > approach, >> and >> > is something that most other softwares can not do. The adoption of >> > containers by bigger organizations will be only part of their >> > workload, they will still be running VMs for the foreseeable >> > future. Being able to appeal to that market is going to be important for >> > us. >> > >> > Hopefully I don't have too many strong opinions here, but I do >> > think we need to be thinking about how we move forward in a world >> > which is >> adopting >> > containers in a very real way. >> > >> >> Understood. I just want to prevent that we add more features to >> CloudStack which are poorly maintained. Not judging Murali here at >> all, but I'd rather see CloudStack loose code then have it added. >> >> Thinking about LXC, would like to see that removed together with >> various other network plugins which I think are rarely used. >> >> Now, the idea of Murali was misunderstood by me. I think it would be >> worth more if we would improve our cloud-init support and integration >> in various tools which makes it much easier to deploy VMs running >> containers ON CloudStack. >> >> Not so much changing CloudStack code, but rather tooling around it. >> >> If we have CloudStack talking to Kubernetes we suddenly have to >> maintain that API integration. Who's going to do that. Realistically. >> >> That's my main worry in this regard. >> >> We have so much more work to do in ACS in total that I don't know if >> this is the realistic route. I talk to many people who are not >> looking at containers and are still working with VMs. >> >> There is not a single truth which is true, it really depends on who >> you ask. >> >> Wido >> >> > Cheers, >> > >> > Will >> > >> > *Will STEVENS* >> > Lead Developer >> > >> > <https://goo.gl/NYZ8KK> >> > >> > On Fri, Jan 27, 2017 at 5:38 AM, Erik Weber <terbol...@gmail.com> wrote: >> > >> > > On Fri, Jan 27, 2017 at 7:20 AM, Murali Reddy >> > > <muralimmre...@gmail.com >> > >> > > wrote: >> > > > All, >> > > > >> > > > I would like propose native functionality into CloudStack to >> > > > provide >> a >> > > container service through which users out-of-the box can use to >> > > launch container based application. Idea is to support ability to >> > > orchestrate >> the >> > > resources and automate aspects of setting up container >> > > orchestrator >> through >> > > CloudStack. Public IAAS service providers AWS with its ECS [1] >> > > and >> google >> > > with GKE [2] already provides ability container applications. >> Competitive >> > > cloud orchestration platforms already have native support for >> > > container service. Users of CloudStack both as public cloud >> > > providers and users >> with >> > > private clouds will benefit with such functionality. >> > > > >> > > > While container orchestrator of user choice can be provisioned >> > > > on >> top of >> > > CloudStack (with out CloudStack being involved) with tools like >> > > TerraForm[3], Ansible[4] etc, advantage of having native >> > > orchestration >> is >> > > giving user a nice cohesive integration. This proposal would like >> > > add a notion of first class CloudStack entity called container >> > > cluster which >> can >> > > be used to provision resources, scale up, scale down, start and >> > > stop >> the >> > > cluster of VM’s on which containerised applications can be run. >> > > For >> actual >> > > container orchestration we will still need container orchestrator >> > > like docker swarm, marathon, kubernetes, but CloudStack container >> > > service >> can >> > > automate setting up of control place automatically. >> > > > >> > > >> > > To be honest I'm torn on this one. >> > > >> > > Containers are a rapid changing thing, and while docker swam, >> > > kubernetes, rancher or whatnot is popular today, they might not >> > > be tomorrow. >> > > They might use CoreOS today, but might not tomorrow. >> > > >> > > We have a rather poor track record of staying up to date with new >> > > features/versions, and adding a feature that is so rapidly >> > > changing is, I fear, going to be hard to maintain. >> > > Want an example, look at xenserver. It is one of the most used >> > > hypervisors we support, yet it took 6 months or so for us to >> > > support the latest release. >> > > Or IPv6... >> > > >> > > I don't mean to bash at maintainers/implementers of those >> > > features, I appreciate all the work being done in every aspect, >> > > but I believe we should be realistic and realize that we have >> > > issues with keeping stuff up to date. >> > > >> > > I'd say focus on making sure other tools can do their job well >> > > against CloudStack (kops, rancher, ++), but that does not mean I >> > > will -1 the idea if anyone really wants to go through with it. >> > > >> > > -- >> > > Erik >> > > >> murali.re...@shapeblue.com www.shapeblue.com 53 Chandos Place, Covent Garden, London WC2N 4HSUK @shapeblue DISCLAIMER ========== This e-mail may contain privileged and confidential information which is the property of Accelerite, a Persistent Systems business. It is intended only for the use of the individual or entity to which it is addressed. If you are not the intended recipient, you are not authorized to read, retain, copy, print, distribute or use this message. If you have received this communication in error, please notify the sender and delete all copies of this message. Accelerite, a Persistent Systems business does not accept any liability for virus infected mails.
