only allow one Metacontroller instance to be active at a time

[crimsonfaith91]

Addressing issue #5 partially. In the long run, we still have to resolve leader election issue when the user spawns more than one Metacontroller instance.

Metacontroller uses apps/v1beta1 Deployment under the hood. This PR changes the Metacontroller to use apps/v1beta2 StatefulSet to maintain the invariant of only one Metacontroller instance is active at a time.

[crimsonfaith91]

@enisoc Do we want to use StatefulSet for hooks too? For example, do we need to change this?

[enisoc]

Let's use apps/v1beta2 for now so we maintain compatibility with k8s 1.8.

[crimsonfaith91]

noted!

[crimsonfaith91]

@enisoc When should we change the version from apps/v1beta2 to apps/v1? When the v1.8 is not supported anymore?

[enisoc]

I'd like to keep support for 1.8 until we have a compelling use case for relying on a feature that only exists in 1.9+.

[enisoc]

Do we want to use StatefulSet for hooks too?

For hooks, we don't need to guarantee that there isn't more than one, but we do ideally want to ensure that all replicas are the same version at a given moment. One way to do that would be to set hook Deployments to use the Recreate update strategy, but that causes unnecessary downtime. Ironically, the ideal way to update a hook deployment would probably be to use a blue-green strategy with a Service that atomically switches to a fully-deployed set of replicas.

You can file an issue to track this if you like, but I don't think it's a high-priority problem for hooks. There may be some "flip-flopping" during rolling updates of hook code, but it should converge on its own once the rollout completes. Most hooks will probably work fine with just one replica, so they should update quickly.

[crimsonfaith91]

@enisoc +1 on using blue-green strategy with the Service to ensure all hook replicas have the same version at one moment, which may become important when there are quite a number of hooks.

As the hook codes converge eventually (i.e., not a high priority issue) and most hooks work fine with 1 replica, I will not file an issue tracking this.

[enisoc]

@crimsonfaith91 Since we don't have automatic e2e tests yet, this is a good chance to make sure you can run the tests manually. Once you have a cluster set up, I recommend using the Skaffold flow to run your dev build. Since you just changed the Deployment to a StatefulSet, it's especially important to make sure you don't have any previous version of Metacontroller running; you can delete the metacontroller namespace first to ensure that.

I tried skaffold run locally and found one problem so far:

error: error validating "STDIN": error validating data: ValidationError(StatefulSet.spec): missing required field "serviceName" in io.k8s.api.apps.v1beta2.StatefulSetSpec; if you choose to ignore these errors, turn validation off with --validate=false

In this case, I think you can set serviceName: "" since we don't need per-pod DNS entries.

Once your dev build of Metacontroller is running, go into the examples dir and run ./test.sh to manually kick off the smoke tests.

[crimsonfaith91]

@enisoc I see. I thought it suffices to get the Travis CI build passed (the build does run unit tests). Thanks for the pointer!

[enisoc]

Since args.yaml and image.yaml contain partial objects, used for kustomize overlays, we don't need to mention fields here if the value is the same as the base manifest.

[enisoc]

LGTM