I wouldn't say that #Knative is overly complex. I'd say that it is early. Even the team says that it is only a 0.1 version. There is a lot to like there. Let me use all the nuance of Twitter to try and put some color on it.
First, the install story is more complex than it should be. YAML files with thousands of lines is huge. Much of that is HTML and such for dashboards. We need to find a better way. This also applies to Istio -- the install has config maps with quoted yaml that has quoted yaml.
Taking a hard dependency on Istio brings a lot of value but adds a lot to the complexity of the system. Installing Istio is complex especially depending on things like mutable admission controllers and auto sidecar injection.
The dependency on Istio means that Knative can't take advantage of the ecosystem that has built up around Ingress. This includes things like cert-manager or external DNS. And for many of the use cases for Knative that n/s traffic stuff is *critical*.
I really really like the break down of the project into decomposed parts including serving, building and events. That hints at a "voltron" where the independent parts are useful on their own but they come together to make something super powerful.
This idea of Voltron systems has worked well for Kubernetes. Those independent parts can then be put together in new ways that we never foresaw when we started. It is a super powerful ecosystem foundation. The ecosystem around Ingress is a great example.
(Ingress itself is pretty flawed to be sure. But I'd rather see the #Knative stuff work with the community to evolve/improve that vs. throwing all of that out. We can talk about how we are looking at evolving Ingress with @HeptioContour for example.)
@HeptioContour But I feel like, because it is still early, Knative doesn't quite nail the Voltron approach yet. It needs to evolve on two fronts. First, the parts are still too coupled both at the tech level (I think -- still learning) and at the install level.
@HeptioContour The second place where it doesn't meet the voltron test is that it is still too DIY when things are put together. The minimal install includes YAML that is deeply nested. See github.com/knative/docs/b…
@HeptioContour As an infrastructure engineer I like this as you can see what is happening under the covers, but it doesn't meet the bar yet to make it feel as easy breezy as people want serverless (whatever that is) to be.
@HeptioContour As an aside, serverless isn't necessarily less complex. It is often moving the pieces around on the plate and finding a different sweet spot. For example look at Lambda + IAM policy + CloudFormation.
@HeptioContour Looking into the crystal ball I think that there is room to complete the voltron but it'll take time and focus. And they'll need to be disciplined about both making sure the parts are independently useful *and* add come together into something awesome. That's hard.
@HeptioContour So -- in summary -- I really like it a lot. Inside of Google I pushed *hard* to build a decomposed AppEngine on top of GCE (at first) and then Kubernetes. That is what this is. So I have a hard time complaining, honestly. I'm just anxious for it to get where it is going /thread
@HeptioContour One last tweet -- if you want to see me drone on for 2 hours playing with it check out j.hept.io/tgik044. Watch on 1.5x speed. I really just scratch the surface and hope to do more #TGIK8s on it in the future.
