The limitation of only being able to run Containers I think will be fleeting — as docker and alternatives mature, it really won't make sense to ever use anything else when trying to get the scale that Kubernetes and Mesos is going for by abstracting out the underlying hardware and providing a framework to run sophisticated apps on undifferentiated hardware.
Disclosure: I work on Kubernetes at Google (and wrote the doc you linked to).
From my point of view, the benefits of the two level scheduling are actually quite limited with respect to how the whole story is usually told. Some Mesos framework always use all the resources from the clusters and it might get tricky to really have multiple frameworks to run at the same time on Mesos. Also, sometimes those frameworks don't really offer so many additional features to justify changing the way you are already using Spark, Cassandra and so on.
[0]https://github.com/coreos/prometheus-operator [1]https://github.com/coreos/etcd-operator [2]https://github.com/huawei-cloudfederation/redis-operator [3]https://github.com/CrunchyData/postgres-operator [4]https://github.com/krallistic/kafka-operator [5]https://github.com/pegerto/cassandra-operator
So for the developers responsible for the uptime it could matter a lot if they are more comfortable with one of codebase or language they use.
Off course if you are more like a consumer app type of startup, just put all energy on UX and stuff and hope for the best. But lets hope every bank that uses this sort of stuff has a developer on call that can actually dive into the codebase.
Edit: I apologize to the readers, this is not the right place for this kind of talk
It is true that Mesos works on a different level, but, most of all, the two level scheduling is just a different take at the problem of abstracting physical/virtual resources. In the end, both Mesos/Marathon and Kubernetes aim at the same goal: allow developers to stop thinking about servers.
Kubernetes' great advantages is the community (which is unbelievable) and the extensibility it proposes: Third Party Resource or Custom Resource Definition, pluggable webhooks in the API Server and a number of other things that are simply not there in Marathon or any competitor which allow companies to make Kubernetes work best for their use cases.