I've run 2 Mesos stacks in production and have experience setting up a k8s stack (on prem). First off in my experience k8s ops is way more complex that the DC/OS stack. I recently setup a new DC/OS deployment (80% of the cluster resources was Spark, which works natively with Mesos and I'd rather run the ancillary services on Marathon, then spend another 80% of my time on k8s). If I didn't have the Spark requirement I would have went k8s.
Despite going with mesos I really had to contend with the fact that k8s just has way, way more developer support - there are so many rich applications in the k8s sphere. Meanwhile I can probably name all the well supported Mesos frameworks offhand. Next, marathon "feels" dead. They recently killed their UI interface as I imagine that they are having trouble giving resources to marathon. 3 years ago I wanted a reverse proxy solution that integrated with mesos as well as non-mesos services so I hacked Caddy to make that work [1]. 3 years later, I was looking for a similar solution and found traefik. It claimed to work with mesos/marathon, but the marathon integration was broken and the mesos integration required polling even though mesos had an events API, so I hacked traefik to make that work [2]. On the other side of the fence, you have companies like Buoyant who rewrote a core piece of their tech (Linkerd) just to support K8s (and only K8s). This has a compounding effect, where over the years things will just become more reliant on assuming you are running k8s.
That "cost" you pay to setup Mesos/k8s is usually a one time cost on the order of a month. I feel however, that k8s is going to give you a better ROI (Unless, you are managing 100s of nodes with Spark/YARN/HDFS, then Mesos continues to be the clear winner).
[1] https://github.com/mholt/caddy/pull/40
[1] https://github.com/containous/traefik/pull/2617
