Interesting set of suggestions! Here are a few comments:

  *   Mesos feels simple to deploy (only very few components: zookeeper, 
masters and agents), customization is done mostly through configuration files. 
I don't think there is a strong need to make it easier (even though I've used 
Mesos for years, so I'm pretty used to the difficulty if any)
  *   Having to manage Zookeeper adds some complexity but since Zookeeper piece 
is required to operate Marathon (which is our main framework), I don't see much 
value in the investment required to get rid of this dependency.
  *   Taking advantage of NUMA topology by default would be a good addition 
although I don't see it as strategic (at least we have solved this on our 
clusters with custom modules)
  *   I would love to see improvement on masters scalability for large clusters 
(our largest cluster is 3500 nodes and may start to suffer from the actor model)

Something that I see as a very significant drawback to the ecosystem at large 
is the difficulty to write frameworks. In addition to this, most open-source 
frameworks feel abandoned. Without good frameworks, Mesos value really 
decreases a lot (although it is very technically strong).
I think, making Mesos thrive would necessarily go through a solution to this 

Something that I'd see as strategic would be the ability to deploy complex 
workloads on Mesos without having to write a new framework. Random idea: make 
Mesos really usable as a backend for Kubernetes (as a virtual kubelet). This 
would remove a lot of barriers to use Mesos as a strong engine to operate a 
fleet of servers while allowing to use the Kubernetes API that apparently 
everybody loves.

What do you think?

Grégoire Seux

Reply via email to