Right now, a combination of Puppet and a simple custom-built tool written in Ruby. The tool inspects the metadata from Puppet about which nodes should run what apps, and the tool uses SSH connections to clone from Git, write config files, restart daemons, etc.
At the same time, developers run a local Vagrant VM that is configured identically to the production/staging clusters, and can run every single app (controlled using the same deployment tool). We also made it simple to mount an app you're working on, with code hotloading automatically taken care of. If you want to work on a microservice or an app, you just modify the code on your local machine and reload the page (or do an API request).
It works really well. It's not perfect, though. Since we're deploying directly to the servers, every node has to install packages, compile modules, package assets, etc., and everything has to wait on the slowest node; we'd like to transition to a system where we build a "release tarball" (or even a Docker image) once, and then push it to the cluster. Also, we're looking into using something like Mesos to better orchestrate apps and move away from role-based nodes; Puppet is way too rigid and too "opsy"; we'd like to use Puppet for controlling the base OS, and let the microservice/apps world be more dynamic.
Ansible doesn't give us any benefits over our current system. (Salt would be a better choice. Not a fan of YAML templating, to be honest, nor the tight integration with Python.)
Configuration management systems like Ansible and Salt and Puppet are fairly rigid; they are basically modular recipes for placing files and starting services on remote servers. What you get from Mesos (combined with its frameworks) is something that can run your services, keep them running, and handle the several state transitions that affect any running application: Transitioning from one version to another, for example, or restarting a failing app.
We've tried using supervisord as a stopgap solution, but it, frankly, sucks. Doesn't follow forks (no cgroups), isn't capable of cleaning reloading its config, can't do master/worker replacement, doesn't support syslog in any meaningful way.
I don't know Kong very well, but having look at the documentation, feels a little too much like a framework to me.
The nice thing about Mesos etc. is that it makes almost no assumptions about what your system is. You pick and mix the functionality you need. For example, Marathon is basically just a process manager. You wire it up the way you want; the application can be anything, from a web server to a one-off script.
Kong seems to be "Rails for microservices". Which is probably useful to someone, of course.
