My company is running it in production. We started with the monitoring stack and then our application, which is a SaaS product. The product itself is a stack deployed on a per customer basis. With rough configuration management using chef and terraform it took an ops person hours to complete a deployment a year ago. Today, it takes a minute and it's self-service for engineers.

Efficiency is increased a great deal. Previously, I'd have idle instances costing money when they were unused. Now my resource consumption is amortized across my entire infrastructure.

It also drives an architecture that lends itself to better availability. A pod in k8s may be rescheduled at any moment. If it's a stateless service, one may just increase the number of replicas to prevent a service interruption. If it's a stateful service, one is forced to think about how to persist data and gracefully resume operations.

It took a while for everyone on my team to get familiar with it, but once we started to really grok it, the sense of safety went way up. I feel pretty good about the fact that if someone accidentally blew away a cluster with thousands of pods in it, that I could easily replicate the entire thing and get it back up and running in tens of minutes without a lot of hassle.