3 ms·
Stop Making K8s Happen
- gryfft 3y agoAt this point I'm feeling pretty good about my assessment that k8s' major utility has been for organizations that felt their engineering departments were too comfortable, organized and productive and needed to distract the nuts-and-bolts types so the executives can prance around and play silly games. I don't think we can go back, mind you. We're stuck with it as an abstraction layer that's never going to go away. But Jesus, containers and orchestration have caused way more problems than they ever solved, in my experience. Throwing away decades of architectural mental models had-- shock!!!!-- downsides.
- mrkeen 3y ago> Incident > Then comes Black Friday, and traffic goes up to over 300k per hour, and stays that way for 5 days. All alarms are going off, Pods becoming unhealthy and “dying”. APIs erroring out with 5xx all over the place. Containers coming up, and dying, and new containers coming up. The site experience is slowed to a crawl. All this while the devops team is frantically bringing up new pods. You have load. It happens. There shouldn't be anything frantic about bringing up new pods. Just bump up your replica numbers in your config via git. Or, if you're doing it by hand for some reason: kubectl scale deployment --replicas=5 my-app > You're not Google. Stop making k8s happen. This negates the incident story. You don't need k8s... unless you start losing services because of load and need to react quickly and/or automatically.
- brodouevencode 3y ago> I had submitted this as a talk to a conference, apparently they did not like the message. Maybe because the conclusion was too one-sided. k8s is fine if you have engineers that can support it. This is a tale of that not happening.