Try k8s

[sebbacon]

#19 showed Google Cloud Run isn't going to be a goer.

I'm less leery than I used to be about using kubernetes, given there's plenty of k8s-as-a-service options out there.

I've tried DigitalOcean's which is priced at the same rate as their normal droplets and it seems quite nice.

The Google Cloud Kubernetes Engine is another option but I kept getting gnomic error messages when I tried it.

The process of getting a minimal deployment going is quite easy. You make a YAML configuration that describes your desired deployment (here's the one I made), then run

kubectl apply -f config.yaml

and then

kubectl expose deployment hello-world --type=LoadBalancer --name=my-service

[sebbacon]

Next steps:

• Automate with Github Actions
• Check all the k8s settings; what can we use to spin down an instance? See autoscaling for more.

[sebbacon]

Note autoscaling has a dependency on metrics-server

[sebbacon]

Autoscaling to zero is the blocker here. We want to reap processes which are no longer being used.

There are some indications this would be possible. For example, I found a recent k8s update supported zero scaling for external metrics; but then you have to set up a prometheus service (or stackdriver) to log request activity, and configure the autoscaler to inspect that service, and it's all quite hairy. Another nice alternative appeared to be OpenFaaS, which claims to support autoscaling out of the box (using various k8s bits and pieces under the hood) but the installation story for their cloud service is somewhat unpleasant.

I also considered Google App Engine Standard edition, which scales to zero for you, but this requires you to use their runtimes and we have some pretty exotic stuff like R-in-python which don't look supported. Google App Engine Flex edition will only scale to a minimum of one.

All in all, it seems like there's no convenient service for autoscaling