The human-centric platform for production ML & AI

Outerbounds

Does the K8s deployment type get the most love and support from Outerbounds? Is it the recommended way to self host and the place you’d like contributions in?

I ask because I’m thinking of finally learning K8s--literally because I know Metaflow can be deployed onto it.

I once tried to port the terraform scripts to AWS CDK for my custom use case (using a mix of on-prem compute with ECS on-prem and AWS) but failed.

So the question is: what’s the focus of Metaflow?

Or is the AWS-native batch-based approach equally frequently enhanced?

the stance is the same as before, <https://docs.metaflow.org/getting-started/infrastructure#example-stacks|as stated here>
• when it comes to simple self-service options `@batch` + `step-functions` is <https://docs.metaflow.org/getting-started/infrastructure#low-maintenance-full-stack-powered-by-aws|a good place to start>. As with any higher-level AWS service, they can work great as long as you stay within the (tight) constraints set by the services.
• since it's impossible to improve AWS services (like the many limits of `@batch`) from outside, <https://docs.metaflow.org/getting-started/infrastructure#customizable-full-stack-on-aws-powered-by-kubernetes|a Kubernetes-native stack> is the way to go when you graduate from the starter stack
Both of these stacks see regular improvements.

If you don't want to spend many engineering-years optimizing and configuring Kubernetes (come and say hi to us at Kubecon Paris booth J26 today! :meow_wave: ), <https://outerbounds.com/get-started/|just deploy Outerbounds>.