Hi everyone! My name is Eike and I am an software engineer and want to integrate prefect via pygeoapi into an spatial data infrastructure running in an k8s cluster. In the end, I would like to run various processes with different infrastructure requirements (GPU, CPU, RAM, ...).

👋 @Eike - welcome! once you're ready to set up your flows to run on kubernetes, this might be a good resource for you! similarly, here's an toy e2e example of setting up a cluster and worker and running some flows against it, just for demonstration of what pieces are involved in the minimal case

I am somehow facing this problem "Is there a way to have a Prefect GPU agent/worker on (temporary) sleep when it’s not used in Kubernetes?"

in general, folks usually keep their worker alive at all times, though for other infrastructures (cloud run, ECS etc) we have a concept of push pools which do not require you to run a worker.

Okay, push pools sound like the required concept.

correct.

Welcome Eike! Great to have you here 👋 ✨

So, it is not possible to have a work-pool with 0 workers and the prefect server triggers the start of an worker if any flow in the pool should be executed!? This should include the downscaling to 0 workers if not jobs/flows are queued for the according pool.

you may be interested to read the original (somewhat old at this point) ideas behind the hybrid model, but yea you need something in your client-side / runtime infrastructure asking the prefect server if there's work to spin up in the kubernetes case, its a pod that asks the server if it should spin up another pod for a flow run (so it scales down to 1)

just to clarify, in the prefect + k8s scenario you're likely thinking of, "workers" are not the things that run your flows or need to scale in any way. They're essentially a messenger between the Prefect server and your flow run jobs and pods, so a worker can be a relatively small pod, but it should be on all the time. Flow run pods that need things like GPUs will still scale up and down the way you're expecting

I am slowly achieving this understanding. But it is unclear to may how this interacts with k8s autoscaler, k8s jobs and stuff. How does this match with the work-pool per infrastructure type idea?