Helm Chart Volume Configuration "ReadWriteOnce" Doesn't Work with Multiple Pods

[moghtader]

My actions before raising this issue

• Read/searched the docs
• Searched past issues

When using the Helm Chart, the storage.yml ReadWriteOnce option under the backend folder structure doesn't work in my K8s environment. If you have deployments across multiple nodes, ReadWriteOnce will form a race condition causing only one backend pod to start successfully. ReadWriteOnce by definition only allows one node to mount - so how is this supposed to work?

Steps to Reproduce (for bugs)

1. Attempt to deploy the helm chart to Kubernetes with the default storage

Expected Behaviour

1. The defaults for the helm chart should be able to launch CVAT in K8s with the minimal amount of modification.

Current Behaviour

Possible Solution

1. Changing storage.yml to ReadWriteMany
2. Put all backend deployments into the same node (maybe there's a reason this isn't designed this way though)

[westernspion]

Do we actually need these pods to all share storage? Or can they each be granted individual PVCs for each instance of the server/utils/workers pods? Sharing a cache a la volumes seems like a poor design choice so I question if this is just an assumption made in the docker-compose configuration and just found it's way into the helm chart.

[Zanz2]

Is there a reason that ReadWriteOnce is hardcoded? Im not a cloud developer so im just curious if switching it to ReadWriteMany has negative side effects

[moghtader]

I'm unsure why it is needed that all share the same claim. Maybe someone more familiar with the chart knows? It might have some unintended consequences if we remove it.