Should have been tolerations: [] , I'll send a PR soon to fix it.
In the meantime you can solve it by setting the value to k8sagent.podTemplate.tolerations: []
I’ll try and remove the newline for tolerations and nodeSelector
Was a mistake on my end, added an extra ] by accident
No, I see that within the k8s-agent pod when it tries to execute the experiment
❯ k get pod -w NAME READY STATUS RESTARTS AGE clearml-agent-group-cpu-agent-6d9cd7cf9f-hq2dl 1/1 Running 2 2d19h clearml-apiserver-7bbcb75496-64lm7 1/1 Running 6 2d19h clearml-elastic-master-0 1/1 Running 2 2d6h clearml-fileserver-68db5b6dd6-fkv4q 1/1 Running 2 2d19h clearml-id-f7cd2dc3584f443c9b7ae895b03e900c 0/1 ContainerCreating 0 9s clearml-k8sagent-7f584456c5-wf6wd 1/1 Running 0 3m26s clearml-mongodb-7698fc9f84-xbfhf 1/1 Running 2 2d19h clearml-redis-master-0 1/1 Running 2 2d6h clearml-webserver-55bdc98c74-ghpv4 1/1 Running 3 2d19h
I can verify that the pod is not removed, and a new one is not created when an experiment is reset and enqueued
I'll give it another try next week and keep you posted
So you see the issue - it's with the k8s glue pod spec?
The configmap shows this❯ k get configmaps k8sagent-pod-template -oyaml apiVersion: v1 data: template.yaml: "apiVersion: v1\nmetadata:\n namespace: \nspec:\n containers:\n \ - resources:\n {}\n env: \n - name: CLEARML_API_HOST\n value: \" \"\n - name: CLEARML_WEB_HOST\n value: \" \"\n - name: CLEARML_FILES_HOST\n value: \" \"\n \ - name: CLEARML_API_ACCESS_KEY\n valueFrom:\n secretKeyRef:\n \ name: clearml-conf\n key: apiserver_key\n - name: CLEARML_API_SECRET_KEY\n \ valueFrom:\n secretKeyRef:\n name: clearml-conf\n key: apiserver_secret\n tolerations:\n []\n nodeSelector:\n {}\n" kind: ConfigMap metadata: annotations: meta.helm.sh/release-name: clearml meta.helm.sh/release-namespace: clearml creationTimestamp: "2022-02-02T10:25:25Z" labels: app.kubernetes.io/managed-by: Helm name: k8sagent-pod-template namespace: clearml resourceVersion: "4241060" uid: aec0e958-8ce9-4dfc-bd88-11a8b78bfdc1
it has a partial pod template mapped to templates/template.yaml
As you'll probably run into issues as soon you want to start running experiments from private repos
Another possibile issue I encountered is when an experiment completes, it's pod is kept in the Complete phase, and when I Reset and Enqueue the experiment again, no new pod is created, the existing one it being updated but will never actually run the experiment again
now waiting for the newer pod to start
Ah I see it! I made a mistake in the helm chart 🙈
Thank you very much! CostlyFox64 SuccessfulKoala55
No problem! Thank you for finding a bug in the chart 🤓
I have some other improvements to the k8sagent I want to submit a PR for soon, so be sure the monitor the chart repo for updates!
SmugHippopotamus96 that's strange - the pod should be removed
SmugHippopotamus96 the new version of the helm chart should fix all the issues you mentioned!
SmugHippopotamus96 how did this setup work for you? are you using an autoscaling node group for the jobs?
with or without GPU?
Any additional tips on usage?
and the spec.tolerations field there is defined as a map where it should be a slice
and the k8s agent is configured to listen on that queue (see above)
Because if not, the k8sagent pod is still using the old version