default resource overhead for each individual profile

Question: Not sure whether this question has been discussed before, but I am curious about the resource overhead for each individual profile. In the current configuration ( in v1.2, v1.3), when creating one single profile, there are two running pods are created by default:

ml-pipeline-ui-artifact-c8fbcb8f9-k88bz
ml-pipeline-visualizationserver-6b78c9646-s7q4c

each one takes additional istio sidecar container (cost 100m cpu/ 128Mi memory). so in total cpu requested 100m * 2 + 128m * 2 = 0.5 CPU unit, memory requested 128 *2 + 128 * 2 = 512Mi. (even the profile has been created but no further actions were conducted.)

so let say today we have to create 100 profiles (one for each individual participant) but only 20% of them are heavy users (80-20 rule), and our system wasted 100 * 80% * 0.5 CPU unit (and 512Mi + 100 * 80% memory).

This seems to be more like a resource leaking, not the design concept. any thoughts? or documents? thanks

---

transferred from https://github.com/kubeflow/kubeflow/issues/5832

[zijianjoy]

Hello @hsinhoyeh , you can disable the pipeline pods by setting the following label from your profile’s namespace labels: pipelines.kubeflow.org/enabled: false.

You can do this for all the profiles that don’t require Kubeflow Pipelines. For future roadmap, we plan to remove visualization server too.

Does this solve your issue?

[stale[bot]]

This issue has been automatically closed because it has not had recent activity. Please comment “/reopen” to reopen it.