BentoML

Yatai and Kubeflow are not strictly speaking either-or. Yatai can work with Kubeflow and serve as the deployment platform for Kubeflow. Interested to see the community’s response on KServe vs BentoML.

I wrote about how we're using Kubeflow and BentoML <https://medium.com/dkatalis/automated-model-deployment-with-bentoml-and-kubeflow-3ac5bc0d1b3d|https://medium.com/dkatalis/automated-model-deployment-with-bentoml-and-kubeflow-3ac5bc0d1b3d>

Great article <@U03NQ3FCY6A> and thanks for the clarification <@U01GMPKSG0H>.

My current decision then would be between:
• Kubeflow (distributed training) + Yatai + BentoML
• Kubeflow + Kserve
    ◦ (or also adding BentoML here..?)
• Databricks serverless inference (just released)
Also, is adding MLflow on top still the default go-to?

tbh I found Bento/Yatai way easier than KServe

one major pain point (I think this should still be the case) about KServe for me is how it forces you to authenticate via a session cookie which requires you to login to Kubeflow programmatically. <https://github.com/kserve/kserve/blob/master/docs/samples/istio-dex/README.md>

It's more than that though. Bento lets you organize your model(s) into separate Runners etc. A lot of the benefits you're getting from KServe you probably already get from Bento