Amazon SageMaker constricts the use of their own mxnet package and does not offer a strong Kubernetes backbone. At the same time, Kubeflow is still quite buggy and cumbersome to use. Which tool is a better pick for MLOps pipelines (both from the perspective of scalability and depth)?
Depends. I think two factors should drive your decision.
Is your core value proposition in the area of ML? If yes, you'll want to customize training, inference and orchestration. You'll hit the golden Cage of SageMaker fairly quickly. If it isn't you are probably ok with the reasonable - albeit limiting - defaults of SageMaker.
Secondly, is your organization invested in Kubernetes and Open Source? Are you single cloud? If so, how strongly committed are you to AWS?
We used SageMaker for 6 months, then pivoted completely to Kubernetes and KubeFlow. The orchestration layer in KF Pipelines is great, since it allows self service for data scientists (python and fairly simple api) and well Depends. I think two factors should drive your decision.
Is your core value proposition in the area of ML? If yes, you'll want to customize training, inference and orchestration. You'll hit the golden Cage of SageMaker fairly quickly. If it isn't you are probably ok with the reasonable - albeit limiting - defaults of SageMaker.
Secondly, is your organization invested in Kubernetes and Open Source? Are you single cloud? If so, how strongly committed are you to AWS?
We used SageMaker for 6 months, then pivoted completely to Kubernetes and KubeFlow. The orchestration layer in KF Pipelines is great, since it allows self service for data scientists (python and fairly simple api) and well for Ops, since everything is based on reproducible, containerized steps.
