Hello, I've been using the open-source version of BentoML for a while. Recently, we encountered errors with the

bentoml containerize

command. The issue stemmed from using remote buildx without a local installation—buildx was performing a health check by testing local installation of docker. This raises a question: What is the goal of

bentoml build

and

containerize

, and why not simply use a standard Dockerfile? In particular, I am surprised by the new

PythonImage

that enable to define the Dockerfile in python. Sometimes it seems like we end up with a standard Docker image that only contains the models and the corresponding files. What does it solve compared to creating a docker image from a Dockerfile ? I feel like I am missing some of the point. Some of my guesses: Data scientists don't know how to create dockerimage. It helps handling some of the kube orchestration. It helps define a business logic that calls the various services and have it work remotely and locally. Thank you for BentoML 🙂

Hi @Arnault Chazareix - BentoML does generate a standard Dockerfile in the bento folder, which can be used directly for scenarios like this where users need more flexibility

I resulted in using docker directly to solve the issue. That's what arose the question. 🙂

not yet, currently needs a bit manual work but we are working on something like that for OSS users