Hi everyone! I am completely new to Flyte but starting to set up some of our ML pipelines. We have a usecase where we might want to build some container based upon the output of a workflow. Does this make sens, how can it be done? Is this bad practice. Thanks :)

Hi, @ambitious-postman-56732! Welcome to the community. Could you elaborate more on your use case? Why do you want to build containers?

Thanks 🙂 Currently(and we thought about redesigning this) we are building the model in to a container before deployment. In order to automate our manual pipeline today we would like to allow flyte to build this container if the new model is proven to be better than the current one used in prod. This would allow us to reduce the amount of manual steps. However this is an artifact of how we update the models that we serve using [Triton](https://github.com/triton-inference-server/server).

Have you looked at https://docs.flyte.org/projects/cookbook/en/latest/auto/core/containerization/raw_container.html? I’m not entirely sure if this would solve your problem, but I do think this is possible. Lemme know if this would work.

My recommendation is to use Flyte as the model store and make your container download the new model

@freezing-airport-6809 I dont fully follow what this means? As I understand flyte dont have support for model storage?

Flyte has an implicit link to the model generated- take a look at Flyte remote. Cc @tall-lock-23197 can you help

Using FlyteRemote, you can retrieve the inputs and outputs of an execution. See https://docs.flyte.org/projects/flytekit/en/latest/design/control_plane.html#retrieving-inspecting-executions for reference.