Flyte enables production-grade orchestration for machine learning workflows and data processing created to accelerate local workflows to production.

Flyte

Hello,

Question, what is the recommended way to do model serving and model registry with Flyte? I understand Flyte primarily concerns itself with orchestration.

Checkout unionml- it is a thin layer on Flyte.

But you can use any serving system. Deep serving integrations are coming. But if you want to deploy to any serving framework invoke them in python 

For model registry - Flyte is actually a model registry, as you can retrieve artifacts- but many folks use mlflow too

I can't find any official 'registry' chapter in the docs though. Although mlflow sounds like a good way to go

Ya we will
Love to add more
Docs. Cc <@U01DYLVUNJE> 

This is unionml <https://www.union.ai/unionml|https://www.union.ai/unionml>

I looked at unionml but it's a bit confusing. I wonder if I will give up Flyte functionality for something more simple? I can't see the integration between both systems. Also I'm a bit worried it will lead to to much complexity and maintenance.

But I must say this is my second day looking at Flyte so I'm kind of new to the tooling

You will not give
Up anything. It's fully interoperable

hey <@U041TDD8ACC> we’re working on ways to make the `Flyte Training -&gt; UnionML Serving` user experience smoother, but the idea is that you can still create custom/complex training pipelines with Flyte, but then easily plug it into a unionml app for serving purposes.

I added some more detail to this issue: <https://github.com/unionai-oss/unionml/issues/96>

It’s a high-level proposal, but please feel free to add comments/thoughts in there!

Also, to better understand your serving use case, do you need batch prediction, predictions-via-api-call, event-based predictions?

That is something I still need to evaluate myself.

I'm trying to build an overview of tools that I will then evaluate

Which is also depended on other systems/components

I think batch prediction and event based prediction and probably api-calls tooo