Hi all, I'd love to hear how folks are deploying with metaflow in production. Three primary questions: 1. What do the flow structures look like? IE do you have a training flow that pushes the model to your registry and then a scoring/inference flow that does the actual scoring separately? 2. How are you orchestrating the flows? Assuming most folks are using airflow, but curious to hear thoughts here? 3. How do you manage dependencies if you aren't using conda? We've considered creating docker containers and then using the docker operator in airflow, but would love to hear alternatives. Thanks!

hopefully other folks can chime in too but I can start with some common answers: 1. A separate training / scoring flow is a common pattern, since it allows you to run inference and training and separate cadence (e.g. retrain daily, promote to production when certain conditions are met, while running inference with the latest model constantly). Note that you don’t need a separate model registry necessarily when using Metaflow, thanks to Metaflow’s artifact tracking.

Awesome, thank you ville!