Feature Stores Are Optional. Feedback Loops Are Not.

Published 2025-08-07 by Faraz Rahimi

You can wait on fancy ML infrastructure. You cannot wait on a way for the floor to tell you the model was wrong.

Feature Stores Are Optional. Feedback Loops Are Not.

Tags: Machine Learning, MLOps, Computer Vision, Operations, unlisted

I would rather have a CSV of rejects and a weekly train job than a cathedral of features with no users talking back.

The loop is: show a guess, accept a no, store the pair, retrain or adjust rules, ship. If any arrow is missing, you have a demo.

Infrastructure follows volume

When the CSV hurts, then a table. When the table hurts, then more. Buying the end-state first is how you spend a quarter without a reject button.

People are the loop

If operators learned to ignore the tool, your feature store is a museum. Fix the UX before you scale the pipes.

Original post: https://farazrahimi.com/posts/feature-stores-are-optional-feedback-loops-are-not

Feature Stores Are Optional. Feedback Loops Are Not.

Published 2025-08-07 by Faraz Rahimi

You can wait on fancy ML infrastructure. You cannot wait on a way for the floor to tell you the model was wrong.

Feature Stores Are Optional. Feedback Loops Are Not.

Tags: Machine Learning, MLOps, Computer Vision, Operations, unlisted

I would rather have a CSV of rejects and a weekly train job than a cathedral of features with no users talking back.

The loop is: show a guess, accept a no, store the pair, retrain or adjust rules, ship. If any arrow is missing, you have a demo.

Infrastructure follows volume

When the CSV hurts, then a table. When the table hurts, then more. Buying the end-state first is how you spend a quarter without a reject button.

People are the loop

If operators learned to ignore the tool, your feature store is a museum. Fix the UX before you scale the pipes.

Original post: https://farazrahimi.com/posts/feature-stores-are-optional-feedback-loops-are-not