Feast Online Store and Real-Time Feature Serving
Serve Feast Features in Milliseconds, Keep Them Fresh, and Wire Them Into Your FastAPI Model at Inference Time.
By Sandip Das · · 6 min read
💡
Hey — It's Sandip Das 👋
In Day 37 we defined features once and pulled point-in-time training sets from the offline store. Today is the other half: serving the latest value per entity in milliseconds.
We'll keep that online store fresh with materialization and wire it into the iris FastAPI app, so the client sends only an ID and the model gets its features from Feast.
In Day 37 we defined features once and pulled point-in-time training sets from the offline store. Today is the other half: serving the latest value per entity in milliseconds.
We'll keep that online store fresh with materialization and wire it into the iris FastAPI app, so the client sends only an ID and the model gets its features from Feast.
📚 Key Learnings
- What the online store holds, and the request path at inference time
- Online store options: SQLite, Redis, DynamoDB
- Materialization, freshness and the TTL window
- Hands-on: FastAPI + Feast for iris-mlops, verified end to end
🧠 Learn here
The online store in one request
The offline store keeps every feature value with its timestamp. The online store keeps one row per entity k Read the full article on learnxops.com. Topics: 100 Days of MLOps, MLOps, ML, GitOps