Feast.
Feast is an open-source feature store providing a shared definition of features, an offline store for training data and an online store for low-latency serving, so both paths compute features identically.
Where Feast fits on a long engagement.
Training-serving skew is the defect a feature store exists to prevent. When features are computed by a training script and reimplemented for the serving path, the two drift, and the model in production receives inputs subtly different from what it learned on. The result is a model that scored well and performs poorly, with nothing in the logs to explain it.
Point-in-time correctness is the other essential property. Training data must reflect feature values as they were at the moment of the event, not as they are now; joining current values into historical training rows leaks future information and produces an offline score that production will never reproduce.
What an assigned team does with Feast.
A feature store is infrastructure with real operational weight — an online store to keep available, materialisation jobs to run, and consistency to maintain. It is justified when several models share features or when skew has already caused a problem, and premature otherwise.
Judging that threshold honestly rather than adopting it by default is part of the architectural work under ai engineering services.
What we use Feast for.
- One feature definition for both paths Training and serving computing the same thing, which removes skew by construction.
- Point-in-time correct training sets Historical values as they were, so offline scores are achievable in production.
- Features reused across models Shared definitions rather than each project reimplementing the same computation.
How Feast capacity is assigned.
Feature infrastructure is assigned inside AI capacity, adopted when sharing or skew justifies it rather than as a default.
Tell us what your roadmap needs Feast for.
A service delivery manager replies with the disciplines we would assign, the monthly capacity and what the first month looks like.
Loading the contact form… You can also email hello@azendo.co.
We reply within one working day. No obligation, and no newsletter.