Managing features centrally.
WHAT ONE IS
A system storing feature definitions and values, serving both training and prediction.
WHAT IT PROVIDES
One definition used by both paths Point-in-time correct historical retrieval Low-latency serving of current values Reuse of features across models Documentation and lineage
WHY POINT-IN-TIME RETRIEVAL MATTERS MOST
It is the capability hardest to build correctly, and the source of the worst errors.
WHAT THE TWO STORES ARE
Offline: history, for training
Online: current values, for prediction
WHAT MUST BE TRUE
They are populated from the same definition.
WHAT IT COSTS
Substantial infrastructure and operation.
WHEN IT IS WARRANTED
Several models sharing features Several teams Genuine low-latency serving requirements
WHEN IT IS NOT
One or two models, with batch prediction.
WHAT TO DO INSTEAD IN THAT CASE
Compute features in the warehouse, and materialise them for serving.
WHAT MATTERS MORE THAN THE TOOL
That training and serving use the same definition.
WHAT TO TEST
That values retrieved for training match those served at prediction.
