Delivering inputs at prediction time.
WHAT THE PATTERNS ARE
- Batch prediction: scoring everything on a schedule, storing results
- Online prediction: computing on request
- Streaming prediction: scoring events as they arrive
WHAT BATCH SUITS
Predictions that do not need to reflect the last moment: churn risk, segmentation, recommendations refreshed daily.
WHY IT IS PREFERABLE WHERE IT FITS
No latency requirement, no serving infrastructure, easy to reprocess.
WHAT ONLINE PREDICTION REQUIRES
Features available within the latency budget A serving store, not the warehouse Fallback when features are missing
WHY NOT THE WAREHOUSE
It is not built for many small low-latency queries.
WHAT TO PRECOMPUTE
Anything not depending on the request itself.
WHAT TO COMPUTE AT REQUEST TIME
Only what depends on the request.
WHAT TO ALWAYS HANDLE
A feature being unavailable.
WHAT THE OPTIONS ARE
A default value, recorded as such Refusing to predict Falling back to a simpler model
WHAT TO NEVER DO
Substitute silently and proceed.
WHAT TO LOG
Every prediction, with the features used.
WHY
It is the only way to investigate a wrong prediction later.
