Engineering, production, governance, cost, communication · insight-0286
Use batch scoring when decisions are made on a schedule
- Evidence type
- practitioner heuristic
- Difficulty
- easy
- Roles
- ML engineer, Data scientist
- Industries
- telecom, retail, banking
Summary: situation, action, reason, boundary
Situation
A team plans a low-latency online endpoint for a churn model whose scores are used in a weekly retention campaign.
Action
Match serving to the decision cadence: precompute scores in batch when decisions are scheduled, and build online inference only when the decision needs information that arrives at request time.
Reason
Batch scoring is simpler to operate, easier to audit and cheaper when freshness requirements are measured in hours or days.
Boundary
Campaign targeting, daily pricing, planning forecasts, risk reviews.
Symptom or risk
Online serving adds latency budgets, availability targets and on-call duty for a decision that happens once a week.
Exceptions
Fraud checks, search ranking and real-time pricing need request-time features and online inference.
Trade-offs
Batch scores go stale between runs; document their age at use.
Pitfalls
Building online serving for the demo and then running it unchanged in production.