A fraud model scores each payment as it happens. Which endpoint fits?
Why B is the answer
A real-time online endpoint
Someone is waiting, so online scoring fits.
In detail
- Fraud must be scored during the payment.
- That needs a real-time endpoint.
- Test latency under load.
Use case
An illustrative scenario, not a real incident.
A fraud check must finish before a card payment completes. A real-time online endpoint answers within the request.
See it in different ways
flowchart TD
A{"Is a caller waiting?"} -->|Yes| B["Online endpoint"]
A -->|No| C["Batch endpoint"]
B --> D["Low-latency scoring"]
C --> E["Scores large datasets in jobs"]stateDiagram-v2 [*] --> WhenNeeded WhenNeeded --> Online: during a request WhenNeeded --> Batch: large volume, can wait Online --> Deployments: stable URL, several deployments Batch --> [*] Deployments --> [*]
flowchart TD
Q{"Who is waiting?"}
Q -->|"A live transaction"| A0["Online endpoint"]
Q -->|"Nobody, it can wait"| A1["Batch endpoint"]Key point: Decisions during a request need real-time endpoints.
Why the other options are wrong
- A. A batch endpoint that runs every night
- A nightly run is far too late.
- C. A training job with a long schedule
- A training job does not serve predictions.
- D. A compute instance used for notebooks
- A notebook instance is not a serving layer.