引導式測驗 · 中等 等級
模型端點負載測試測驗
使用到達模式、代表性輸入、延遲百分位數、資源測量和安全測試控制來規劃實際的端點負載測試。
問題 1 的 8
Which metric set gives a broader view of model endpoint capacity than average latency alone?
預覽每個問題8 問題
當您參加測驗時,就會出現答案和解釋。
- Which metric set gives a broader view of model endpoint capacity than average latency alone?
- Training accuracy and validation loss
- Mean request size and model parameter count
- Throughput, latency percentiles, errors, and resource use
- Source-file count and deployment age
- How can a closed-loop virtual-user test hide overload?
- It creates an unlimited arrival rate regardless of response time
- Waiting users reduce their request rate as responses slow
- It records only successful requests
- It applies a fixed arrival rate independent of user behavior
- Why include several payload sizes in a model endpoint test?
- Payload length determines the endpoint license
- All tensor shapes require identical compute and memory
- Large payloads cannot be tested safely
- Input dimensions or sequence lengths can change compute and memory
- What distinguishes a soak test from a brief load test?
- It measures only startup latency before any request
- It repeats one request until its response is cached
- It runs long enough to reveal gradual degradation such as leaks
- It varies model architecture while keeping duration brief
- Why test cold and warm serving paths separately?
- Cold requests always have lower latency than warm ones
- Warm requests bypass the model and use a different API
- Startup and model loading can add latency beyond steady inference
- Cold and warm requests differ only in labels
- What can make a load generator provide misleading results?
- Recording latency makes the generator invalid
- The generator can saturate on CPU, network, or serialization before reaching target load
- Python scenarios cannot generate concurrent requests
- A load generator cannot affect measured throughput
- Why define stop conditions and monitor dependent services during a stress test?
- Stop conditions guarantee that the endpoint cannot fail
- Monitoring dependencies removes them from the test
- Stress traffic should be bounded to avoid affecting other systems
- A stress test should disable all infrastructure metrics
- What does Locust provide for a load test?
- Python-defined scenarios with throughput, response-time, and error reporting
- An automatic proof that production capacity will be sufficient
- A pretrained model that serves predictions
- A guarantee that the generator has unlimited capacity