rohitg00--ai-engineering-from-scratch
wehub-resource-sync
79 行
2.5 KiB
JSON
79 行
2.5 KiB
JSON
{
|
|
"lesson": "22-load-testing-llm-apis",
|
|
"title": "Load Testing LLM APIs — Why k6 and Locust Lie",
|
|
"questions": [
|
|
{
|
|
"stage": "pre",
|
|
"question": "What is the GIL trap in Locust-based LLM load testing?",
|
|
"options": [
|
|
"Locust only works on Windows",
|
|
"Locust does not support HTTP",
|
|
"Client-side tokenization runs under the Python GIL and queues behind request generation, inflating reported inter-token latency",
|
|
"Locust requires CUDA"
|
|
],
|
|
"correct": 2,
|
|
"explanation": ""
|
|
},
|
|
{
|
|
"stage": "check",
|
|
"question": "What is the prompt-uniformity trap?",
|
|
"options": [
|
|
"Uniform prompts always slow the server down",
|
|
"Sampling from a real distribution under-represents long prompts",
|
|
"Uniform prompts require streaming",
|
|
"Looping the same prompt makes prefix caching look like full concurrent decode, inflating reported throughput"
|
|
],
|
|
"correct": 3,
|
|
"explanation": ""
|
|
},
|
|
{
|
|
"stage": "check",
|
|
"question": "Which four load patterns does the lesson recommend?",
|
|
"options": [
|
|
"Constant 1 RPS for 10 days",
|
|
"Manual click tests",
|
|
"Steady-state, ramp, spike, soak",
|
|
"Burst only"
|
|
],
|
|
"correct": 2,
|
|
"explanation": ""
|
|
},
|
|
{
|
|
"stage": "check",
|
|
"question": "How does the lesson recommend building a realistic prompt distribution?",
|
|
"options": [
|
|
"Sample from a real distribution using mean and stddev (for example LLMPerf's --mean-input-tokens / --stddev-input-tokens) or replay real traffic",
|
|
"Random characters per request",
|
|
"Hand-write 5 prompts and shuffle",
|
|
"Always use the same prompt to maximize cache hits"
|
|
],
|
|
"correct": 0,
|
|
"explanation": ""
|
|
},
|
|
{
|
|
"stage": "post",
|
|
"question": "Which 2026 tool combination is positioned as best for CI/CD SLA gates and Kubernetes-native distributed runs?",
|
|
"options": [
|
|
"k6 v2026.1.0 with the k6 Operator 1.0 GA (TestRun / PrivateLoadZone CRDs)",
|
|
"guidellm only",
|
|
"Vegeta only",
|
|
"Locust 2.43.3 stock"
|
|
],
|
|
"correct": 0,
|
|
"explanation": ""
|
|
},
|
|
{
|
|
"stage": "post",
|
|
"question": "Which failure mode does the soak load pattern catch?",
|
|
"options": [
|
|
"Memory leaks, connection-pool drift, and observability overflow over hours",
|
|
"Cold-start tail",
|
|
"Tokenizer GIL contention",
|
|
"Cache eviction storms"
|
|
],
|
|
"correct": 0,
|
|
"explanation": ""
|
|
}
|
|
]
|
|
}
|