Theoretical Quiz: MLOps and LLMOps
Question 1
Why is Server-Sent Events (SSE) preferred over standard request-response HTTP completions for user-facing LLM applications?
- A) SSE encrypts the tokens with quantum cryptography.
- B) SSE reduces Time-To-First-Token (TTFT) perception by streaming tokens interactively as they are sampled, rather than blocking the client until generation finishes.
- C) SSE eliminates API token usage.
- D) SSE automatically validates SQL code.
Answer: B
Explanation: Streaming delivers the first token within hundreds of milliseconds, giving users immediate feedback.
Question 2
What is the primary architectural limitation of deploying heavy open-source models (like Llama-3-70B) on AWS Lambda?
- A) AWS Lambda only supports JavaScript.
- B) AWS Lambda has no GPU hardware accelerator support and has a 10 GB maximum container image limit.
- C) AWS Lambda cannot connect to the internet.
- D) AWS Lambda requires daily restarts.
Answer: B
Explanation: AWS Lambda does not support GPUs and has a 10GB container image limit, making it unsuitable for hosting 70B parameter models.
Question 3
How does semantic caching differ from traditional key-value caching in an LLM pipeline?
- A) Semantic caching only works with MongoDB.
- B) Semantic caching compares the cosine similarity of query embedding vectors rather than exact string equality, allowing semantically identical rephrasings to hit the cache.
- C) Semantic caching generates new responses automatically.
- D) Semantic caching increases OpenAI billings.
Answer: B
Explanation: Natural language queries vary in phrasing; semantic caching matches queries based on meaning rather than exact character equality.