PRODUCTION SCENARIO
A Generative AI Engineer is handed one request ID for an agent that took 31 seconds to answer. The agent calls a retriever, two tools and the LLM, and the engineer needs to know which of those steps consumed the time.
Which capability gives the engineer that breakdown?
Answering here is anonymous. Nothing is saved unless you sign in.
Show answer and explanation
Answer: MLflow Tracing, which records each step's inputs, outputs and duration
MLflow Tracing records inputs, outputs, intermediate steps and metadata so you can see exactly how the app behaves, which is what turns one slow request into a per-step breakdown. Inference tables capture incoming requests and outgoing responses but nothing between them, the gateway usage table reports one total latency per request rather than per step, and an evaluation run scores answer quality instead of attributing time.
Official Source: https://docs.databricks.com/aws/en/mlflow3/genai/tracing/ ↗