Diagnose Model Serving Bottleneck

Tell me how you would diagnose and resolve a bottleneck in a model serving pipeline.

Implement a function that analyzes stage latencies and throughput samples from a serving pipeline and returns the stage most likely causing end-to-end slowdown. If multiple stages are tied, choose the earliest stage in the pipeline.

Function Signature

def diagnose_bottleneck(stages, samples):

Input

• stages: list of stage names in pipeline order
• samples: list of dicts, each with stage, latency_ms, and throughput_rps

Output

Return the name of the bottleneck stage.