dsh-latency-pct

In model-inference pipelines, average latency often masks the real experience of tail requests, yet percentile metrics such as P50, P95, and P99 are what actually define your SLO. Computing them by hand is tedious and error-prone, especially during rapid debugging cycles where you need a trustworthy number within seconds rather than minutes of scribbling throwaway scripts.

dsh-latency-pct ships a ready-in-session tool called latency_pct that computes latency percentiles with precise, parameter-driven calculation and returns detailed, structured results. The entire implementation is pure Node.js with zero network calls, so it runs reliably in fully offline environments and sidesteps the version drift and supply-chain risks that accompany external dependencies.

It is a good fit for ML engineers tuning or stress-testing inference pipelines, SREs managing SLO dashboards, and any developer who wants to grab a latency-distribution snapshot inside a conversational workflow. One invocation yields the percentile data you can trust, letting you focus on the model behavior that actually matters.