| #12077: Active KV use is not reusable cache residency | The usage gauge measures active/referenced capacity, while released blocks may remain reusable until eviction. | 3 | Semantics resolve the interpretation; actual reuse requires same-window hit/query observations. | v0.23.0 |
| #35048: Aggregate throughput can hide experience regressions | Token and iteration metrics describe system work; TTFT queue ITL TPOT and E2E describe distinct parts of per-request experience. | 8 | Schema supplies the decomposition; root cause requires controlled A/B or commit bisect. | v0.23.0 |
| #38194: Prefix-cache percentages require comparable definitions | A percentage hides its numerator, denominator, labels, process scope, routing, reset history, tokenization, and time window. | 2 | Schema defines the vLLM side; the other system and matched runtime data are required for comparison. | v0.23.0 |
| #40696: Full-block eligibility creates prefix-cache cliffs | Full-block matching can create sharp eligibility boundaries; zero hits need not mean the counter is broken. | 4 | Semantics explain the mechanism; model geometry and net impact require version-matched measurement. | v0.23.0 |
| #8115: TTFT observations and completed requests count different events | TTFT is observed on first-token production; request_success increments from finished requests, so lifecycle and failure paths can separate them. | 2 | Schema proves different update events; a particular delta needs deployment observations. | v0.23.0 |
| #9346: vLLM metrics do not provide an all-HTTP-request denominator | request_success counts finished engine requests; running requests is a gauge; neither counts every HTTP attempt. | 2 | Lifecycle semantics rule out proposed denominators; complete SLI needs external API-layer telemetry. | v0.23.0 |
Schema.ai· vLLM