You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
otb: perf column tooltips describe the same-dialect passthrough methodology
The lat/rps20/rpsmax tooltips said 'best supported path' (pre-decision). Now they say
the numbers are measured on the gateway's same-dialect passthrough (the Tested-on dialect),
the cell chosen by lowest added latency - pure forwarding. Latency selects the cell;
sustained RPS is the external comparison metric.
id: "lat",label: "Added latency p99 (µs)",desc: false,title: "Gateway p99 minus direct-to-mock p99 at concurrency 1, on the gateway's best supported path",
176
+
id: "lat",label: "Added latency p99 (µs)",desc: false,title: "Gateway p99 minus direct-to-mock p99 at concurrency 1, on the gateway's same-dialect passthrough (the 'Tested on' dialect); the cell is chosen by lowest added latency, pure forwarding",
id: "rps20",label: "Sustained RPS @20ms",desc: true,title: "Sustained requests/sec with a 20 ms mock LLM latency (p99 < 1 s, <0.1% errors), on the gateway's best supported path",
180
+
id: "rps20",label: "Sustained RPS @20ms",desc: true,title: "Sustained requests/sec with a 20 ms mock LLM latency (p99 < 1 s, <0.1% errors), on the gateway's same-dialect passthrough (the 'Tested on' dialect); the cell is chosen by lowest added latency, pure forwarding",
id: "rpsmax",label: "Max proxy RPS",desc: true,title: "Throughput ceiling against an instant mock (p99 < 1 s, <0.1% errors), on the gateway's best supported path",
184
+
id: "rpsmax",label: "Max proxy RPS",desc: true,title: "Throughput ceiling against an instant mock (p99 < 1 s, <0.1% errors), on the gateway's same-dialect passthrough (the 'Tested on' dialect); the cell is chosen by lowest added latency, pure forwarding",
0 commit comments