TRACES · BENCHMARK RESULTSPipeline Performance Benchmarks
Solid bars sustained the offered load. Dashed bars fell short; their labels show achieved requests/sec. N/A means no verified measurement. All tiers are tested independently. One 60-second trial per case; CPU includes processing through final delivery, divided by the 60-second offered-load window, including any final request and delivery drain. Memory is final RSS, not peak usage. These are fixed-load comparisons, not maximum-capacity estimates.
512 spans/request form 256 complete two-span traces. Trace IDs are unique across requests. Filter drops health traces (25%); Mask replaces the secret attribute; Sample retains approximately 50% by trace ID; Tail retains error traces (25%) after a 1-second decision window, with a 4-million-trace buffer. Collectors batch Tail output on a 500ms timer; Rust uses native decision flushing, including ingestion-assisted decisions in the latest build. Actual Tail batch sizes remain native. Resource metadata is stable across requests. Sample checks retention and matching parent/child fingerprints; it does not independently prove the exact vendor-specific retained set or rule out small delivery loss. Other scenarios verify exact output counts. Enrich is excluded.
This lower ladder was selected before the run to leave headroom below the slowest prior Tail result (~700 requests/sec). It compares fixed-load efficiency, not maximum capacity. All five scenarios and all three agents are included.
Measured trials and verification
| Scenario | Offered req/s | Vendor | Reached req/s | CPU cores | Final RSS MB | Delivery / operation | Status |
|---|
Measured: 2026-11-13. Load names are relative to this signal’s ladder; compare the actual input rates shown above.