Test Environment

Machine Intel i5-11320H @ 3.20 GHz, 4 physical cores / 8 logical, 15 GiB RAM
OS Linux 6.14, NVMe storage
Target 127.0.0.1:9889 (loopback)
Protocol HTTP/3 over QUIC → HTTP/2 upstream, TLS (self-signed)
Backends 2 × upstream servers on 127.0.0.1:7001 and 127.0.0.1:7002
Config config/config.development.yaml with benchmark-specific performance overrides (worker_threads=4, per_backend_inflight_limit=256, global_inflight_limit=4096, adaptive admission enabled)
Run ID 20260501T170035Z

Scenarios

Burst — peak concurrency

3,000 requests, 120 concurrent connections.

Metric Value
Throughput 21,235 req/s
Success rate 3000/3000 — 100%
p50 latency 19.1 ms
p95 latency 87.8 ms
p99 latency 102.4 ms

Burst — high concurrency

3,000 requests, 80 concurrent connections.

Metric Value
Throughput 14,691 req/s
Success rate 3000/3000 — 100%
p50 latency 25.1 ms
p95 latency 57.7 ms
p99 latency 64.6 ms

Slow upstream — backend latency simulation

1,000 requests, 80 concurrent connections, upstream introduces delay.

Metric Value
Throughput 9,549 req/s
Success rate 1000/1000 — 100%
p50 latency 26.9 ms
p95 latency 58.1 ms
p99 latency 62.0 ms

QUIC packet loss — lossy network simulation

1,500 requests, 120 concurrent connections, simulated packet loss.

Metric Value
Throughput 12,500 req/s
Success rate 1500/1500 — 100%
p50 latency 44.2 ms
p95 latency 79.7 ms
p99 latency 91.2 ms

Summary

Spooky sustains 21k+ req/s at full concurrency with zero errors across burst, slow-upstream, and packet-loss scenarios. p99 latency stays under 103 ms even at 120 concurrent connections.

The async data plane, per-backend inflight limits, and adaptive admission control ensure that slow or lossy backends do not stall the proxy — fast paths remain fast while degraded upstreams are isolated.