benchmark

greyproxy benchmark – runs 2/3/4

All latencies in ms. Run 4 = no conversation tracking.

Scenario	Mode	Run	TTFB p50	TTFB p95	Total p50	Total p95	Errors	Tput (req/s)
small	direct	2	0.14	0.27	0.40	0.54	0	100.0
small	direct	3	0.19	0.34	0.45	0.63	0	100.0
small	proxy	2	2.68	3.42	3.50	4.28	0	100.0
small	proxy	3	3.08	4.05	3.94	4.94	2	99.9
small	proxy	4 no-conv	3.08	4.27	3.93	5.16	0	100.0
medium	direct	2	0.20	0.35	4.35	5.51	0	100.0
medium	direct	3	0.23	0.37	4.42	5.57	0	100.0
medium	proxy	2	3.55	4.60	16.21	18.33	110	96.3
medium	proxy	3	3.64	4.51	16.38	18.74	113	96.2
medium	proxy	4 no-conv	3.51	4.44	16.10	18.20	79	97.3
large	direct	2	1.44	2.10	26.18	28.95	0	99.9
large	direct	3	1.40	1.93	25.95	28.92	0	99.9
large	proxy	2	6.01	8.97	75.15	85.45	700	76.5
large	proxy	3	5.78	8.53	75.58	85.77	665	77.7
large	proxy	4 no-conv	5.50	7.59	70.08	79.41	603	79.8

Key takeaways:

• TTFB overhead is ~2.5-3.5ms constant across payload sizes (proxy processing cost)
• Total latency overhead scales with body size — points to body buffering in the hot path
• Error rates: 0% small, ~3.7% medium, ~20-23% large — all at 100 req/s
• Disabling conversation tracking helped medium errors (-30%) and large latency (-7%) but is not the primary bottleneck
• Large payload errors remain ~20% even without conversation tracking — body copy path is the likely culprit