Raw streaming capacity tables — 2026-08-30

Backing data for the streaming results in ../soak-results.md. Setup: 1-CPU/1-Gi grpc-jvm pod, driver (clj-grpc.soak.stream-driver) on its own worker, persistent bidi Chat streams, absolute-schedule open-loop pacing, bounded in-flight (256/stream), per-message client-side latency. First step of every run is JIT warmup (fresh pods) — excluded from conclusions.

Summary

Streaming capacity (2026-08-30, bidi Chat echo, same grpc-jvm pod, raw tables in results/): streaming moves ~15,000–16,000 msg/s on one core — ~7.5× the unary gRPC plateau and ~16× REST — with per-message p50 under 2 ms and p99 under 30 ms through 8,000 msg/s. The executors split decisively at streaming's limits: :executor :direct holds p99 2–2.5× lower than the VT default at every rate above 6,000 and keeps delivering at 16,000 (38 deferred/s vs VT's 625/s) — the per-message dispatch the VT executor still pays is the one cost streaming cannot amortize. Full doctrine across every measured regime: VT wins only low-utilization unary tails; :direct wins high-load unary, all streaming, capacity, and CPU — provided handlers never block. And the library's oldest guidance is now cluster-quantified: one stream really does beat N unaries, by ~7.5× per core.

S1 — VT executor, 20 streams, 400–4,800 msg/s

offered msg/s	sent/s	echoed/s	p50ms	p99ms	p999ms	deferred/s
400	400.0	400.0	3.34	2460.29	2460.29	0.0
800	800.0	800.0	2.47	8.22	14.97	0.0
1200	1200.2	1200.2	2.47	8.22	14.97	0.0
1600	1600.0	1600.0	2.47	8.22	20.21	0.0
2000	2000.0	2000.0	1.83	8.22	20.21	0.0
2400	2400.2	2400.2	1.83	8.22	14.97	0.0
2800	2800.2	2800.2	1.83	8.22	27.29	0.0
3200	3200.0	3200.0	1.83	8.22	14.97	0.0
3600	3600.2	3600.2	1.83	11.09	27.29	0.0
4000	4000.0	4000.0	1.83	8.22	14.97	0.0
4400	4400.2	4400.2	1.83	11.09	20.21	0.0
4800	4800.2	4800.2	1.83	11.09	27.29	0.0

S2 — :direct, 20 streams, 400–4,800 msg/s

offered msg/s	sent/s	echoed/s	p50ms	p99ms	p999ms	deferred/s
400	400.0	400.0	3.34	1822.43	2460.29	0.0
800	800.0	800.0	2.47	11.09	122.36	0.0
1200	1200.2	1200.2	2.47	8.22	14.97	0.0
1600	1600.0	1600.0	1.83	8.22	11.09	0.0
2000	2000.0	2000.0	1.83	8.22	14.97	0.0
2400	2400.2	2400.2	2.47	8.22	14.97	0.0
2800	2800.2	2800.2	1.83	8.22	20.21	0.0
3200	3200.0	3200.0	1.36	8.22	20.21	0.0
3600	3600.2	3600.2	1.83	11.09	49.73	0.0
4000	4000.0	4000.0	1.83	8.22	14.97	0.0
4400	4400.2	4400.2	1.83	11.09	20.21	0.0
4800	4800.2	4800.2	1.83	11.09	27.29	0.0

E1 — VT executor, 40 streams, 4,000–16,000 msg/s

offered msg/s	sent/s	echoed/s	p50ms	p99ms	p999ms	deferred/s
4000	3154.9	3154.9	4.51	20105.83	20105.83	845.1
6000	6000.3	6000.3	2.47	27.29	67.14	0.0
8000	8000.0	8000.0	3.34	27.29	67.14	0.0
10000	10000.0	10000.0	4.51	90.64	122.36	0.0
12000	12000.3	12000.3	4.51	165.19	301.06	0.0
14000	14000.3	14000.3	11.09	406.43	406.43	0.0
16000	15375.1	15375.1	27.29	1349.95	2460.29	625.0

E2 — :direct, 40 streams, 4,000–16,000 msg/s

offered msg/s	sent/s	echoed/s	p50ms	p99ms	p999ms	deferred/s
4000	3641.0	3641.0	2.47	14893.21	14893.21	359.0
6000	6000.3	6000.3	1.83	14.97	27.29	0.0
8000	8000.0	8000.0	2.47	20.21	90.64	0.0
10000	10000.0	10000.0	3.34	49.73	122.36	0.0
12000	12000.3	12000.3	3.34	67.14	122.36	0.0
14000	14000.3	14000.3	4.51	165.19	301.06	0.0
16000	15961.7	15961.7	11.09	740.71	740.71	38.3