Raw per-step capacity tables — 2026-08-29/30 campaign

Backing data for the summary in ../soak-results.md (procedure: ../../soak/README.md) and for grpc/grpc-java#13012. Setup: talos-main (arm64), identical 1-CPU/1-Gi Guaranteed pods, one arm per 4-core worker, sequential runs, k6 ramping-arrival-rate 200→2400 req/s in 2-minute steps (10 s climb + 110 s hold), tiny unary echo, warm JVM client on its own worker. Each row: the step's 110 s hold window. grpc-java 1.83.1, non-shaded Netty 4.2.16.Final, Clojure 1.12.5 via clj-grpc 0.1.4.

Known artifacts, disclosed: the first step of every JVM-based run is JIT warmup (the arm rolls immediately before its run). Run 1's arm was liveness-killed at ~00:06:41Z and again near ramp top (probes later relaxed to failureThreshold 60). Run 3's arm was liveness-killed at ~00:56Z. Run 5's 1400-offered window shows a transient inversion (dropped-iterations spike that window) inconsistent with its neighbors.

Summary

Capacity ramps (200→2400 req/s, 2-min steps, one arm at a time, same pods): max sustainable goodput per identical 1-CPU pod —

configuration	plateau	overload behavior
gRPC JVM :direct	~2,140 req/s	graceful (h2 flow-control backpressure)
gRPC JVM VT default	~2,060 req/s	graceful, tails degrade
gRPC native VT	~1,550 req/s	liveness-starvation kills before probe tuning
REST (Jetty, HTTP/1.1)	~960 req/s	queue-death: p50 pins at 1.7 s, then collapse

At equal resources gRPC sustains 2.2× (JVM) / 1.6× (native) REST's throughput. The executor trade inverts with load: VT protects tails ~2× at ~10% utilization; above ~75% utilization :direct wins goodput and tails (p99 382 vs 662 ms at 2000 req/s). Peak heaps under saturation: native 81 MB, gRPC-JVM 61 MB, REST 267 MB.

A grpc-netty patch adding a server-side WriteQueue.drainNow() (upstream prototype) reproduced its JMH signature on-cluster: p99 −20–23% across 1200–2200 req/s and −64% at 2400 (1994→717 ms), goodput unchanged.

Run 1 — grpc-native, VT executor (start 23:49:12Z)

offered	delivered	p50ms	p99ms	dropped/s
200	200.3	2.6	9.1	0
400	400.5	2.7	13.1	0
600	600.0	2.9	23.7	0
800	800.8	3.8	48.9	0
1000	999.5	11.0	81.3	0.3
1200	1197.6	30.9	187.8	2.8
1400	1379.3	82.3	457.7	20.8
1600	1499.7	426.0	2226.4	91.9
1800	1580.6	760.3	2455.1	187.6
2000	1567.6	832.9	2465.4	432.4
2200	1607.7	823.3	2459.5	598.8
2400	1524.8	840.1	2496.5	818.0

Run 2 — grpc-jvm, VT executor (start 00:13:46Z)

offered	delivered	p50ms	p99ms	dropped/s
200	207.6	3.9	4326.1	0.0
400	400.2	2.8	12.8	0.0
600	600.2	3.0	24.9	0.0
800	800.4	3.3	44.5	0.0
1000	1000.6	3.9	59.5	0.0
1200	1201.8	4.9	86.8	0.0
1400	1393.0	10.7	827.9	6.9
1600	1600.7	16.0	238.7	0.0
1800	1791.7	34.9	324.9	0.0
2000	1991.9	69.2	661.8	10.8
2200	2059.3	126.6	972.5	132.4
2400	1984.3	254.6	2069.8	339.7

Run 3 — rest (Pedestal/Jetty 12, HTTP/1.1, platform pool) (start 00:38:20Z)

offered	delivered	p50ms	p99ms	dropped/s
200	201.9	207.1	10000.0	3.8
400	396.8	2692.3	7685.4	9.5
600	601.4	4.6	239.9	0.0
800	799.9	9.1	95.5	0.0
1000	962.5	1634.6	2486.2	30.0
1200	960.7	1697.2	2487.9	239.1
1400	959.1	1700.3	2488.8	441.0
1600	958.3	1696.7	2488.3	642.1
1800	957.1	1689.7	2490.3	843.1
2000	43.2	0.0	9696.5	1942.6
2200	244.6	6572.0	10000.0	1954.6
2400	348.1	5500.6	7460.2	1995.7

Run 4 — grpc-jvm, directExecutor (start 01:04:21Z)

offered	delivered	p50ms	p99ms	dropped/s
200	206.6	3.1	4448.4	0.0
400	400.4	2.6	9.6	0.0
600	599.5	2.8	21.1	0.0
800	800.3	3.0	38.3	0.0
1000	1000.2	3.4	50.4	0.0
1200	1200.7	4.0	86.6	0.0
1400	1403.1	4.9	113.0	0.0
1600	1602.3	9.4	222.5	0.0
1800	1797.5	17.1	289.3	3.7
2000	1991.0	32.4	381.9	4.0
2200	2137.4	60.3	649.7	78.0
2400	1978.8	94.4	1993.9	340.7

Run 5 — grpc-jvm, directExecutor + drainNow patch (start 01:29:48Z)

grpc-netty classes replaced with the wedge/server-drain-now-1.83.1 build (v1.83.1 + server-side WriteQueue.drainNow() in NettyServerHandler.channelReadComplete); everything else identical to run 4, same node.

offered	delivered	p50ms	p99ms	dropped/s
200	203.9	2.9	2125.2	0.0
400	400.3	2.6	9.6	0.0
600	600.6	2.7	18.4	0.0
800	799.9	2.9	32.4	0.0
1000	1000.8	3.4	49.0	0.0
1200	1196.3	4.0	66.6	0.0
1400	1397.7	5.6	260.1	1.1
1600	1594.6	10.1	222.0	5.4
1800	1796.0	20.3	230.7	1.0
2000	1986.6	43.5	391.6	6.5
2200	2116.4	75.5	498.1	77.4
2400	2013.4	104.5	717.1	318.1

Heap (per-minute /metrics sampling across runs 1–4)

arm	idle	peak under saturation
grpc-native	6 MB	81 MB
grpc-jvm	11 MB	61 MB
rest	30 MB	267 MB