librtmp2-server 0.6.3

What's Changed

• Update Rust crate libc to v0.2.190 by @renovate[bot] in #267
• Update Rust crate uuid to v1.27.0 by @renovate[bot] in #268
• fix(bug-hunter): full-scan fixes (2 confirmed bugs) by @AlexanderWagnerDev in #266
• Update Rust crate tokio to v1.53.2 - abandoned by @renovate[bot] in #269
• Update taiki-e/install-action digest to db3874b by @renovate[bot] in #271
• fix(bug-hunter): full-scan fixes (3 confirmed bugs, 1 reverted to manual review) by @AlexanderWagnerDev in #270
• Update taiki-e/install-action digest to 2849dd4 by @renovate[bot] in #272
• fix(bug-hunter): round 6 - cluster health/learner gate, preauth timeout, route rename, sink cleanup, CI exit codes by @AlexanderWagnerDev in #273
• fix(ci): build with --locked by @AlexanderWagnerDev in #274
• fix(bug-hunter): round 7 - subscription refcount guard, non-finite bandwidth config, stale CI diagnostics by @AlexanderWagnerDev in #275
• Update taiki-e/install-action digest to 24b35e7 by @renovate[bot] in #276
• ci(bench): cross-server sweep on every merge and release, compared with the last release by @AlexanderWagnerDev in #277
• fix(bench): build the librtmp2 client without --locked by @AlexanderWagnerDev in #278
• refactor(bench): keep the benchmark data on main instead of a bench-data branch by @AlexanderWagnerDev in #279
• chore(deps): update actions/upload-artifact action to v7.0.2 by @renovate[bot] in #282
• chore(deps): update actions/download-artifact action to v8.0.2 by @renovate[bot] in #281
• chore(deps): update taiki-e/install-action digest to 27cdd32 by @renovate[bot] in #280
• fix(bug-hunter): round 8 - relay restore on publish rename, bench sweep/nginx hardening by @AlexanderWagnerDev in #283
• chore(deps): update rust crate tokio-util to v0.7.20 by @renovate[bot] in #288
• chore(deps): update dtolnay/rust-toolchain digest to 686976e by @renovate[bot] in #287
• fix(logger): reject symlinked LOG_FILE on Unix (O_NOFOLLOW) by @cursor[bot] in #286
• fix(bug-hunter): round 9 - player-reservation, init-cache eviction and subscription-refcount fixes by @AlexanderWagnerDev in #285
• Benchmarks against the newest competitor releases (MediaMTX v1.21.2) by @AlexanderWagnerDev in #284
• docs: fill the Unreleased changelog and re-record the README terminal demo by @AlexanderWagnerDev in #289
• chore(deps): update taiki-e/install-action digest to be76f54 by @renovate[bot] in #290
• chore(deps): update rust crate uuid to v1.28.0 by @renovate[bot] in #291
• Pre-release audit for 0.6.3: release workflow, Docker build, proxy detection, docs by @AlexanderWagnerDev in #292
• release: depend on librtmp2 0.12.0 for 0.6.3 by @AlexanderWagnerDev in #293

Full Changelog: v0.6.2...v0.6.3

Benchmarks

Measured automatically for this release on a CI runner; the machine is described below. Raw data: bench-results-v0.6.3.json.

• Measured: 2026-10-14 22:00:08 UTC · b72fe6c · crate version 0.6.3 · workflow run
• Built on: librtmp2 0.12.0
• CPU: AMD EPYC 9V45 96-Core Processor · 4 vCPUs
• RAM: 15.6 GiB
• System: Linux 6.17.0-1022-azure x86_64 · rustc 1.99.0 · GitHub-hosted ubuntu24 (image 20261004.327.1)
• Noise: shared CI VM — CPU model, neighbouring load and the runner image change between runs, so treat absolute numbers as indicative.
• Comparable vs release v0.6.5: ❌ different hardware (AMD EPYC 9V74 80-Core Processor, 4 vCPUs vs AMD EPYC 9V45 96-Core Processor, 4 vCPUs): only large changes (> 25 %) mean anything

http_api

Benchmark	Time	Throughput	vs release v0.6.5
http_api/create_stream	144.15 µs	—	🟢 −48.7 %
http_api/health	91.80 µs	—	🟢 −37.7 %
http_api/list_streams	566.5 ms	—	⚪ +21.3 %

🟢 faster · 🔴 slower · ⚪ within the noise band (5–25 % depending on how comparable the two environments are, plus both runs' Criterion confidence intervals). Lower time is better.

Cross-server sweep

Same RTMP client and ffmpeg source for every server, one server at a time: load viewers 500 1000 2000, publisher ffmpeg 1280x720@30 H.264 2.5 Mbps + AAC 128 kbps. One sweep per run, no repetitions — the shared runner makes single values noisy, so compare servers within this run rather than across runs.

Servers: librtmp2-server 0.6.3 (librtmp2 0.12.0) · MediaMTX v1.21.2 · LiveForge main @ 4e70fb3 · SRS v8.0-d0 · nginx-rtmp nginx 1.31.6 (nginx-rtmp-module @ 6c7719d)

librtmp2-server vs the baselines

Metric	Now	vs release v0.6.5
publish handshakes/s	10,073	🟢 +41.3 %
publish handshake avg (ms)	2.23	🟢 −33.4 %
publish handshake p95 (ms)	6.54	⚪ +18.1 %
play handshakes/s	11,891	🟢 +108.5 %
play handshake avg (ms)	1.81	🟢 −52.7 %
play handshake p95 (ms)	4.11	🟢 −57.8 %
join, 1 viewer: avg (ms)	0.41	🟢 −28.1 %
join, 1 viewer: p95 (ms)	0.41	🟢 −28.1 %
join, 25 viewers: avg (ms)	1.07	⚪ −22.5 %
join, 25 viewers: p95 (ms)	1.78	⚪ −14.8 %
join, 100 viewers: avg (ms)	1.26	🟢 −58.7 %
join, 100 viewers: p95 (ms)	2.20	🟢 −62.8 %
load, 500 viewers: join avg (ms)	13.93	⚪ −10.9 %
load, 500 viewers: join p95 (ms)	31.45	⚪ −20.2 %
load, 500 viewers: frames/viewer	73.15	⚪ +0.1 %
load, 500 viewers: CPU (% of one core)	26.30	🟢 −30.8 %
load, 500 viewers: peak RSS (MiB)	22.20	⚪ −1.3 %
load, 1000 viewers: join avg (ms)	8.44	🟢 −57.2 %
load, 1000 viewers: join p95 (ms)	26.24	🟢 −50.5 %
load, 1000 viewers: frames/viewer	73.15	⚪ +0.0 %
load, 1000 viewers: CPU (% of one core)	51.90	🟢 −26.9 %
load, 1000 viewers: peak RSS (MiB)	31.80	⚪ −1.9 %
load, 2000 viewers: join avg (ms)	29.89	🟢 −65.6 %
load, 2000 viewers: join p95 (ms)	99.52	🟢 −65.8 %
load, 2000 viewers: frames/viewer	73.14	⚪ −0.1 %
load, 2000 viewers: CPU (% of one core)	98.60	⚪ −11.9 %
load, 2000 viewers: peak RSS (MiB)	52.80	⚪ −0.6 %

Handshake latency

Server	publish /s	avg ms	p95	p99	play /s	avg ms	p95	p99
librtmp2-server	10,073	2.23	6.54	7.01	11,891	1.81	4.11	5.42
MediaMTX	9,114	2.74	4.78	6.03	8,462	2.44	4.21	6.75
LiveForge	9,904	2.23	3.86	5.13	8,376	2.84	5.30	5.88
SRS	594.6	47.32	51.71	51.79	625.9	45.92	50.90	51.17
nginx-rtmp	707.8	41.53	42.83	43.04	358.0	83.08	85.16	85.80

Join latency (avg / p95 ms)

Server	1 viewer	25 viewers	100 viewers
librtmp2-server	0.41 / 0.41	1.07 / 1.78	1.26 / 2.20
MediaMTX	0.62 / 0.62	2.48 / 3.62	1.83 / 3.93
LiveForge	2.37 / 2.37	4.54 / 8.19	6.52 / 18.80
SRS	41.97 / 41.97	44.31 / 45.60	56.93 / 64.44
nginx-rtmp	82.02 / 82.02	82.27 / 82.88	84.52 / 87.66

Many viewers on one stream

Viewers	Server	join avg / p95 ms	CPU % (one core)	peak RSS MiB	frames/viewer
500	librtmp2-server	13.93 / 31.45	26.30	22.20	73.15
500	MediaMTX	20.08 / 37.89	62.80	112.6	73.12
500	LiveForge	65.19 / 124.8	40.20	96.00	73.10
500	SRS	146.8 / 175.2	5.70	139.6	72.38
500	nginx-rtmp	87.19 / 93.51	40.80	23.90	73.06
1000	librtmp2-server	8.44 / 26.24	51.90	31.80	73.15
1000	MediaMTX	30.14 / 70.42	127.5	163.4	73.08
1000	LiveForge	90.84 / 201.7	83.60	151.5	73.13
1000	SRS	408.6 / 1,133	10.10	184.2	72.96
1000	nginx-rtmp	106.5 / 136.4	88.50	29.80	73.08
2000	librtmp2-server	29.89 / 99.52	98.60	52.80	73.14
2000	MediaMTX	120.2 / 248.2	196.1	300.2	69.90
2000	LiveForge	400.3 / 1,058	151.4	275.7	73.07
2000	SRS	581.2 / 1,191	21.50	268.0	73.14
2000	nginx-rtmp	194.0 / 227.7	95.00	35.00	55.62

A competitor delivering fewer frames per viewer than the others was overloaded at that step (the delivered rate, not just latency, is the result).