Call videoSrc.EnableKeyFrameCache() immediately after binding the video
UDP Source in allocAdjacentPair(). Audio sources are intentionally left
uncached — IDR detection on Opus packets would accumulate the entire
audio stream without ever resetting the burst.
promauto is not in the vendor tree. Replace promauto.With(reg).NewXxx()
with prometheus.NewXxx() + reg.MustRegister() — functionally identical
but uses only the already-vendored prometheus/client_golang/prometheus
package. Fixes the vendor-mode build error:
cannot find module providing package .../prometheus/promauto
Conflict resolution: keep M3's full handler.go rewrite (per-stream
index, error matrix, PATCH, CORS, auto-cleanup) and re-apply the
swagger annotations from #7 onto the new function declarations,
including a fresh annotation for the M3-introduced Trickle endpoint.
Swagger docs regenerated to pick up all three.
Race-clean: go test -race ./app/webrtc/... green.
M5 / final M2-stack work. The fork now identifies itself unambiguously
in logs, the API, and the README without changing the Go module path
(internal imports stay at github.com/datarhei/core/v16 — see NOTES.md
for the rationale).
Identity surfaces:
- app/version.go gains Variant ('dragonfork') and Fork ('Datarhei —
Dragon Fork') as vars (overridable via -ldflags for downstream
re-packagers).
- api.About + the /api endpoint expose 'variant' and 'fork' fields;
Swagger docs regenerated.
- Startup banner logs 'variant' + 'fork' alongside the existing
application + version fields, so a TrueNAS sysadmin tail-following
/var/log can tell at a glance which fork is running.
Documentation:
- README.md rewritten with a Dragon Fork header and Quick start; the
upstream feature surface is summarised in 'From upstream Datarhei'
with a clear additivity statement. Sample process JSON, multi-input
pipeline guidance, link to the design + testing docs.
- NOTICE: Apache 2.0 §4(d) attribution to upstream datarhei Core,
Pion, Echo, FFmpeg.
- CREDITS: enumerated dependency list with licenses.
- CHANGELOG.md prepended with a 'Datarhei — Dragon Fork' section
starting at v0.1.0-dragonfork; upstream's '# Core' history preserved
below.
Module path stays github.com/datarhei/core/v16 by design — the fork is
distinguished by repo location and branch history, not import path.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Adds an end-to-end RTP-arrival latency probe that runs as a dedicated
CI job and asserts p95 < 50ms.
Implementation
--------------
A build-tagged test (-tags latency, off by default) sends 1000
synthetic RTP packets at 60Hz into corewebrtc.Source and reads them
back via a Pion subscriber's track.ReadRTP(). Each packet's payload
starts with the publisher's UnixNano send time; the subscriber diffs
against time.Now() at arrival and accumulates p50/p95/p99.
This exercises every link of the egress hop: Source UDP read,
subscriber fan-out, forwardRTPSplit, Pion's TrackLocalStaticRTP
write, DTLS-SRTP encrypt, ICE socket write, decrypt at the
subscriber, RTP unmarshal at ReadRTP. Pure server-side; no FFmpeg
or codecs involved.
Why not glass-to-glass
----------------------
The design's §7 calls for FFmpeg drawtext frame counters + decode-
side pixel sampling, p95<300ms RTMP / <200ms SRT. Implementing that
in pure Go needs a cgo H.264 decoder or an FFmpeg sidecar pipe — a
significantly bigger lift for a marginal regression-detection win
(encode/decode latency is roughly fixed by the codec stack and
isn't moved by Core code changes). The server-hop measurement
captures everything Core code can actually regress.
Threshold
---------
50ms p95. Locally observed on a quiet host:
p50=110µs, p95=237µs, p99=318µs.
The 50ms gate is ~200x headroom — generous enough to absorb CI
runner noise without false alarms, tight enough to catch a real
slowdown.
Race-clean: latencySamples uses a sync.Mutex around the slice append
(initial draft had a slice racing with the receive goroutine; vet
caught it).
Documented in test/TESTING.md and wired to .forgejo/workflows/test.yml
as the latency-gate job (depends on lint-and-vet, parallel with test
and webrtc-smoke).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The WHEP routes were mounted by http/server.go via the app/webrtc
Handler.Register(), but Subscribe and Unsubscribe carried no swag
annotations. The Swagger UI at /api/swagger/index.html therefore
didn't list /api/v3/whep/* — programmatic API consumers and humans
browsing the docs couldn't discover the endpoints.
Adds the standard upstream-shaped @Summary / @Tags / @ID / @Router
annotations on Subscribe and Unsubscribe (matching the rtmp.go and
srt.go pattern) and regenerates docs/{docs.go,swagger.json,swagger.yaml}
via 'make swagger'. Verified: swagger.json now contains both paths,
swagger UI renders them under the v16.16.0 tag.
Closes#3.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
BuildArgs hardcoded -map 0✌️0 / -map 0🅰️0 for the two RTP legs.
Correct for production RTMP/SRT publishers (single combined input),
but breaks any process whose audio lives on a different input index
— multi-input lavfi test scaffolds, multi-camera pipelines, SDI +
file-audio mixes, etc.
Adds VideoMap and AudioMap fields to ConfigWebRTC (and the API DTO),
defaulting to the prior literals so existing deployments are
unaffected. BuildArgs reads them.
Tests:
- TestBuildArgs_DefaultMaps locks the empty-string default behavior
- TestBuildArgs_CustomMaps drives the multi-input override path
- TestProcessConfigWebRTCMapsRoundtrip extends the DTO roundtrip
Closes#2.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
TestIntegration_FiveViewerFanout drives the M3 acceptance criterion
in the wide direction: spin up the subsystem, register one process,
attach 5 Pion subscribers in parallel via the real Echo handler,
spray synthetic RTP at the allocated UDP ports, and assert each
subscriber's video + audio track receive at least one packet inside
a 15s window. After onProcessStop, the per-stream peer index must
drain to zero within 3s.
TestSubsystem_TeardownHookFiresOnProcessStop is the unit-level
counterpart — confirms the callback registered via
SetTeardownHook actually fires when a process is torn down, even
without a full Pion handshake.
Together these cover the acceptance language: '5 concurrent viewers,
all error paths correct, clean teardown'.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Covers each new code path that the design's §6 table requires:
- Subscribe -> 406 on non-H264 / non-Opus offer (TestHandler_Subscribe_406OnCodecMismatch)
- Subscribe -> 503 when total cap exhausted (TestHandler_Subscribe_503OnTotalCap)
- Subscribe -> 503 when per-stream cap exhausted (TestHandler_Subscribe_503OnPerStreamCap)
- Trickle -> 404 on unknown resource (TestHandler_Trickle_404WhenUnknown)
- preflight -> 204 + CORS headers (TestHandler_PreflightCORS)
- Register installs all 5 routes (TestHandler_RegisterMountsAllRoutes)
- Close drains the index without panicking (TestHandler_Close_DrainsPeers)
- requireH264AndOpus table-driven (TestRequireH264AndOpus)
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Major Handler rewrite implementing the design's M3 acceptance
criteria ('5 concurrent viewers, all error paths correct, clean
teardown'):
Multi-viewer correctness:
- streamID -> resourceID -> Peer two-level index (was flat)
- per-stream peer cap alongside total cap, defaults match the
design's '5–8 viewer' target (8/stream, total from corewebrtc)
- per-peer awaitPeerClose goroutine watches Peer.Done() so ICE
failures yank the index entry + decrement the counter (no leaks)
- tearDownStreamPeers callback (registered with Subsystem in
NewHandler) drives all peer closes when the source process stops
Error matrix from design §6:
- 406 on codec mismatch (offer missing H264 or Opus rtpmap)
- 504 on ICE gathering timeout (passthrough from CreatePeerFromSources)
- 204 on DELETE unknown resource (idempotent per WHEP spec; was 404)
- 503 on per-stream cap reached (separate body from total-cap 503)
- 400 on missing/empty body (unchanged)
- 404 on unknown stream (unchanged)
WHEP spec compatibility:
- PATCH /whep/:id/:resource for trickle-ICE
- OPTIONS preflight on every WHEP path
- CORS Allow-Origin/Methods/Headers + Expose-Headers (Location, ETag)
- ETag header on Subscribe response
Defensive nil-peer guards in tearDown / Close paths so a partial
state doesn't panic.
Refactor: 134 -> 341 lines on handler.go but the surface is the
same (NewHandler/Register/Subscribe/Unsubscribe/Close); existing
callers continue to work. Pre-M3 test 'Unsubscribe_404WhenUnknown'
renamed and updated to the new 204 expectation.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Subsystem.SetTeardownHook installs a callback the subsystem invokes
just before closing per-stream Sources in onProcessStop. Used by the
WHEP Handler in M3 to drain its per-stream peer index before the
underlying Sources go away — closes the 'subscribers fan out into a
closed channel' race the design's §6 error matrix calls out as
'Publisher disconnects / FFmpeg exits'.
Single consumer by design (one subsystem, one handler). Calling
SetTeardownHook again replaces the previous callback; nil detaches.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
End-to-end exercise of the M2 pipeline — subsystem hook, port
allocation, two-track forwarding, WHEP handshake — without
spinning up a full Core HTTP server:
- Fire onProcessStart directly to get the two RTP legs back
- Parse video + audio UDP ports out of the leg addresses,
assert adjacency
- Mount the Handler on an Echo httptest server
- Build a Pion PeerConnection (recvonly video + audio), POST
its offer, feed the answer back in
- Spray synthetic RTP packets at both loopback sockets
- Assert both OnTrack callbacks fire and each delivers at least
one RTP packet within 10s
- DELETE via the returned Location header to confirm teardown
Passes cleanly under -race in ~1s. Catches regressions across
the whole M2 wiring from a single fixture.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Installs the WebRTC egress subsystem at Core boot when
cfg.WebRTC.Enable is true and the subsystem constructs cleanly:
- http.Config gains an optional WebRTC *appwebrtc.Handler field;
server.setRoutesV3 mounts its WHEP routes on the JWT-protected
/api/v3 group.
- api.start() constructs the Subsystem, registers its ProcessHooks
with the restreamer, and builds a Handler. A construction failure
is logged and Core continues without WebRTC — consistent with
disabling the subsystem outright.
- api.stop() closes the Handler (tearing down active peers) before
closing the Subsystem (releasing per-process UDP sockets), mirroring
the RTMP/SRT teardown pattern.
Verified: go build ./... clean; go test ./app/webrtc/...
./core/webrtc/... ./restream/... ./http/... all pass.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Introduces the HTTP surface the browser (or OBS WebRTC clients)
target when subscribing to a process's egress:
POST /whep/:id -> answer SDP + Location header
DELETE /whep/:id/:resource -> tear down a specific peer
The handler looks up the per-process stream pair via the Subsystem,
validates SDP offer shape, and delegates peer creation to the core
PeerFactory's CreatePeerFromSources (two-source forwarding).
WHEP routes are left unauthenticated in M2 — browsers and OBS don't
carry the Core JWT, and per-process signed-URL tokens are an M3
enhancement. Deployments should place the endpoint behind an
authenticated reverse-proxy for now.
Tests cover:
- 404 for POSTs against unregistered streams
- 400 for empty/invalid SDP offers once a stream is registered
- 404 for DELETE against unknown resource ids
Introduces the subsystem layer that sits alongside api.API and wires
the M1 core/webrtc primitives into the per-process restream lifecycle.
app/webrtc/subsystem.go:
- Subsystem struct holding the global WebRTC config, core PeerFactory,
per-process stream map, and logger
- New(config.DataWebRTC, logger) constructor
- Enabled(), Hooks(), Close(), lookup() methods
app/webrtc/lifecycle.go:
- onProcessStart: allocates an adjacent UDP port pair, binds two
Pion Sources (video on V, audio on V+1), registers them under the
process id, and returns the two RTP output legs to append to the
FFmpeg command.
- onProcessStop: tears down the pair.
- allocAdjacentPair: retries up to 10 times to find a free (V, V+1)
pair since the kernel's ephemeral picker can hand us an odd port.
- splitRTPLegs: converts BuildArgs' flat []string into two ConfigIO
entries by splitting on the second -map token.
core/webrtc/peer.go + forward.go:
- Adds PeerFactory.CreatePeerFromSources for the M2 two-source
forwarding mode (video and audio on separate UDP ports, no
payload-type sniffing). Leaves CreatePeer intact for the M1 PoC.
- Adds forwardRTPSplit companion goroutine.
config/data.go:
- Promote anonymous WebRTC struct to named type DataWebRTC so
app/webrtc can accept it by value.
Adds Alloc(), the ephemeral loopback UDP port grabber the subsystem
uses to pick the RTP port it will hand to FFmpeg and then re-bind with
core/webrtc.NewSourceOn. Covered by a 100x rebind test.
Adds BuildArgs(), which emits the -f rtp output fragments (video on
the passed port, audio on port+1) with copy codecs by default and an
H.264 baseline / libopus re-encode leg when ForceTranscode is set.
Covered by three unit tests.
If purging is enabled, overwriting a file with a file of the same
or smaller size will not result in an error.
It is now possible to change the purging mode on an existing sized
filesystem.
