Compare commits

...

14 Commits

Author SHA1 Message Date
Cédric Verstraeten
bd5df30de3 Merge pull request #284 from kerberos-io/feature/align-pps-sps-in-mp4-construct
feature/align-pps-sps-in-mp4-construct
2026-06-11 18:34:45 +02:00
Cédric Verstraeten
2c063c39c6 Potential fix for pull request finding
Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
2026-06-11 18:33:33 +02:00
Cédric Verstraeten
2f0f29ce8c Detect and isolate upstream loop seam in MP4
Add logic to detect an upstream source loop/restart seam (premature IDR) and force a fragment flush so the seam IDR starts its own fragment. Introduces SeamGapDivisor constant and two MP4 fields (LastKeyframeRawPTS, LastKeyframeGapMs) to track recent keyframe timing; when a keyframe interval is significantly shorter than the previous interval (gap*SeamGapDivisor < previousGap) a fragment boundary is forced and a warning is logged. This approach derives the threshold from the observed GOP cadence to avoid false positives on short-GOP or all-intra streams.

Also add tests (machinery/src/video/mp4_loopseam_test.go) that synthesize loop-seam scenarios across multiple GOP sizes (15, 30, 60 frames) to verify the seam is isolated, and include a sample MP4 reproducer (machinery/thales_1781183923_3-758_2top_0-0-0-0_-1_30221.mp4).
2026-06-11 18:32:23 +02:00
Cédric Verstraeten
a05acb7fc8 Fix SPS/PPS prepend and warn on missing PS
When prepending H.264 parameter sets to keyframes, build a fresh buffer instead of appending into existing slices to avoid corrupting shared backing arrays (which could produce an invalid avcC and trigger FFmpeg "non-existing PPS 0 referenced" errors). Also add explicit error logs in mp4.Close() to surface incomplete H.264/H.265 parameter sets (avcC/hvcC) so missing VPS/SPS/PPS conditions are easier to diagnose.
2026-06-11 17:06:45 +02:00
Cédric Verstraeten
2b88c0ff93 Merge pull request #283 from kerberos-io/feature/add-bump-release-workflow
feature/add-bump-release-workflow
2026-06-10 12:13:06 +02:00
Cédric Verstraeten
4aa2b6e51a Refactor release bump workflow to support multi-architecture builds and improve Docker image handling 2026-06-10 10:01:39 +00:00
Cédric Verstraeten
0c439e34c7 Merge pull request #282 from kerberos-io/feature/add-bump-release-workflow
feature/add-bump-release-workflow
2026-06-10 11:25:44 +02:00
Cédric Verstraeten
d57bea3079 Add release bump workflow for semantic versioning 2026-06-10 10:37:59 +02:00
Cédric Verstraeten
46a48db080 Merge pull request #281 from kerberos-io/feature/add-backpressure-rtsp-logging
feature/add-backpressure-rtsp-logging
2026-06-10 09:49:45 +02:00
Cédric Verstraeten
b7fe9947c2 Potential fix for pull request finding
Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
2026-06-10 09:43:16 +02:00
Cédric Verstraeten
f214a09826 Potential fix for pull request finding
Co-authored-by: Copilot Autofix powered by AI <175728472+Copilot@users.noreply.github.com>
2026-06-10 09:42:54 +02:00
Cédric Verstraeten
7059503ac1 Expose required secrets to PR description workflow
Replace 'secrets: inherit' with an explicit secrets mapping in the PR description workflow. This exposes TOKEN and the Azure/OpenAI-related secrets (AZURE_OPENAI_API_KEY, OPENAI_MODEL, AZURE_OPENAI_ENDPOINT, AZURE_OPENAI_VERSION) to the job so it can authenticate against the OpenAI/Azure services when generating or updating PR descriptions.
2026-06-10 09:29:10 +02:00
Cédric Verstraeten
7ee79cc063 Use reusable PR description workflow
Replace the inline OpenAI-based PR description job with a reusable workflow call (uug-ai/workflows/.github/workflows/pr-description.yml@main). Removed unused env vars and the checkout/action steps, and pass pr_number and overwrite_description inputs while inheriting repository secrets. This centralizes PR description logic and simplifies the workflow file.
2026-06-10 09:05:51 +02:00
Cédric Verstraeten
9bfbe4ee0f Add RTSP stream health and watchdog logs
Introduce streamHealth instrumentation for Golibrtsp to track per-stream metrics (write durations, gaps, lost packets, decode errors) and emit immediate warnings and periodic summaries to distinguish downstream back-pressure from upstream camera/network stalls. Wire gortsplib Client hooks (OnPacketsLost, OnDecodeError) to the health logger, label streams, and measure WritePacket blocking time around packet writes. Add packetAgeString helper and augment Kerberos ControlAgent logs with stalled counters and last-packet age to provide more context when triggering restarts.
2026-06-10 08:55:39 +02:00
6 changed files with 611 additions and 27 deletions

View File

@@ -2,25 +2,16 @@ name: Autofill PR description
on: pull_request
env:
ORGANIZATION: uugai
PROJECT: ${{ github.event.repository.name }}
PR_NUMBER: ${{ github.event.number }}
jobs:
openai-pr-description:
runs-on: ubuntu-22.04
steps:
- uses: actions/checkout@v4
- name: Autofill PR description if empty using OpenAI
uses: cedricve/azureopenai-pr-description@master
with:
github_token: ${{ secrets.TOKEN }}
openai_api_key: ${{ secrets.OPENAI_API_KEY }}
azure_openai_api_key: ${{ secrets.AZURE_OPENAI_API_KEY }}
azure_openai_endpoint: ${{ secrets.AZURE_OPENAI_ENDPOINT }}
azure_openai_version: ${{ secrets.AZURE_OPENAI_VERSION }}
openai_model: ${{ secrets.OPENAI_MODEL }}
pull_request_url: https://pr${{ env.PR_NUMBER }}.api.kerberos.lol
overwrite_description: true
uses: uug-ai/workflows/.github/workflows/pr-description.yml@main
with:
pr_number: ${{ github.event.number }}
pull_request_url: ""
overwrite_description: true
secrets:
TOKEN: ${{ secrets.TOKEN }}
AZURE_OPENAI_API_KEY: ${{ secrets.AZURE_OPENAI_API_KEY }}
OPENAI_MODEL: ${{ secrets.OPENAI_MODEL }}
AZURE_OPENAI_ENDPOINT: ${{ secrets.AZURE_OPENAI_ENDPOINT }}
AZURE_OPENAI_VERSION: ${{ secrets.AZURE_OPENAI_VERSION }}

168
.github/workflows/release-bump.yml vendored Normal file
View File

@@ -0,0 +1,168 @@
name: Bump release
on:
workflow_dispatch:
inputs:
bump:
description: "Which part of the version to bump"
required: true
default: patch
type: choice
options:
- major
- minor
- patch
permissions:
contents: write
env:
REPO: kerberos/agent
jobs:
# Determine the next version, create the GitHub release and expose the tag.
bump-release:
uses: uug-ai/workflows/.github/workflows/release-bump.yml@main
with:
bump: ${{ github.event.inputs.bump }}
secrets: inherit
# Publish the platform image to the uug-ai GitHub Container Registry
# (ghcr.io/uug-ai/agent-platform).
release:
needs: bump-release
uses: uug-ai/workflows/.github/workflows/release-create.yml@main
with:
organization: uug-ai
project: ${{ github.event.repository.name }}
tag: ${{ needs.bump-release.outputs.tag }}
docker_context: "."
create_gitops_pr: false
runner_matrix: >-
[
{"architecture":"amd64","runner":"ubuntu-24.04"},
{"architecture":"arm64","runner":"ubuntu-24.04-arm"}
]
secrets: inherit
# Everything below mirrors the agent's own release-create.yml pipeline and
# publishes the multi-arch image to the kerberos/agent Docker Hub repo, driven
# by the freshly bumped tag instead of a `release: created` event.
build-amd64:
needs: bump-release
runs-on: ubuntu-24.04
permissions:
contents: write
strategy:
matrix:
architecture: [amd64]
steps:
- name: Login to DockerHub
uses: docker/login-action@v2
with:
username: ${{ secrets.DOCKER_USERNAME }}
password: ${{ secrets.DOCKER_PASSWORD }}
- name: Checkout
uses: actions/checkout@v3
- uses: benjlevesque/short-sha@v2.1
id: short-sha
with:
length: 7
- name: Run Build
run: |
docker build --provenance=false --build-arg VERSION=${{ needs.bump-release.outputs.tag }} -t ${{matrix.architecture}} .
CID=$(docker create ${{matrix.architecture}})
docker cp ${CID}:/home/agent ./output-${{matrix.architecture}}
docker rm ${CID}
- name: Strip binary
run: tar -cf agent-${{matrix.architecture}}.tar -C output-${{matrix.architecture}} . && rm -rf output-${{matrix.architecture}}
- name: Build and push Docker image
run: |
docker tag ${{matrix.architecture}} $REPO-arch:arch-${{matrix.architecture}}-${{ needs.bump-release.outputs.tag }}
docker push $REPO-arch:arch-${{matrix.architecture}}-${{ needs.bump-release.outputs.tag }}
- name: Upload artifact
uses: actions/upload-artifact@v4
with:
name: agent-${{matrix.architecture}}.tar
path: agent-${{matrix.architecture}}.tar
build-arm64:
needs: bump-release
runs-on: ubuntu-24.04-arm
permissions:
contents: write
strategy:
matrix:
architecture: [arm64]
steps:
- name: Login to DockerHub
uses: docker/login-action@v2
with:
username: ${{ secrets.DOCKER_USERNAME }}
password: ${{ secrets.DOCKER_PASSWORD }}
- name: Checkout
uses: actions/checkout@v3
- uses: benjlevesque/short-sha@v2.1
id: short-sha
with:
length: 7
- name: Run Build
run: |
docker build --provenance=false --build-arg VERSION=${{ needs.bump-release.outputs.tag }} -t ${{matrix.architecture}} -f Dockerfile.arm64 .
CID=$(docker create ${{matrix.architecture}})
docker cp ${CID}:/home/agent ./output-${{matrix.architecture}}
docker rm ${CID}
- name: Strip binary
run: tar -cf agent-${{matrix.architecture}}.tar -C output-${{matrix.architecture}} . && rm -rf output-${{matrix.architecture}}
- name: Build and push Docker image
run: |
docker tag ${{matrix.architecture}} $REPO-arch:arch-${{matrix.architecture}}-${{ needs.bump-release.outputs.tag }}
docker push $REPO-arch:arch-${{matrix.architecture}}-${{ needs.bump-release.outputs.tag }}
- name: Upload artifact
uses: actions/upload-artifact@v4
with:
name: agent-${{matrix.architecture}}.tar
path: agent-${{matrix.architecture}}.tar
create-manifest:
runs-on: ubuntu-24.04
needs: [bump-release, build-amd64, build-arm64]
steps:
- name: Login to DockerHub
uses: docker/login-action@v2
with:
username: ${{ secrets.DOCKER_USERNAME }}
password: ${{ secrets.DOCKER_PASSWORD }}
- name: Create and push multi-arch manifest
run: |
docker manifest create $REPO:${{ needs.bump-release.outputs.tag }} \
$REPO-arch:arch-amd64-${{ needs.bump-release.outputs.tag }} \
$REPO-arch:arch-arm64-${{ needs.bump-release.outputs.tag }}
docker manifest push $REPO:${{ needs.bump-release.outputs.tag }}
- name: Create and push latest manifest
run: |
docker manifest create $REPO:latest \
$REPO-arch:arch-amd64-${{ needs.bump-release.outputs.tag }} \
$REPO-arch:arch-arm64-${{ needs.bump-release.outputs.tag }}
docker manifest push $REPO:latest
create-release:
runs-on: ubuntu-24.04
needs: [bump-release, build-amd64, build-arm64]
permissions:
contents: write
steps:
- name: Download all artifacts
uses: actions/download-artifact@v4
- name: Create a release
uses: ncipollo/release-action@v1
with:
latest: true
allowUpdates: true
name: ${{ needs.bump-release.outputs.tag }}
tag: ${{ needs.bump-release.outputs.tag }}
generateReleaseNotes: false
omitBodyDuringUpdate: true
artifacts: "agent-*.tar/agent-*.tar"

View File

@@ -95,6 +95,129 @@ type Golibrtsp struct {
keyframeBufferSize int
keyframeBufferIndex int
keyframeMutex sync.Mutex
// Stream health instrumentation. Used to pinpoint the root cause behind
// "RTP packets lost" + watchdog restarts by separating downstream
// back-pressure from upstream network/camera stalls.
health *streamHealth
streamLabel string
}
// streamHealth instruments the RTSP read path. gortsplib delivers every RTP
// packet on a single read goroutine; queue.WritePacket() is synchronous, so if
// a downstream consumer (recording, muxing, WebRTC) is slow or the process is
// CPU-starved, WritePacket() blocks, the TCP socket is not drained, and the
// camera advances RTP sequence numbers -> "RTP packets lost". This type makes
// the two failure modes distinguishable:
// - large writeMax / writeAvg => downstream back-pressure (our side).
// - large gapMax with fast writes => upstream network / camera stall.
type streamHealth struct {
mu sync.Mutex
windowStart time.Time
lastPacket time.Time
frames int64
writeSum time.Duration
writeMax time.Duration
gapMax time.Duration
lost uint64
decodeErrs int64
}
const (
streamHealthWindow = 10 * time.Second
streamHealthWriteWarn = 150 * time.Millisecond
streamHealthGapWarn = 1500 * time.Millisecond
)
func newStreamHealth() *streamHealth {
now := time.Now()
return &streamHealth{windowStart: now, lastPacket: now}
}
// observePacket records one processed video frame: the wall-clock gap since the
// previous frame (arrival cadence) and how long WritePacket() blocked
// (back-pressure). It emits an immediate warning when either side stalls and a
// periodic summary every streamHealthWindow.
func (h *streamHealth) observePacket(streamType string, writeDur time.Duration) {
if h == nil {
return
}
h.mu.Lock()
defer h.mu.Unlock()
now := time.Now()
var gap time.Duration
if h.frames == 0 {
// First frame: initialize timing to avoid counting RTSP setup time as a stall.
h.windowStart = now
h.lastPacket = now
gap = 0
} else {
gap = now.Sub(h.lastPacket)
h.lastPacket = now
}
h.frames++
h.writeSum += writeDur
if writeDur > h.writeMax {
h.writeMax = writeDur
}
if gap > h.gapMax {
h.gapMax = gap
}
if writeDur >= streamHealthWriteWarn {
log.Log.Warning(fmt.Sprintf(
"capture.golibrtsp.health(%s): WritePacket blocked %dms — downstream back-pressure / CPU starvation",
streamType, writeDur.Milliseconds()))
}
if gap >= streamHealthGapWarn {
log.Log.Warning(fmt.Sprintf(
"capture.golibrtsp.health(%s): %dms since previous frame — upstream network / camera stall",
streamType, gap.Milliseconds()))
}
if now.Sub(h.windowStart) >= streamHealthWindow {
elapsed := now.Sub(h.windowStart).Seconds()
var avgWriteMs float64
if h.frames > 0 {
avgWriteMs = float64(h.writeSum.Milliseconds()) / float64(h.frames)
}
log.Log.Info(fmt.Sprintf(
"capture.golibrtsp.health(%s): %.0fs window — frames=%d (%.1f/s) writeAvg=%.1fms writeMax=%dms gapMax=%dms lost=%d decodeErrs=%d",
streamType, elapsed, h.frames, float64(h.frames)/elapsed, avgWriteMs,
h.writeMax.Milliseconds(), h.gapMax.Milliseconds(), h.lost, h.decodeErrs))
h.windowStart = now
h.frames = 0
h.writeSum = 0
h.writeMax = 0
h.gapMax = 0
h.lost = 0
h.decodeErrs = 0
}
}
// observeLost is invoked by gortsplib when RTP sequence numbers skip. On a TCP
// transport this means the sender (camera) dropped packets because we were not
// reading fast enough, not loss on the wire.
func (h *streamHealth) observeLost(streamType string, lost uint64) {
if h == nil {
return
}
h.mu.Lock()
h.lost += lost
h.mu.Unlock()
log.Log.Warning(fmt.Sprintf(
"capture.golibrtsp.health(%s): %d RTP packet(s) lost — sender-side gap (receiver not draining TCP fast enough)",
streamType, lost))
}
// observeDecodeError is invoked by gortsplib on incomplete/invalid access units,
// which are a downstream symptom of the loss reported by observeLost.
func (h *streamHealth) observeDecodeError(streamType string, err error) {
if h == nil {
return
}
h.mu.Lock()
h.decodeErrs++
h.mu.Unlock()
log.Log.Debug(fmt.Sprintf("capture.golibrtsp.health(%s): decode error: %s", streamType, err.Error()))
}
// fpsTracker holds per-stream state for PTS-based FPS calculation.
@@ -195,9 +318,20 @@ func (g *Golibrtsp) Connect(ctx context.Context, ctxOtel context.Context) (err e
defer span.End()
transport := gortsplib.TransportTCP
g.health = newStreamHealth()
g.Client = gortsplib.Client{
RequestBackChannels: false,
Transport: &transport,
// Route gortsplib's packet-loss / decode-error reporting through our
// structured logger with stream context (replaces its plain stdout
// logging). These hooks are what let us tell whether the camera is
// dropping packets because we can't drain the socket fast enough.
OnPacketsLost: func(lost uint64) {
g.health.observeLost(g.streamLabel, lost)
},
OnDecodeError: func(err error) {
g.health.observeDecodeError(g.streamLabel, err)
},
}
// parse URL
@@ -550,6 +684,12 @@ func compositionOffsetMs(ext dtsExtractor, au [][]byte, pts int64, clockRate int
func (g *Golibrtsp) Start(ctx context.Context, streamType string, queue *packets.Queue, configuration *models.Configuration, communication *models.Communication) (err error) {
log.Log.Debug("capture.golibrtsp.Start(): started")
// Label this client's loss/decode/health logging with the stream type.
g.streamLabel = streamType
if g.health == nil {
g.health = newStreamHealth()
}
// called when a MULAW audio RTP packet arrives
if g.AudioG711Media != nil && g.AudioG711Forma != nil {
g.Client.OnPacketRTP(g.AudioG711Media, g.AudioG711Forma, func(rtppkt *rtp.Packet) {
@@ -813,15 +953,38 @@ func (g *Golibrtsp) Start(ctx context.Context, streamType string, queue *packets
pkt.Data = pkt.Data[4:]
if pkt.IsKeyFrame {
annexbNALUStartCode := func() []byte { return []byte{0x00, 0x00, 0x00, 0x01} }
pkt.Data = append(annexbNALUStartCode(), pkt.Data...)
pkt.Data = append(g.VideoH264Forma.PPS, pkt.Data...)
pkt.Data = append(annexbNALUStartCode(), pkt.Data...)
pkt.Data = append(g.VideoH264Forma.SPS, pkt.Data...)
pkt.Data = append(annexbNALUStartCode(), pkt.Data...)
// Prepend SPS/PPS (when available) in front of every keyframe so the
// access unit is self-contained. Downstream decoders (and the MP4 writer's
// in-band parameter-set recovery) rely on this; a recording whose first
// frame lacks SPS/PPS produces an MP4 with an empty avcC, which makes FFmpeg
// report "non-existing PPS 0 referenced".
//
// Build the payload in a freshly allocated buffer. The previous code
// did append(g.VideoH264Forma.PPS, pkt.Data...): because the SPS/PPS
// slices are sub-slices of the RTP reassembly buffer (spare capacity),
// that append wrote into - and corrupted - the shared parameter-set
// backing arrays, occasionally poisoning the SPS/PPS stored for the
// recording.
startCode := []byte{0x00, 0x00, 0x00, 0x01}
out := make([]byte, 0, len(g.VideoH264Forma.SPS)+len(g.VideoH264Forma.PPS)+len(pkt.Data)+12)
if len(g.VideoH264Forma.SPS) > 0 {
out = append(out, startCode...)
out = append(out, g.VideoH264Forma.SPS...)
}
if len(g.VideoH264Forma.PPS) > 0 {
out = append(out, startCode...)
out = append(out, g.VideoH264Forma.PPS...)
}
out = append(out, startCode...)
out = append(out, pkt.Data...)
pkt.Data = out
}
writeStart := time.Now()
queue.WritePacket(pkt)
// Records WritePacket() blocking time and frame arrival cadence so
// we can tell back-pressure from a network/camera stall.
g.health.observePacket(streamType, time.Since(writeStart))
// This will check if we need to stop the thread,
// because of a reconfiguration.
@@ -983,7 +1146,11 @@ func (g *Golibrtsp) Start(ctx context.Context, streamType string, queue *packets
}
}
writeStart := time.Now()
queue.WritePacket(pkt)
// Records WritePacket() blocking time and frame arrival cadence so
// we can tell back-pressure from a network/camera stall.
g.health.observePacket(streamType, time.Since(writeStart))
// This will check if we need to stop the thread,
// because of a reconfiguration.

View File

@@ -2,6 +2,7 @@ package components
import (
"context"
"fmt"
"os"
"strconv"
"sync/atomic"
@@ -446,6 +447,37 @@ func RunAgent(configDirectory string, configuration *models.Configuration, commu
return status
}
// packetAgeString returns a human readable age (e.g. "12s") since the last
// packet timestamp stored in the given atomic.Value, or "unknown" when no
// packet has been received yet. Used to add context to watchdog restart logs.
func packetAgeString(timer *atomic.Value) string {
if timer == nil {
return "unknown"
}
// atomic.Value panics on Load() if it was never initialized via Store().
var v any
func() {
defer func() {
if recover() != nil {
v = nil
}
}()
v = timer.Load()
}()
last, ok := v.(int64)
if !ok || last == 0 {
return "unknown"
}
age := time.Now().Unix() - last
if age < 0 {
age = 0
}
return strconv.FormatInt(age, 10) + "s"
}
// ControlAgent will check if the camera is still connected, if not it will restart the agent.
// In the other thread we are keeping track of the number of packets received, and particular the keyframe packets.
// Once we are not receiving any packets anymore, we will restart the agent.
@@ -480,7 +512,8 @@ func ControlAgent(communication *models.Communication) {
// After 15 seconds without activity this is thrown..
if occurence == 3 {
log.Log.Info("components.Kerberos.ControlAgent(): Restarting machinery because of blocking mainstream.")
log.Log.Info(fmt.Sprintf("components.Kerberos.ControlAgent(): Restarting machinery because of blocking mainstream. (stalledKeyframeCounter=%d, lastPacket=%s ago, isConfiguring=%t)",
packetsR, packetAgeString(communication.LastPacketTimer), communication.IsConfiguring.IsSet()))
select {
case communication.HandleBootstrap <- "restart":
log.Log.Info("components.Kerberos.ControlAgent(): Restarting machinery because of blocking substream.")
@@ -507,6 +540,8 @@ func ControlAgent(communication *models.Communication) {
// After 15 seconds without activity this is thrown..
if occurenceSub == 3 {
log.Log.Info(fmt.Sprintf("components.Kerberos.ControlAgent(): substream stalled (stalledKeyframeCounter=%d, lastPacket=%s ago, isConfiguring=%t)",
packetsSubR, packetAgeString(communication.LastPacketTimerSub), communication.IsConfiguring.IsSet()))
select {
case communication.HandleBootstrap <- "restart":
log.Log.Info("components.Kerberos.ControlAgent(): Restarting machinery because of blocking substream.")

View File

@@ -32,6 +32,16 @@ const MacEpochOffset uint64 = 2082844800
// resulting in ~3 second fragments (assuming a typical GOP interval).
const FragmentDurationMs = 3000
// SeamGapDivisor controls loop-seam detection. A keyframe is treated as an
// upstream loop/restart seam when it arrives in less than (previous keyframe
// interval / SeamGapDivisor) — i.e. far sooner than the established keyframe
// cadence. Comparing against the *previous* interval (rather than a fixed
// millisecond threshold) makes the check scale automatically with the camera's
// configured GOP size: it works the same whether keyframes are 0.5s, 1s, 2s or
// more apart, and does not misfire on legitimately short-GOP or all-intra
// streams (where every interval is similar, so none looks anomalously short).
const SeamGapDivisor = 2
type MP4 struct {
// FileName is the name of the file
FileName string
@@ -74,6 +84,8 @@ type MP4 struct {
TotalKeyframesWritten int // Total keyframes written to trun boxes
FragmentKeyframeCount int // Keyframes in the current fragment
PendingSampleIsKeyframe bool // Whether the pending video sample is a keyframe
LastKeyframeRawPTS uint64 // Raw PTS of the most recently seen keyframe (across fragments)
LastKeyframeGapMs uint64 // Interval (ms) between the two most recent keyframes; reference cadence for seam detection
}
// NewMP4 creates a new MP4 object.
@@ -298,6 +310,42 @@ func (mp4 *MP4) AddSampleToTrack(trackID uint32, isKeyframe bool, data []byte, p
}
shouldFlush := !mp4.Start || elapsed >= FragmentDurationMs
// Detect an upstream source-loop / restart discontinuity. When a source
// MP4 is looped through virtual-rtsp (ffmpeg `-stream_loop -1 -re`) the
// loop seam emits a fresh IDR much sooner than a normal GOP would. PTS
// keeps growing monotonically, so the timing-only `elapsed` check above
// does not catch it and the seam IDR ends up as a mid-fragment sync
// sample. macOS VideoToolbox rejects such a fragment with
// kVTVideoDecoderBadDataErr (-12909) and MSE players (Video.js /
// Chromium / Firefox) report media corruption, because the inner IDR
// resets frame_num/POC inside what they expect to be a single GOP. This
// is exactly what freezes playback around the seam (e.g. the ~10s mark).
// Force a fragment boundary so the seam IDR starts its own fragment.
//
// The GOP size is configurable per camera, so we do NOT compare against a
// fixed millisecond threshold. Instead we compare this keyframe interval
// to the previous one and only flag a *sudden* shortening: a seam IDR
// arrives in less than (previous interval / SeamGapDivisor). Deriving the
// threshold from the observed cadence keeps detection correct for any
// configured GOP (0.5s, 1s, 2s, ...) and avoids false positives on
// steady short-GOP / all-intra streams (where consecutive intervals are
// similar, so none looks anomalously short). Because the reference is the
// immediately preceding interval, a burst of close keyframes only forces
// a single flush instead of cascading into many tiny fragments.
if trackID == uint32(mp4.VideoTrack) && mp4.Start &&
mp4.LastKeyframeRawPTS > 0 && pts > mp4.LastKeyframeRawPTS {
gap := pts - mp4.LastKeyframeRawPTS
if !shouldFlush && mp4.LastKeyframeGapMs > 0 &&
gap*SeamGapDivisor < mp4.LastKeyframeGapMs {
log.Log.Warning(fmt.Sprintf("mp4.AddSampleToTrack(): forcing fragment flush at unexpectedly close keyframe (interval=%d ms, previous interval=%d ms, fragment elapsed=%d ms) - likely upstream loop/restart discontinuity", gap, mp4.LastKeyframeGapMs, elapsed))
shouldFlush = true
}
mp4.LastKeyframeGapMs = gap
}
if trackID == uint32(mp4.VideoTrack) {
mp4.LastKeyframeRawPTS = pts
}
if shouldFlush {
// Write the previous segment to the file
if mp4.Start {
@@ -571,6 +619,13 @@ func (mp4 *MP4) Close(config *models.Config) {
includePS := true
spsNALUs, ppsNALUs := normalizeH264ParameterSets(mp4.SPSNALUs, mp4.PPSNALUs)
log.Log.Debug("mp4.Close(): AVC parameter sets: SPS=" + formatNaluDebug(spsNALUs) + ", PPS=" + formatNaluDebug(ppsNALUs))
if len(spsNALUs) == 0 || len(ppsNALUs) == 0 {
// An avcC without both SPS and PPS is invalid: downstream FFmpeg-based
// pipelines decoding this file will report "non-existing PPS 0 referenced"
// and fail to extract any frame. Surface it loudly so the capture-side
// parameter-set handling can be diagnosed.
log.Log.Error(fmt.Sprintf("mp4.Close(): incomplete H264 parameter sets (SPS=%d, PPS=%d) - the avcC will be invalid and downstream decoders will report 'non-existing PPS 0 referenced'", len(spsNALUs), len(ppsNALUs)))
}
err := init.Moov.Traks[0].SetAVCDescriptor("avc1", spsNALUs, ppsNALUs, includePS)
if err != nil {
log.Log.Error("mp4.Close(): error setting AVC descriptor: " + err.Error())
@@ -597,6 +652,11 @@ func (mp4 *MP4) Close(config *models.Config) {
includePS := true
vpsNALUs, spsNALUs, ppsNALUs := normalizeH265ParameterSets(mp4.VPSNALUs, mp4.SPSNALUs, mp4.PPSNALUs)
log.Log.Debug("mp4.Close(): HEVC parameter sets: VPS=" + formatNaluDebug(vpsNALUs) + ", SPS=" + formatNaluDebug(spsNALUs) + ", PPS=" + formatNaluDebug(ppsNALUs))
if len(vpsNALUs) == 0 || len(spsNALUs) == 0 || len(ppsNALUs) == 0 {
// An hvcC missing VPS/SPS/PPS is invalid and downstream FFmpeg-based
// pipelines will fail to decode the recording. Surface it loudly.
log.Log.Error(fmt.Sprintf("mp4.Close(): incomplete H265 parameter sets (VPS=%d, SPS=%d, PPS=%d) - the hvcC will be invalid and downstream decoders will fail to process the recording", len(vpsNALUs), len(spsNALUs), len(ppsNALUs)))
}
err := init.Moov.Traks[0].SetHEVCDescriptor("hvc1", vpsNALUs, spsNALUs, ppsNALUs, [][]byte{}, includePS)
if err != nil {
log.Log.Error("mp4.Close(): error setting HEVC descriptor: " + err.Error())

View File

@@ -0,0 +1,163 @@
package video
import (
"os"
"testing"
mp4ff "github.com/Eyevinn/mp4ff/mp4"
"github.com/kerberos-io/agent/machinery/src/models"
)
// runLoopSeamScenario builds a fragmented MP4 that reproduces the loop-seam
// pattern observed in the failing virtual-rtsp recordings (see
// thales_1781183923_3-758_2top_0-0-0-0_-1_30221.mp4): a steady GOP cadence,
// but at the source-MP4 loop boundary an IDR arrives prematurely - far sooner
// than a normal GOP. Without the fix this seam IDR ends up bunched into the
// same fragment as the prior GOP's IDR, producing a mid-fragment sync sample
// that trips macOS VideoToolbox (kVTVideoDecoderBadDataErr / -12909) and
// MSE-based players, freezing playback around the seam (~10s in the original
// file).
//
// gopFrames is the number of frames per GOP, so the same scenario can be
// exercised at different (configurable) camera GOP sizes. The fix derives its
// threshold from the observed keyframe cadence, so the seam must be isolated
// into its own fragment regardless of GOP size.
func runLoopSeamScenario(t *testing.T, gopFrames int) {
t.Helper()
tmpFile, err := os.CreateTemp("", "test_loop_seam_*.mp4")
if err != nil {
t.Fatalf("create temp: %v", err)
}
tmpFile.Close()
defer os.Remove(tmpFile.Name())
sps := []byte{0x67, 0x42, 0xc0, 0x1e, 0xd9, 0x00, 0xa0, 0x47, 0xfe, 0xc8}
pps := []byte{0x68, 0xce, 0x38, 0x80}
mp4Video := NewMP4(tmpFile.Name(), [][]byte{sps}, [][]byte{pps}, nil, 60)
mp4Video.SetWidth(1920)
mp4Video.SetHeight(1080)
v := mp4Video.AddVideoTrack("H264")
mk := func(k bool) []byte {
nt := byte(0x01)
if k {
nt = 0x65
}
f := []byte{0, 0, 0, 1, nt}
for i := 0; i < 200; i++ {
f = append(f, byte(i))
}
return f
}
frameDur := uint64(33)
normalGOPms := uint64(gopFrames) * frameDur
pts := uint64(0)
emitFrame := func(isKey bool) {
// compositionOffset is 0: synthetic stream has no B-frames.
mp4Video.AddSampleToTrack(v, isKey, mk(isKey), pts, 0)
pts += frameDur
}
emitP := func(n int) {
for i := 0; i < n; i++ {
emitFrame(false)
}
}
// emitGOP emits one GOP: a leading keyframe followed by gopFrames-1 P-frames.
emitGOP := func() {
emitFrame(true)
emitP(gopFrames - 1)
}
// Several healthy GOPs to establish the cadence and fill a couple of
// fragments, then the last "good" keyframe followed by a few P-frames so we
// are clearly mid-fragment when the seam arrives.
for g := 0; g < 9; g++ {
emitGOP()
}
emitFrame(true)
seamLead := gopFrames / 5 // seam IDR arrives ~20% into the GOP
if seamLead < 1 {
seamLead = 1
}
emitP(seamLead)
// Loop seam: the source recording restarts, emitting a fresh IDR far sooner
// than the normal GOP. Without the fix this lands as a mid-fragment sync
// sample and freezes playback at the seam.
emitFrame(true)
emitP(gopFrames - 1) // the seam IDR opens a fresh, healthy GOP
// The recording continues with normal GOPs to the end.
for g := 0; g < 10; g++ {
emitGOP()
}
mp4Video.Close(&models.Config{Signing: &models.Signing{PrivateKey: ""}})
f, err := os.Open(tmpFile.Name())
if err != nil {
t.Fatalf("open: %v", err)
}
defer f.Close()
parsed, err := mp4ff.DecodeFile(f)
if err != nil {
t.Fatalf("decode: %v", err)
}
// A healthy fragment only ever contains keyframes spaced ~normalGOPms apart.
// If any fragment contains two keyframes closer than half a normal GOP, the
// premature seam IDR was not isolated and the file will freeze on playback.
maxBunchMs := normalGOPms / 2
fragIdx := 0
for _, seg := range parsed.Segments {
for _, fr := range seg.Fragments {
for _, traf := range fr.Moof.Trafs {
if traf.Tfhd.TrackID != 1 {
continue
}
tfdt := traf.Tfdt.BaseMediaDecodeTime()
offset := uint64(0)
var keys []uint64
for _, trun := range traf.Truns {
for _, s := range trun.Samples {
// sample_depends_on == 2 => "does not depend on others" => IDR/sync.
if (s.Flags>>24)&0x03 == 0x02 {
keys = append(keys, offset)
}
offset += uint64(s.Dur)
}
}
t.Logf("gop=%dframes frag %d tfdt=%d samples_dur=%d keys@%v", gopFrames, fragIdx, tfdt, offset, keys)
for i := 1; i < len(keys); i++ {
gap := keys[i] - keys[i-1]
if gap < maxBunchMs {
t.Errorf("gop=%dframes frag %d (tfdt=%d): two IDRs only %d ms apart in same fragment (< %d) - seam was not isolated",
gopFrames, fragIdx, tfdt, gap, maxBunchMs)
}
}
fragIdx++
}
}
}
}
// TestMP4LoopSeamIsolation exercises the ~1s GOP case (30 frames @ ~33ms),
// matching the original failing recording.
func TestMP4LoopSeamIsolation(t *testing.T) {
runLoopSeamScenario(t, 30)
}
// TestMP4LoopSeamIsolationLargeGOP exercises a larger ~2s GOP (60 frames). The
// GOP size is configurable per camera; this guards against regressing to a
// fixed-millisecond threshold that would only work for ~1s GOPs.
func TestMP4LoopSeamIsolationLargeGOP(t *testing.T) {
runLoopSeamScenario(t, 60)
}
// TestMP4LoopSeamIsolationShortGOP exercises a short ~0.5s GOP (15 frames),
// where a fixed ~1s threshold would misfire on every keyframe. The relative
// detection must only isolate the genuine premature seam IDR.
func TestMP4LoopSeamIsolationShortGOP(t *testing.T) {
runLoopSeamScenario(t, 15)
}