fix(gateway): recover bounded pacing debt
Verify Data Plane / gateway (push) Failing after 3m59s

This commit is contained in:
sechmachine
2026-08-09 20:54:15 +07:00
parent 55afea72a1
commit 122080ab34
8 changed files with 334 additions and 3 deletions
+121
View File
@@ -41,6 +41,83 @@ func TestFairPacerBoundsCatchupAfterHostStall(t *testing.T) {
if next.Before(resumed.Add(-fairPacerMaximumCatchup)) || next.After(resumed.Add(10*time.Millisecond)) {
t.Fatalf("post-stall reservation = %s, want bounded catchup near %s", next, resumed)
}
pacer.mu.Lock()
debt := pacer.flows["one"].debt
pacer.mu.Unlock()
if debt <= 0 || debt > nativeApolloVideoQueueLatency-fairPacerMaximumCatchup {
t.Fatalf("post-stall debt = %s, want bounded valid schedule debt", debt)
}
}
func TestFairPacerRepaysBoundedDebtAfterHostStall(t *testing.T) {
start := time.Date(2026, time.January, 1, 0, 0, 0, 0, time.UTC)
pacer := newFairPacer(8000)
next := make(map[string]time.Time)
deliveries := runSyntheticPacerWithStall(
pacer, start, start.Add(6*time.Second), []string{"one"}, next,
start.Add(time.Second), 100*time.Millisecond,
)
if total := syntheticDeliveryBytes(deliveries); total < 5_990_000 || total > 6_010_000 {
t.Fatalf("post-stall delivery bytes = %d, want nominal throughput after bounded debt repayment", total)
}
pacer.mu.Lock()
remaining := pacer.flows["one"].debt
pacer.mu.Unlock()
if remaining != 0 {
t.Fatalf("post-stall debt = %s after repayment, want zero", remaining)
}
assertSyntheticCap(t, deliveries, 1_000_000)
t.Logf("single-flow debt repaid: bytes=%d remaining=%s", syntheticDeliveryBytes(deliveries), remaining)
}
func TestFairPacerRepaysSimultaneousEightFlowDebtAcrossCapacitySteps(t *testing.T) {
start := time.Date(2026, time.January, 1, 0, 0, 0, 0, time.UTC)
flows := []string{"one", "two", "three", "four", "five", "six", "seven", "eight"}
tests := []struct {
name string
kbps int64
bytesPerSecond int64
minimumBytes int64
}{
{name: "baseline", kbps: 8000, bytesPerSecond: 1_000_000, minimumBytes: 9_980_000},
{name: "quarter", kbps: 6000, bytesPerSecond: 750_000, minimumBytes: 7_480_000},
{name: "half", kbps: 4000, bytesPerSecond: 500_000, minimumBytes: 4_980_000},
}
for _, test := range tests {
pacer := newFairPacer(8000)
next := make(map[string]time.Time, len(flows))
_ = runSyntheticPacer(pacer, start, start.Add(time.Second), flows, next)
resumed := start.Add(1100 * time.Millisecond)
for _, flow := range flows {
next[flow] = pacer.reserveAt(resumed, flow, 1000)
}
assertSyntheticDebt(t, pacer, flows, true)
pacer.setKbps(test.kbps)
deliveries := runSyntheticPacer(pacer, resumed, resumed.Add(10*time.Second), flows, next)
if total := syntheticDeliveryBytes(deliveries); total < test.minimumBytes {
t.Fatalf("%s post-stall delivery bytes = %d, want at least %d", test.name, total, test.minimumBytes)
}
assertSyntheticFairness(t, deliveries, flows)
assertSyntheticCap(t, deliveries, test.bytesPerSecond)
assertSyntheticDebt(t, pacer, flows, false)
t.Logf("%s eight-flow debt repaid: bytes=%d cap=%d", test.name, syntheticDeliveryBytes(deliveries), test.bytesPerSecond*5*105/100)
}
}
func assertSyntheticDebt(t *testing.T, pacer *fairPacer, flows []string, wantDebt bool) {
t.Helper()
pacer.mu.Lock()
defer pacer.mu.Unlock()
for _, flow := range flows {
debt := pacer.flows[flow].debt
if wantDebt && (debt <= 0 || debt > nativeApolloVideoQueueLatency-fairPacerMaximumCatchup) {
t.Fatalf("flow %s active debt = %s, want bounded nonzero debt", flow, debt)
}
if !wantDebt && debt != 0 {
t.Fatalf("flow %s debt = %s after repayment, want zero", flow, debt)
}
}
}
func runSyntheticPacer(pacer *fairPacer, start, end time.Time, flows []string, next map[string]time.Time) []syntheticPacerDelivery {
@@ -67,6 +144,50 @@ func runSyntheticPacer(pacer *fairPacer, start, end time.Time, flows []string, n
}
}
func runSyntheticPacerWithStall(pacer *fairPacer, start, end time.Time, flows []string, next map[string]time.Time, stallAt time.Time, stall time.Duration) []syntheticPacerDelivery {
const packetBytes = 1000
for _, flow := range flows {
if next[flow].IsZero() {
next[flow] = pacer.reserveAt(start, flow, packetBytes)
}
}
now := start
stalled := false
var deliveries []syntheticPacerDelivery
for {
flow := ""
target := end.Add(time.Nanosecond)
for _, candidate := range flows {
if next[candidate].Before(target) {
flow, target = candidate, next[candidate]
}
}
if target.After(end) {
return deliveries
}
if !stalled && !target.Before(stallAt) {
now = stallAt.Add(stall)
stalled = true
}
if now.Before(target) {
now = target
}
if now.After(end) {
return deliveries
}
deliveries = append(deliveries, syntheticPacerDelivery{at: now, flow: flow, bytes: packetBytes})
next[flow] = pacer.reserveAt(now, flow, packetBytes)
}
}
func syntheticDeliveryBytes(deliveries []syntheticPacerDelivery) int64 {
var total int64
for _, delivery := range deliveries {
total += delivery.bytes
}
return total
}
func assertSyntheticFairness(t *testing.T, deliveries []syntheticPacerDelivery, flows []string) {
t.Helper()
counts := make(map[string]int64, len(flows))
+178
View File
@@ -719,6 +719,184 @@ func TestQualificationCarriesCompleteLargeFramesThroughPublicPath(t *testing.T)
}
}
func TestQualificationPacerRepaysThreeMediaLoopStallsThroughPublicPath(t *testing.T) {
profile := qualificationMediaProfiles()[0]
const totalFrames = 480
barriers := []int64{30, 180, 330}
type stall struct {
public chan struct{}
blocked chan uint64
release chan struct{}
}
stalls := make([]stall, len(barriers))
for index := range stalls {
stalls[index] = stall{public: make(chan struct{}), blocked: make(chan uint64, 1), release: make(chan struct{})}
}
ctx, cancel := context.WithTimeout(context.Background(), 15*time.Second)
defer cancel()
var path *qualificationPath
var completed int64
var observationMu sync.Mutex
var maximumQueueDelay time.Duration
observer := func(observation mediaTimingObservation) {
observationMu.Lock()
maximumQueueDelay = max(maximumQueueDelay, observation.QueueDelay)
completed++
current := completed
observationMu.Unlock()
for index, barrier := range barriers {
if current != barrier {
continue
}
select {
case <-stalls[index].public:
case <-ctx.Done():
return
}
baseline := path.session.mediaRecovered.Load()
select {
case stalls[index].blocked <- baseline:
case <-ctx.Done():
return
}
select {
case <-stalls[index].release:
case <-ctx.Done():
}
return
}
}
path = newQualificationPathWithNativeBackendAndObserver(
t, profile, qualificationFramePacerKbps(profile), nil, NewNativeApolloBackend(), observer,
)
defer path.Close()
spacing := time.Second / time.Duration(profile.FPS)
started := time.Now().Add(10 * time.Millisecond)
producerDone := make(chan error, 1)
go func() {
var nextRelease time.Time
for index := int64(0); index < totalFrames; index++ {
release := qualificationBoundedRelease(
started.Add(time.Duration(index)*spacing), nextRelease, time.Now(), spacing,
)
if err := qualificationWaitContext(ctx, release); err != nil {
producerDone <- err
return
}
nextRelease = release.Add(spacing)
if _, err := path.emit(t, qualificationFramePayload(profile, index)); err != nil {
producerDone <- err
return
}
}
producerDone <- nil
}()
var expectedVersePackets uint64
receiverDone := make(chan error, 1)
go func() {
for index := int64(0); index < totalFrames; index++ {
payload, err := path.receivePayload(ctx)
if err != nil {
receiverDone <- err
return
}
expected := qualificationFramePayload(profile, index)
if !bytes.Equal(payload, expected) {
receiverDone <- fmt.Errorf("public payload changed at frame %d", index)
return
}
fragments := (len(payload) + frameV2PayloadSize - 1) / frameV2PayloadSize
expectedVersePackets += uint64(fragments)
for barrierIndex, barrier := range barriers {
if index+1 == barrier {
close(stalls[barrierIndex].public)
}
}
}
receiverDone <- nil
}()
for index := range stalls {
var baseline uint64
select {
case baseline = <-stalls[index].blocked:
case <-ctx.Done():
t.Fatalf("media loop did not block at public frame %d: %v", barriers[index], ctx.Err())
}
target := baseline + 6
for path.session.mediaRecovered.Load() < target {
select {
case <-ctx.Done():
t.Fatalf("stall %d recovered %d, want at least %d: %v", index, path.session.mediaRecovered.Load(), target, ctx.Err())
default:
runtime.Gosched()
}
}
close(stalls[index].release)
}
if err := <-producerDone; err != nil {
t.Fatalf("source fixture: %v", err)
}
if err := <-receiverDone; err != nil {
t.Fatalf("independent client: %v; expected=%d sent=%d source=%d ingress=%d recovered=%d enqueued=%d provider_drops=%d gateway_drops=%d queue_count=%d queue_bytes=%d",
err, path.expectedSourceUDP.Load(), path.fixture.sentPackets.Load(), path.sourceUDP.Load(),
path.session.mediaIngress.Load(), path.session.mediaRecovered.Load(), path.session.mediaEnqueued.Load(),
path.session.Telemetry().MediaDrops, path.server.Metrics().MediaDrops,
path.session.mediaQueueMaximum.Load(), path.session.mediaQueueMaximumBytes.Load())
}
for {
observationMu.Lock()
observed := completed
observationMu.Unlock()
if observed == totalFrames {
break
}
select {
case <-ctx.Done():
t.Fatalf("media observer completed %d frames, want %d: %v", observed, totalFrames, ctx.Err())
default:
runtime.Gosched()
}
}
if source := path.expectedSourceUDP.Load(); source != path.fixture.sentPackets.Load() || source != path.sourceUDP.Load() || source != path.session.mediaIngress.Load() {
t.Fatalf("source accounting expected=%d sent=%d source=%d ingress=%d", source, path.fixture.sentPackets.Load(), path.sourceUDP.Load(), path.session.mediaIngress.Load())
}
metrics := path.server.Metrics()
observationMu.Lock()
completedFrames := completed
queueDelay := maximumQueueDelay
observationMu.Unlock()
if path.backend.setups.Load() != 1 || path.backend.opens.Load() != 1 || completedFrames != totalFrames ||
path.session.mediaRecovered.Load() != totalFrames || path.session.mediaEnqueued.Load() != totalFrames ||
metrics.ProcessingSamples != totalFrames || metrics.MediaPackets != expectedVersePackets ||
path.server.pacer.reservations.Load() != expectedVersePackets ||
path.session.Telemetry().MediaDrops != 0 || path.server.Metrics().MediaDrops != 0 {
t.Fatalf("media accounting setup=%d open=%d completed=%d recovered=%d enqueued=%d processing=%d media=%d pacer=%d provider_drops=%d gateway_drops=%d",
path.backend.setups.Load(), path.backend.opens.Load(), completedFrames,
path.session.mediaRecovered.Load(), path.session.mediaEnqueued.Load(), metrics.ProcessingSamples,
metrics.MediaPackets, path.server.pacer.reservations.Load(), path.session.Telemetry().MediaDrops, metrics.MediaDrops)
}
if path.session.mediaQueueMaximum.Load() > nativeApolloVideoQueuePackets ||
path.session.mediaQueueMaximumBytes.Load() > nativeApolloVideoQueueBytes || queueDelay > nativeApolloVideoQueueLatency {
t.Fatalf("queue bounds count=%d bytes=%d residence=%s", path.session.mediaQueueMaximum.Load(), path.session.mediaQueueMaximumBytes.Load(), queueDelay)
}
t.Logf("three-stall public path: frames=%d source=%d ingress=%d recovered=%d enqueued=%d drops=%d/%d queue=%d/%d residence=%s verse_packets=%d",
totalFrames, path.sourceUDP.Load(), path.session.mediaIngress.Load(), path.session.mediaRecovered.Load(),
path.session.mediaEnqueued.Load(), path.session.Telemetry().MediaDrops, metrics.MediaDrops,
path.session.mediaQueueMaximum.Load(), path.session.mediaQueueMaximumBytes.Load(), queueDelay, metrics.MediaPackets)
path.Close()
select {
case <-path.session.readDone:
case <-time.After(time.Second):
t.Fatal("native media workers did not stop")
}
}
func TestQualificationLateSourceBatchDoesNotCollapseThroughPublicPath(t *testing.T) {
profile := qualificationMediaProfiles()[2]
path := newQualificationPath(t, profile, profile.BitrateKbps)
+5 -1
View File
@@ -1059,6 +1059,10 @@ func newQualificationPathWithImpairment(t *testing.T, profile qualificationMedia
}
func newQualificationPathWithNativeBackend(t *testing.T, profile qualificationMediaProfile, pacerKbps int64, impairment *qualificationImpairmentProfile, native *NativeApolloBackend) *qualificationPath {
return newQualificationPathWithNativeBackendAndObserver(t, profile, pacerKbps, impairment, native, nil)
}
func newQualificationPathWithNativeBackendAndObserver(t *testing.T, profile qualificationMediaProfile, pacerKbps int64, impairment *qualificationImpairmentProfile, native *NativeApolloBackend, observer func(mediaTimingObservation)) *qualificationPath {
t.Helper()
serverTLS, clientTLS := testTLS(t)
fixture := newQualificationApolloFixture(t, serverTLS, clientTLS, "qualification-session", profile)
@@ -1081,7 +1085,7 @@ func newQualificationPathWithNativeBackend(t *testing.T, profile qualificationMe
ListenAddress: "127.0.0.1:0", TLSConfig: serverTLS, GatewayID: authority.GatewayID,
Capabilities: DefaultCapabilities(), ProviderCapabilities: DefaultCapabilities(),
Admission: admission, ProviderStateReporter: &recordingProviderStateReporter{},
Provider: provider, PacerKbps: pacerKbps,
Provider: provider, PacerKbps: pacerKbps, mediaObserver: observer,
})
if err != nil {
t.Fatal(err)
+6
View File
@@ -126,6 +126,7 @@ type fairPacer struct {
type fairPacerFlow struct {
next time.Time
lastSeen time.Time
debt time.Duration
}
const fairPacerMaximumCatchup = 5 * time.Millisecond
@@ -180,9 +181,14 @@ func (p *fairPacer) reserveAt(now time.Time, flow string, bytes int) time.Time {
base = now
} else if lag := now.Sub(base); lag > fairPacerMaximumCatchup {
base = now.Add(-fairPacerMaximumCatchup)
state.debt = min(state.debt+lag-fairPacerMaximumCatchup, nativeApolloVideoQueueLatency-fairPacerMaximumCatchup)
}
numerator := int64(bytes) * int64(len(p.flows)) * int64(time.Second)
delay := time.Duration((numerator + p.bytesPerSecond - 1) / p.bytesPerSecond)
if repayment := min(delay/21, state.debt); repayment > 0 {
delay -= repayment
state.debt -= repayment
}
state.next = base.Add(delay)
p.flows[flow] = state
return state.next