feat(472): sub-split the HLS cold-start startup phase #482

Merged
timothy merged 4 commits from feat/472-startup-subsplit into main 2026-07-19 23:59:23 +02:00
5 changed files with 377 additions and 3 deletions
@@ -131,9 +131,19 @@ public class StartFFmpegSessionHandler : IRequestHandler<StartFFmpegSession, Eit
long startupMs = (long)segments.ProcessStartup.TotalMilliseconds;
long fillMs = (long)segments.SegmentFill.TotalMilliseconds;
long setupMs = Math.Max(0, totalMs - startupMs - fillMs);
// #472 sub-splits the startup work (81% of total, all of the variance) into the ErsatzTV-side
// prep before FFmpeg is launched, FFmpeg's own init (input open+probe and decoder/encoder
// init), and the wait for the playlist once FFmpeg is reporting progress. splitKind says how
// much of that was actually observable for this sample. NOTE these buckets span the worker's
// Run entry rather than the startup stopwatch, so they do NOT sum to startupMs — prep overlaps
// the tail of setup. The log says "spans runEntry" so a reader can't miss it.
// See ColdStartStartupSplit for the full set of caveats.
ColdStartStartupSplit split = segments.StartupSplit;
_logger.LogInformation(
"HLS cold-start channel {Channel} mode {Mode}: total {TotalMs}ms " +
"(setup {SetupMs}ms + startup {ProcessStartupMs}ms + fill {SegmentFillMs}ms), " +
"startup split {SplitKind} spans runEntry (prep {PrepMs}ms + ffmpegInit {FFmpegInitMs}ms " +
"+ firstGop {FirstGopMs}ms), " +
"segments {SegmentsReached}/{InitialSegmentCount}, " +
"deadlineExpired {DeadlineExpired}, subtitleBurnIn {SubtitleBurnIn}, hwaccel {HwAccel}",
request.ChannelNumber,
@@ -142,6 +152,10 @@ public class StartFFmpegSessionHandler : IRequestHandler<StartFFmpegSession, Eit
setupMs,
startupMs,
fillMs,
split.Kind,
(long)split.Prep.TotalMilliseconds,
(long)split.FFmpegInit.TotalMilliseconds,
(long)split.FirstGop.TotalMilliseconds,
segments.SegmentsReached,
segments.InitialSegmentCount,
segments.DeadlineExpired,
@@ -61,6 +61,14 @@ public class HlsSessionWorker : IHlsSessionWorker
// segments cannot exist until this process ran) — volatile for cross-thread visibility.
private volatile string _coldStartFFmpegArguments;
// Stopwatch timestamps of the cold-start milestones used to sub-split the "startup" phase (#472).
// Each is written once on the sequential Run loop and read on the handler thread from
// WaitForPlaylistSegments; long fields cannot be volatile, so access goes through Volatile/
// Interlocked. Zero means "never reached", which ColdStartStartupSplit degrades gracefully on.
private long _coldStartRunTicks;
private long _coldStartProcessLaunchedTicks;
private long _coldStartFirstProgressTicks;
public HlsSessionWorker(
IServiceScopeFactory serviceScopeFactory,
IGraphicsEngine graphicsEngine,
@@ -187,6 +195,10 @@ public class HlsSessionWorker : IHlsSessionWorker
{
_cancellationTokenSource = CancellationTokenSource.CreateLinkedTokenSource(incomingCancellationToken);
// anchor for the cold-start startup sub-split (#472); this runs before any later milestone,
// so every sub-phase derived from it is non-negative by construction
Volatile.Write(ref _coldStartRunTicks, Stopwatch.GetTimestamp());
try
{
_channelNumber = channelNumber;
@@ -314,6 +326,7 @@ public class HlsSessionWorker : IHlsSessionWorker
var sw = Stopwatch.StartNew();
var processStartup = TimeSpan.Zero;
var startupSplit = ColdStartStartupSplit.Unavailable;
var segmentCount = 0;
try
{
@@ -329,6 +342,13 @@ public class HlsSessionWorker : IHlsSessionWorker
_logger.LogDebug("Playlist exists");
processStartup = sw.Elapsed;
// #472: sub-split the phase that #350 measured as 81% of cold-start and all of its variance
startupSplit = ColdStartStartupSplit.FromTimestamps(
Volatile.Read(ref _coldStartRunTicks),
Volatile.Read(ref _coldStartProcessLaunchedTicks),
Volatile.Read(ref _coldStartFirstProgressTicks),
Stopwatch.GetTimestamp());
// start the segment-wait deadline only after the playlist file appears,
// so slow pipeline setup (e.g. h264 profile probing) doesn't consume the budget
DateTimeOffset finish = DateTimeOffset.Now.AddSeconds(8);
@@ -362,7 +382,8 @@ public class HlsSessionWorker : IHlsSessionWorker
segmentCount,
initialSegmentCount,
segmentCount < initialSegmentCount,
ColdStartFeatures.FromFFmpegArguments(_coldStartFFmpegArguments));
ColdStartFeatures.FromFFmpegArguments(_coldStartFFmpegArguments),
startupSplit);
}
finally
{
@@ -576,10 +597,30 @@ public class HlsSessionWorker : IHlsSessionWorker
var progressParser = new FFmpegProgress();
// #472: the first -progress line is the only cold-start milestone FFmpeg gives us
// for free (the pipeline runs -loglevel error -nostats, so stderr stays silent on a
// healthy run). It means the input is open and probed and the decoder/encoder are
// initialized. Record-once, so only the session's first process is measured.
void ParseProgressLine(string line)
{
// the read short-circuits the timestamp call for every line after the first,
// which is every line for the life of the session
if (Volatile.Read(ref _coldStartFirstProgressTicks) == 0)
{
Interlocked.CompareExchange(ref _coldStartFirstProgressTicks, Stopwatch.GetTimestamp(), 0);
}
progressParser.ParseLine(line);
}
// everything before this point is ErsatzTV-side "prep" (playout item resolution,
// pipeline build, graphics engine spawn); FFmpeg's own clock starts here
Interlocked.CompareExchange(ref _coldStartProcessLaunchedTicks, Stopwatch.GetTimestamp(), 0);
CommandResult commandResult = await processWithPipe
.WithWorkingDirectory(_workingDirectory)
.WithStandardErrorPipe(PipeTarget.ToStringBuilder(stdErrBuffer))
.WithStandardOutputPipe(PipeTarget.ToDelegate(progressParser.ParseLine))
.WithStandardOutputPipe(PipeTarget.ToDelegate(ParseProgressLine))
.WithValidation(CommandResultValidation.None)
.ExecuteAsync(linkedCts.Token);
@@ -0,0 +1,171 @@
using System.Diagnostics;
using ErsatzTV.Core.FFmpeg;
using NUnit.Framework;
using Shouldly;
namespace ErsatzTV.Core.Tests.FFmpeg;
[TestFixture]
public class ColdStartStartupSplitTests
{
// milestones are Stopwatch.GetTimestamp() values; build them from a base + millisecond offsets
private const long Base = 1_000_000_000;
private static long At(double milliseconds) =>
Base + (long)(milliseconds / 1000.0 * Stopwatch.Frequency);
[Test]
public void Should_Split_Three_Ways_When_All_Milestones_Present()
{
ColdStartStartupSplit split = ColdStartStartupSplit.FromTimestamps(
At(0),
At(150),
At(1200),
At(1600));
split.Kind.ShouldBe(ColdStartStartupSplitKind.ThreeWay);
split.Prep.TotalMilliseconds.ShouldBe(150, 1);
split.FFmpegInit.TotalMilliseconds.ShouldBe(1050, 1);
split.FirstGop.TotalMilliseconds.ShouldBe(400, 1);
}
[Test]
public void Sub_Phases_Should_Sum_To_Run_Entry_Through_Playlist()
{
// deliberately NOT "should sum to startup": the buckets span the worker's Run entry, which
// begins before the request thread's startup stopwatch, so prep overlaps the tail of setup
ColdStartStartupSplit split = ColdStartStartupSplit.FromTimestamps(
At(0),
At(150),
At(1200),
At(1600));
(split.Prep + split.FFmpegInit + split.FirstGop).TotalMilliseconds.ShouldBe(1600, 1);
}
[Test]
public void Should_Be_Unavailable_When_The_Playlist_Predates_The_Process_Launch()
{
// a stale live.m3u8 survives when the handler's pre-session folder wipe fails (EmptyFolder
// swallows the failure into a warning). Every bucket would be meaningless, so report nothing
// rather than a plausible-looking sample with a prep that exceeds the whole measured phase
ColdStartStartupSplit split = ColdStartStartupSplit.FromTimestamps(
At(0),
At(1600),
0,
At(150));
split.ShouldBe(ColdStartStartupSplit.Unavailable);
}
[Test]
public void Stale_Playlist_Guard_Should_Take_Precedence_Over_The_Progress_Branches()
{
// without the guard, this input would be classified TwoWayLateProgress; the guard must be
// evaluated first. (It can never preempt a ThreeWay: that requires processLaunched <=
// playlistExists, which is exactly the negation of the guard condition.)
ColdStartStartupSplit split = ColdStartStartupSplit.FromTimestamps(
At(0),
At(1600),
At(1700),
At(150));
split.ShouldBe(ColdStartStartupSplit.Unavailable);
}
[Test]
public void Should_Fall_Back_To_Two_Way_Split_When_Progress_Predates_The_Process_Launch()
{
// a progress timestamp older than the launch cannot belong to this process
ColdStartStartupSplit split = ColdStartStartupSplit.FromTimestamps(
At(100),
At(150),
At(120),
At(1600));
split.Kind.ShouldBe(ColdStartStartupSplitKind.TwoWay);
split.FFmpegInit.TotalMilliseconds.ShouldBe(1450, 1);
split.FirstGop.ShouldBe(TimeSpan.Zero);
}
[Test]
public void Should_Stay_Three_Way_When_Progress_Coincides_With_A_Boundary()
{
ColdStartStartupSplit atLaunch = ColdStartStartupSplit.FromTimestamps(At(0), At(150), At(150), At(1600));
atLaunch.Kind.ShouldBe(ColdStartStartupSplitKind.ThreeWay);
atLaunch.FFmpegInit.ShouldBe(TimeSpan.Zero);
atLaunch.FirstGop.TotalMilliseconds.ShouldBe(1450, 1);
ColdStartStartupSplit atPlaylist = ColdStartStartupSplit.FromTimestamps(At(0), At(150), At(1600), At(1600));
atPlaylist.Kind.ShouldBe(ColdStartStartupSplitKind.ThreeWay);
atPlaylist.FFmpegInit.TotalMilliseconds.ShouldBe(1450, 1);
atPlaylist.FirstGop.ShouldBe(TimeSpan.Zero);
}
[Test]
public void Should_Fall_Back_To_Two_Way_Split_When_FFmpeg_Never_Reported_Progress()
{
// no -progress output before the playlist appeared: ffmpegInit must absorb the remainder
// rather than the split inventing a firstGop boundary that was never observed
ColdStartStartupSplit split = ColdStartStartupSplit.FromTimestamps(
At(0),
At(150),
0,
At(1600));
split.Kind.ShouldBe(ColdStartStartupSplitKind.TwoWay);
split.Prep.TotalMilliseconds.ShouldBe(150, 1);
split.FFmpegInit.TotalMilliseconds.ShouldBe(1450, 1);
split.FirstGop.ShouldBe(TimeSpan.Zero);
}
[Test]
public void Should_Report_Late_Progress_Distinctly_When_Progress_Arrived_After_The_Playlist()
{
// the playlist is observed on the request thread while progress is recorded on the worker
// thread; a progress milestone outside the phase must not produce a negative bucket
ColdStartStartupSplit split = ColdStartStartupSplit.FromTimestamps(
At(0),
At(150),
At(1800),
At(1600));
split.Kind.ShouldBe(ColdStartStartupSplitKind.TwoWayLateProgress);
split.FFmpegInit.TotalMilliseconds.ShouldBe(1450, 1);
split.FirstGop.ShouldBe(TimeSpan.Zero);
}
[TestCase(0L, 150L, 1200L, 1600L, TestName = "Run never started")]
[TestCase(100L, 0L, 0L, 1600L, TestName = "Process never launched")]
[TestCase(100L, 150L, 1200L, 0L, TestName = "Playlist never appeared")]
public void Should_Be_Unavailable_When_A_Required_Milestone_Is_Missing(
long runStarted,
long processLaunched,
long firstProgress,
long playlistExists)
{
ColdStartStartupSplit split = ColdStartStartupSplit.FromTimestamps(
runStarted == 0 ? 0 : At(runStarted),
processLaunched == 0 ? 0 : At(processLaunched),
firstProgress == 0 ? 0 : At(firstProgress),
playlistExists == 0 ? 0 : At(playlistExists));
split.ShouldBe(ColdStartStartupSplit.Unavailable);
}
[Test]
public void Should_Clamp_Rather_Than_Report_A_Negative_Prep()
{
// defensive: launch cannot precede Run entry, but telemetry must never show a negative
ColdStartStartupSplit split = ColdStartStartupSplit.FromTimestamps(
At(500),
At(150),
At(1200),
At(1600));
split.Prep.ShouldBe(TimeSpan.Zero);
split.Kind.ShouldBe(ColdStartStartupSplitKind.ThreeWay);
split.FFmpegInit.TotalMilliseconds.ShouldBe(1050, 1);
split.FirstGop.TotalMilliseconds.ShouldBe(400, 1);
}
}
@@ -0,0 +1,146 @@
using System.Diagnostics;
namespace ErsatzTV.Core.FFmpeg;
/// <summary>
/// How finely a cold-start's startup work could be broken down (#472).
/// </summary>
public enum ColdStartStartupSplitKind
{
/// <summary>
/// No split available: the FFmpeg process was never launched, the playlist never appeared, or the
/// playlist was observed before FFmpeg was launched (a stale playlist left behind because the
/// pre-session transcode-folder wipe failed — it logs a warning and continues).
/// </summary>
Unavailable = 0,
/// <summary>
/// Two-way split: <c>prep</c> + <c>ffmpegInit</c>, because FFmpeg emitted no progress output at all
/// before the playlist appeared. <c>ffmpegInit</c> therefore runs to the playlist.
/// </summary>
TwoWay = 1,
/// <summary>
/// Two-way split, distinguished because FFmpeg <em>did</em> report progress but only after the
/// playlist was observed. Same buckets as <see cref="TwoWay"/>; kept separate because it means the
/// playlist appeared before the first progress report rather than FFmpeg being silent, which is a
/// different story about the pipeline (and is also what the 100ms playlist poll can manufacture).
/// </summary>
TwoWayLateProgress = 2,
/// <summary>Three-way split: <c>prep</c> + <c>ffmpegInit</c> + <c>firstGop</c>.</summary>
ThreeWay = 3
}
/// <summary>
/// Sub-split of the HLS cold-start startup work (#472), which #350's measurement showed to be 81% of
/// tune-in latency and to carry 100% of its variance while remaining a single opaque bucket.
/// <para>
/// <see cref="Prep"/> is ErsatzTV-side work before FFmpeg exists: playout-item resolution, pipeline
/// build, graphics-engine spawn. <see cref="FFmpegInit"/> is FFmpeg from launch until it first reports
/// progress — input open + probe (the NFS hypothesis) plus decoder/encoder init (the VAAPI-contention
/// hypothesis). <see cref="FirstGop"/> is from that first progress report until <c>live.m3u8</c> exists.
/// </para>
/// <para>
/// <b>These buckets span the session worker's <c>Run</c> entry to the playlist appearing, which is NOT
/// exactly the logged <c>startup</c> phase</b>: the worker is launched fire-and-forget slightly before
/// the request thread starts the <c>startup</c> stopwatch, so <see cref="Prep"/> overlaps the tail of
/// the logged <c>setup</c> bucket (in practice one config read). Do not expect
/// <c>prep + ffmpegInit + firstGop</c> to equal <c>startup</c> — it is a superset by that overlap.
/// </para>
/// <para>
/// Because the pipeline runs <c>-loglevel error -nostats -hide_banner</c>, a healthy FFmpeg writes
/// nothing to stderr, so input-open and encoder-init cannot be separated from each other without
/// changing the FFmpeg command — which this instrumentation deliberately does not do. The
/// <c>-progress</c> stream on stdout is therefore the only zero-cost milestone available, and
/// <see cref="FFmpegInit"/> necessarily lumps those two candidates together. #472 accepts this: a
/// large <see cref="Prep"/> vs a large <see cref="FFmpegInit"/> is itself the first discrimination,
/// and it is honest about what it cannot yet see.
/// </para>
/// <para>
/// Three further caveats when reading these numbers. The playlist is detected by a 100ms poll, so its
/// timestamp is up to 100ms late and that error lands entirely in <see cref="FirstGop"/> — the
/// smallest bucket — and can also flip a sample between <see cref="ColdStartStartupSplitKind.ThreeWay"/>
/// and <see cref="ColdStartStartupSplitKind.TwoWayLateProgress"/>. And if the session's first FFmpeg
/// process fails and a second one produces the playlist, <see cref="FFmpegInit"/> spans the first
/// process's whole lifetime plus the retry while still being labelled as one process's init. And the
/// stale-playlist guard below is best-effort rather than a proof: if the folder wipe failed, whether
/// the stale playlist is observed before or after the launch milestone is a scheduling race, so an
/// unlucky sample could still slip through as an implausibly fast one (most often
/// <see cref="ColdStartStartupSplitKind.TwoWay"/>, since FFmpeg has usually not reported progress
/// that early).
/// </para>
/// </summary>
public readonly record struct ColdStartStartupSplit(
TimeSpan Prep,
TimeSpan FFmpegInit,
TimeSpan FirstGop,
ColdStartStartupSplitKind Kind)
{
public static readonly ColdStartStartupSplit Unavailable =
new(TimeSpan.Zero, TimeSpan.Zero, TimeSpan.Zero, ColdStartStartupSplitKind.Unavailable);
/// <summary>
/// Builds the split from four <see cref="Stopwatch.GetTimestamp"/> milestones; <c>0</c> means the
/// milestone never happened. Milestones are recorded on different threads (the session worker
/// records the launch and progress ones; the request thread observes the playlist), so ordering is
/// validated rather than assumed: any out-of-order or missing milestone degrades the result to a
/// coarser <see cref="ColdStartStartupSplitKind"/> instead of producing a negative or invented bucket.
/// </summary>
public static ColdStartStartupSplit FromTimestamps(
long runStarted,
long processLaunched,
long firstProgress,
long playlistExists)
{
if (runStarted <= 0 || processLaunched <= 0 || playlistExists <= 0)
{
return Unavailable;
}
if (processLaunched > playlistExists)
{
// the playlist was observed before FFmpeg was even launched, so it is a stale file: the
// handler wipes the transcode folder before starting the session, but that wipe swallows
// its failures into a warning (LocalFileSystem.EmptyFolder) and continues.
// Every bucket would be meaningless; report nothing rather than a plausible-looking sample
return Unavailable;
}
// the worker's Run entry strictly precedes every later milestone; clamp anyway so a clock
// oddity can never surface as a negative duration in telemetry
TimeSpan prep = Elapsed(runStarted, processLaunched);
if (firstProgress <= 0 || firstProgress < processLaunched)
{
// FFmpeg reported no usable progress before the playlist appeared: fall back to the two-way
// split #472 explicitly accepts, rather than inventing a boundary that was never observed
return new ColdStartStartupSplit(
prep,
Elapsed(processLaunched, playlistExists),
TimeSpan.Zero,
ColdStartStartupSplitKind.TwoWay);
}
if (firstProgress > playlistExists)
{
return new ColdStartStartupSplit(
prep,
Elapsed(processLaunched, playlistExists),
TimeSpan.Zero,
ColdStartStartupSplitKind.TwoWayLateProgress);
}
return new ColdStartStartupSplit(
prep,
Elapsed(processLaunched, firstProgress),
Elapsed(firstProgress, playlistExists),
ColdStartStartupSplitKind.ThreeWay);
}
private static TimeSpan Elapsed(long from, long to)
{
TimeSpan elapsed = Stopwatch.GetElapsedTime(from, to);
return elapsed < TimeSpan.Zero ? TimeSpan.Zero : elapsed;
}
}
@@ -6,6 +6,7 @@ namespace ErsatzTV.Core.FFmpeg;
/// FFmpeg process spawn + probe + libass/encoder init + first GOP, since the wait begins right
/// after the fire-and-forget worker is launched); <see cref="SegmentFill"/> is Phase B (playlist
/// exists -&gt; the requested number of segments are present, or the 8s deadline).
/// <see cref="StartupSplit"/> breaks Phase A down further (#472).
/// </summary>
public readonly record struct PlaylistSegmentsResult(
TimeSpan ProcessStartup,
@@ -13,4 +14,5 @@ public readonly record struct PlaylistSegmentsResult(
int SegmentsReached,
int InitialSegmentCount,
bool DeadlineExpired,
ColdStartFeatures Features);
ColdStartFeatures Features,
ColdStartStartupSplit StartupSplit);