feat(server): add hls playlist hint (#29150)

* add playlist hint

* update api

* unused import

* update crf in tests

mobile/openapi/lib/api/assets_api.dart

@@ -982,7 +982,9 @@ class AssetsApi {
   /// * [String] key:
   ///
   /// * [String] slug:
-  Future<Response> getMediaPlaylistWithHttpInfo(String id, String sessionId, int variantIndex, { String? key, String? slug, Future<void>? abortTrigger, }) async {
+  ///
+  /// * [num] xImmichHlsPos:
+  Future<Response> getMediaPlaylistWithHttpInfo(String id, String sessionId, int variantIndex, { String? key, String? slug, num? xImmichHlsPos, Future<void>? abortTrigger, }) async {
     // ignore: prefer_const_declarations
     final apiPath = r'/assets/{id}/video/stream/{sessionId}/{variantIndex}/playlist.m3u8'
       .replaceAll('{id}', id)
@@ -1003,6 +1005,10 @@ class AssetsApi {
       queryParams.addAll(_queryParams('', 'slug', slug));
     }
 
+    if (xImmichHlsPos != null) {
+      headerParams[r'x-immich-hls-pos'] = parameterToString(xImmichHlsPos);
+    }
+
     const contentTypes = <String>[];
 
 
@@ -1033,8 +1039,10 @@ class AssetsApi {
   /// * [String] key:
   ///
   /// * [String] slug:
-  Future<String?> getMediaPlaylist(String id, String sessionId, int variantIndex, { String? key, String? slug, Future<void>? abortTrigger, }) async {
+  ///
-    final response = await getMediaPlaylistWithHttpInfo(id, sessionId, variantIndex, key: key, slug: slug, abortTrigger: abortTrigger,);
+  /// * [num] xImmichHlsPos:
+  Future<String?> getMediaPlaylist(String id, String sessionId, int variantIndex, { String? key, String? slug, num? xImmichHlsPos, Future<void>? abortTrigger, }) async {
+    final response = await getMediaPlaylistWithHttpInfo(id, sessionId, variantIndex, key: key, slug: slug, xImmichHlsPos: xImmichHlsPos, abortTrigger: abortTrigger,);
     if (response.statusCode >= HttpStatus.badRequest) {
       throw ApiException(response.statusCode, await _decodeBodyBytes(response));
     }

open-api/immich-openapi-specs.json

@@ -4924,6 +4924,15 @@
               "maximum": 9007199254740991,
               "type": "integer"
             }
+          },
+          {
+            "name": "x-immich-hls-pos",
+            "required": false,
+            "in": "header",
+            "schema": {
+              "minimum": 0,
+              "type": "number"
+            }
           }
         ],
         "responses": {

packages/sdk/src/fetch-client.ts

@@ -4452,12 +4452,13 @@ export function endSession({ id, key, sessionId, slug }: {
 /**
  * Get HLS media playlist
  */
-export function getMediaPlaylist({ id, key, sessionId, slug, variantIndex }: {
+export function getMediaPlaylist({ id, key, sessionId, slug, variantIndex, xImmichHlsPos }: {
     id: string;
     key?: string;
     sessionId: string;
     slug?: string;
     variantIndex: number;
+    xImmichHlsPos?: number;
 }, opts?: Oazapfts.RequestOpts) {
     return oazapfts.ok(oazapfts.fetchBlob<{
         status: 200;
@@ -4466,7 +4467,10 @@ export function getMediaPlaylist({ id, key, sessionId, slug, variantIndex }: {
         key,
         slug
     }))}`, {
-        ...opts
+        ...opts,
+        headers: oazapfts.mergeHeaders(opts?.headers, {
+            "x-immich-hls-pos": xImmichHlsPos
+        })
     }));
 }
 /**

server/src/constants.ts

@@ -223,6 +223,12 @@ export const SUPPORTED_HWA_CODECS: Record<TranscodeHardwareAcceleration, VideoCo
 export const HLS_BACKPRESSURE_PAUSE_SEGMENTS = 30;
 export const HLS_BACKPRESSURE_RESUME_SEGMENTS = 15;
 export const HLS_CLEANUP_INTERVAL_MS = 60 * 1000;
+export const HLS_CRF: Record<VideoCodec, number> = {
+  [VideoCodec.H264]: 23,
+  [VideoCodec.Hevc]: 28,
+  [VideoCodec.Vp9]: 31,
+  [VideoCodec.Av1]: 35,
+};
 export const HLS_INACTIVITY_TIMEOUT_MS = 5 * 60 * 1000;
 export const HLS_LEASE_DURATION_MS = 30 * 60 * 1000;
 export const HLS_PLAYLIST_CONTENT_TYPE = 'application/vnd.apple.mpegurl';

server/src/controllers/video-stream.controller.ts

@@ -6,6 +6,7 @@ import { HLS_PLAYLIST_CONTENT_TYPE } from 'src/constants';
 import { Endpoint, HistoryBuilder } from 'src/decorators';
 import { AuthDto } from 'src/dtos/auth.dto';
 import {
+  HlsPlaylistHeaderDto,
   HlsSegmentHeaderDto,
   HlsSegmentParamDto,
   HlsSessionParamDto,
@@ -50,8 +51,17 @@ export class VideoStreamController {
     description: 'Returns an HLS media playlist for one variant of the streaming session.',
     history: new HistoryBuilder().added('v3').alpha('v3'),
   })
-  getMediaPlaylist(@Auth() auth: AuthDto, @Param() { id, sessionId }: HlsVariantParamDto) {
+  getMediaPlaylist(
-    return this.service.getMediaPlaylist(auth, id, sessionId);
+    @Auth() auth: AuthDto,
+    @Param() { id, sessionId, variantIndex }: HlsVariantParamDto,
+    @Headers() headers: HlsPlaylistHeaderDto,
+  ) {
+    try {
+      headers = HlsPlaylistHeaderDto.create(headers);
+    } catch (error) {
+      throw new ZodValidationException(error);
+    }
+    return this.service.getMediaPlaylist(auth, id, sessionId, variantIndex, headers[ImmichHeader.HlsPosition]);
   }
 
   @Get(':id/video/stream/:sessionId/:variantIndex/:filename')

server/src/dtos/streaming.dto.ts

@@ -32,3 +32,11 @@ const HlsSegmentHeaderSchema = z.object({
 });
 
 export class HlsSegmentHeaderDto extends createZodDto(HlsSegmentHeaderSchema) {}
+
+const HlsPlaylistHeaderSchema = z.object({
+  // Lets the client hint at which segment will be loaded after the playlist.
+  // A position rather than a segment index since indices aren't comparable across variants.
+  [ImmichHeader.HlsPosition]: z.coerce.number().min(0).optional(),
+});
+
+export class HlsPlaylistHeaderDto extends createZodDto(HlsPlaylistHeaderSchema) {}

server/src/enum.ts

@@ -25,6 +25,7 @@ export enum ImmichHeader {
   Checksum = 'x-immich-checksum',
   CorrelationId = 'X-Correlation-ID',
   HlsInitSegment = 'x-immich-hls-msn',
+  HlsPosition = 'x-immich-hls-pos',
 }
 
 export enum ImmichQuery {

server/src/services/hls.service.spec.ts

@@ -8,6 +8,7 @@ import { newTestService, ServiceMocks } from 'test/utils';
 // EXTINF values come from FFmpeg's playlist to enforce an exact match
 const eiffelExpectedMediaPlaylist = `#EXTM3U
 #EXT-X-VERSION:7
+#EXT-X-INDEPENDENT-SEGMENTS
 #EXT-X-TARGETDURATION:2
 #EXT-X-MEDIA-SEQUENCE:0
 #EXT-X-PLAYLIST-TYPE:VOD
@@ -41,6 +42,7 @@ seg_11.m4s
 
 const waterfallExpectedMediaPlaylist = `#EXTM3U
 #EXT-X-VERSION:7
+#EXT-X-INDEPENDENT-SEGMENTS
 #EXT-X-TARGETDURATION:2
 #EXT-X-MEDIA-SEQUENCE:0
 #EXT-X-PLAYLIST-TYPE:VOD
@@ -62,6 +64,7 @@ seg_5.m4s
 
 const trainExpectedMediaPlaylist = `#EXTM3U
 #EXT-X-VERSION:7
+#EXT-X-INDEPENDENT-SEGMENTS
 #EXT-X-TARGETDURATION:2
 #EXT-X-MEDIA-SEQUENCE:0
 #EXT-X-PLAYLIST-TYPE:VOD
@@ -95,6 +98,7 @@ const sessionId = '00000000-0000-0000-0000-000000000000';
 
 const eiffelExpectedMasterDisabled = `#EXTM3U
 #EXT-X-VERSION:7
+#EXT-X-INDEPENDENT-SEGMENTS
 #EXT-X-STREAM-INF:BANDWIDTH=1000000,RESOLUTION=480x852,CODECS="av01.0.04M.08,mp4a.40.2",VIDEO-RANGE=SDR,FRAME-RATE=24.910
 ${sessionId}/0/playlist.m3u8
 #EXT-X-STREAM-INF:BANDWIDTH=1200000,RESOLUTION=480x852,CODECS="hvc1.1.6.L90.B0,mp4a.40.2",VIDEO-RANGE=SDR,FRAME-RATE=24.910
@@ -117,6 +121,7 @@ ${sessionId}/8/playlist.m3u8
 
 const eiffelExpectedMasterRkmpp = `#EXTM3U
 #EXT-X-VERSION:7
+#EXT-X-INDEPENDENT-SEGMENTS
 #EXT-X-STREAM-INF:BANDWIDTH=1200000,RESOLUTION=480x852,CODECS="hvc1.1.6.L90.B0,mp4a.40.2",VIDEO-RANGE=SDR,FRAME-RATE=24.910
 ${sessionId}/1/playlist.m3u8
 #EXT-X-STREAM-INF:BANDWIDTH=2500000,RESOLUTION=480x852,CODECS="avc1.64001e,mp4a.40.2",VIDEO-RANGE=SDR,FRAME-RATE=24.910
@@ -133,6 +138,7 @@ ${sessionId}/8/playlist.m3u8
 
 const waterfallExpectedMasterDisabled = `#EXTM3U
 #EXT-X-VERSION:7
+#EXT-X-INDEPENDENT-SEGMENTS
 #EXT-X-STREAM-INF:BANDWIDTH=1000000,RESOLUTION=480x852,CODECS="av01.0.04M.08,mp4a.40.2",VIDEO-RANGE=SDR,FRAME-RATE=29.830
 ${sessionId}/0/playlist.m3u8
 #EXT-X-STREAM-INF:BANDWIDTH=1200000,RESOLUTION=480x852,CODECS="hvc1.1.6.L90.B0,mp4a.40.2",VIDEO-RANGE=SDR,FRAME-RATE=29.830
@@ -218,12 +224,58 @@ describe(HlsService.name, () => {
     it.each(fixtures)('matches FFmpeg for $data.originalPath', async ({ data, playlist }) => {
       mocks.access.asset.checkOwnerAccess.mockResolvedValue(new Set([assetId]));
       mocks.videoStream.getForMediaPlaylist.mockResolvedValue(data);
-      await expect(sut.getMediaPlaylist(auth, assetId, sessionId)).resolves.toBe(playlist);
+      await expect(sut.getMediaPlaylist(auth, assetId, sessionId, 0)).resolves.toBe(playlist);
     });
 
     it('throws NotFoundException when the session/asset cannot be loaded', async () => {
       mocks.access.asset.checkOwnerAccess.mockResolvedValue(new Set([assetId]));
-      await expect(sut.getMediaPlaylist(auth, assetId, sessionId)).rejects.toBeInstanceOf(NotFoundException);
+      await expect(sut.getMediaPlaylist(auth, assetId, sessionId, 0)).rejects.toBeInstanceOf(NotFoundException);
+    });
+
+    it('prewarms transcoding at the segment containing the hinted position', async () => {
+      mocks.access.asset.checkOwnerAccess.mockResolvedValue(new Set([assetId]));
+      mocks.videoStream.getForMediaPlaylist.mockResolvedValue(eiffelTower);
+      await sut.getMediaPlaylist(auth, assetId, sessionId, 1, 10.5);
+      expect(mocks.websocket.serverSend).toHaveBeenCalledWith('HlsSegmentRequest', {
+        sessionId,
+        assetId,
+        variantIndex: 1,
+        segmentIndex: 5,
+      });
+    });
+
+    it('prewarms from the last requested segment when no hint is given', async () => {
+      mocks.access.asset.checkOwnerAccess.mockResolvedValue(new Set([assetId]));
+      mocks.videoStream.getSession.mockResolvedValue({ id: sessionId, assetId } as never);
+      mocks.storage.checkFileExists.mockResolvedValue(true);
+      await sut.getSegment(auth, assetId, sessionId, 0, 'seg_5.m4s');
+
+      mocks.videoStream.getForMediaPlaylist.mockResolvedValue(eiffelTower);
+      await sut.getMediaPlaylist(auth, assetId, sessionId, 1);
+      expect(mocks.websocket.serverSend).toHaveBeenCalledWith('HlsSegmentRequest', {
+        sessionId,
+        assetId,
+        variantIndex: 1,
+        segmentIndex: 6,
+      });
+    });
+
+    it('does not prewarm without a hint or prior segment traffic', async () => {
+      mocks.access.asset.checkOwnerAccess.mockResolvedValue(new Set([assetId]));
+      mocks.videoStream.getForMediaPlaylist.mockResolvedValue(eiffelTower);
+      await sut.getMediaPlaylist(auth, assetId, sessionId, 1);
+      expect(mocks.websocket.serverSend).not.toHaveBeenCalled();
+    });
+
+    it('does not prewarm the variant the session is already playing', async () => {
+      mocks.access.asset.checkOwnerAccess.mockResolvedValue(new Set([assetId]));
+      mocks.videoStream.getSession.mockResolvedValue({ id: sessionId, assetId } as never);
+      mocks.storage.checkFileExists.mockResolvedValue(true);
+      await sut.getSegment(auth, assetId, sessionId, 1, 'seg_5.m4s');
+
+      mocks.videoStream.getForMediaPlaylist.mockResolvedValue(eiffelTower);
+      await sut.getMediaPlaylist(auth, assetId, sessionId, 1, 12.5);
+      expect(mocks.websocket.serverSend).not.toHaveBeenCalledWith('HlsSegmentRequest', expect.anything());
     });
   });
 
@@ -314,7 +366,7 @@ describe(HlsService.name, () => {
       );
     });
 
-    it('uses the target segment for init.mp4 when provided', async () => {
+    it('uses the initSegment hint for init.mp4', async () => {
       await sut.getSegment(auth, assetId, sessionId, variantIndex, 'init.mp4', 7);
       expect(mocks.websocket.serverSend).toHaveBeenCalledWith('HlsHeartbeat', {
         sessionId,
@@ -323,18 +375,18 @@ describe(HlsService.name, () => {
       });
     });
 
-    it('prefers the target segment over the lastRequested + 1 fallback', async () => {
+    it('prefers the initSegment hint over the lastRequested + 1 fallback', async () => {
       await sut.getSegment(auth, assetId, sessionId, variantIndex, 'seg_5.m4s'); // fallback would be 6
       mocks.websocket.serverSend.mockClear();
-      await sut.getSegment(auth, assetId, sessionId, variantIndex, 'init.mp4', 12);
+      await sut.getSegment(auth, assetId, sessionId, variantIndex, 'init.mp4', 10);
       expect(mocks.websocket.serverSend).toHaveBeenCalledWith('HlsHeartbeat', {
         sessionId,
         variantIndex,
-        segmentIndex: 12,
+        segmentIndex: 10,
       });
     });
 
-    it('ignores the target segment for media segment requests (the filename wins)', async () => {
+    it('ignores the initSegment hint for media segment requests (the filename wins)', async () => {
       await sut.getSegment(auth, assetId, sessionId, variantIndex, 'seg_5.m4s', 99);
       expect(mocks.websocket.serverSend).toHaveBeenCalledWith('HlsHeartbeat', {
         sessionId,

server/src/services/hls.service.ts

@@ -21,6 +21,7 @@ import { ImmichFileResponse } from 'src/utils/file';
 import { getOutputSize } from 'src/utils/media';
 
 type AssetWithStreamInfo = { videoStream: VideoStreamInfo & { timeBase: number }; packets: VideoPacketInfo };
+type Segmentation = { fps: number; framesPerSegment: number; segmentCount: number; segmentDuration: number };
 type ApiSession = { lastRequestedSegment: number | null; lastVariantIndex: number | null };
 
 @Injectable()
@@ -71,7 +72,7 @@ export class HlsService extends BaseService {
     return this.generateMainPlaylist(sessionId, ffmpeg, asset);
   }
 
-  async getMediaPlaylist(auth: AuthDto, assetId: string, sessionId: string) {
+  async getMediaPlaylist(auth: AuthDto, assetId: string, sessionId: string, variantIndex: number, position?: number) {
     await this.requireAccess({ auth, permission: Permission.AssetView, ids: [assetId] });
 
     const asset = await this.videoStreamRepository.getForMediaPlaylist(assetId, sessionId);
@@ -79,7 +80,11 @@ export class HlsService extends BaseService {
       throw new NotFoundException('Asset not found or metadata not yet ready for streaming');
     }
 
-    return this.generateMediaPlaylist(asset);
+    const segmentation = this.getSegmentation(asset);
+    const hintedSegment = position === undefined ? undefined : this.positionToSegmentIndex(segmentation, position);
+    this.prewarmVariant(assetId, sessionId, variantIndex, hintedSegment);
+
+    return this.generateMediaPlaylist(asset, segmentation);
   }
 
   async getSegment(
@@ -129,7 +134,7 @@ export class HlsService extends BaseService {
     const fps = ((asset.packets.packetCount * asset.videoStream.timeBase) / asset.packets.totalDuration).toFixed(3);
     const sourceResolution = Math.min(asset.videoStream.height, asset.videoStream.width);
     const targetResolution = Math.max(sourceResolution, HLS_VARIANTS[0].resolution);
-    const lines = ['#EXTM3U', `#EXT-X-VERSION:${HLS_VERSION}`];
+    const lines = ['#EXTM3U', `#EXT-X-VERSION:${HLS_VERSION}`, '#EXT-X-INDEPENDENT-SEGMENTS'];
     for (let i = 0; i < HLS_VARIANTS.length; i++) {
       const { resolution, bitrate, codec, codecString } = HLS_VARIANTS[i];
       if (resolution > targetResolution || !SUPPORTED_HWA_CODECS[ffmpeg.accel].includes(codec)) {
@@ -143,24 +148,33 @@ export class HlsService extends BaseService {
     }
     lines.push('');
 
-    if (lines.length === 3) {
+    if (lines.length === 4) {
       throw new NotFoundException('No supported variants for this video');
     }
 
     return lines.join('\n');
   }
 
-  private generateMediaPlaylist({ videoStream, packets }: AssetWithStreamInfo) {
+  private getSegmentation({ videoStream, packets }: AssetWithStreamInfo): Segmentation {
     const fps = (packets.packetCount * videoStream.timeBase) / packets.totalDuration;
     const framesPerSegment = Math.ceil(HLS_SEGMENT_DURATION * fps);
-    const fullSegmentDuration = framesPerSegment / fps;
     const segmentCount = Math.ceil(packets.outputFrames / framesPerSegment);
+    return { fps, framesPerSegment, segmentCount, segmentDuration: framesPerSegment / fps };
+  }
+
+  private positionToSegmentIndex({ segmentDuration, segmentCount }: Segmentation, position: number) {
+    return Math.min(Math.max(Math.floor(position / segmentDuration), 0), segmentCount - 1);
+  }
+
+  private generateMediaPlaylist({ packets }: AssetWithStreamInfo, segmentation: Segmentation) {
+    const { fps, framesPerSegment, segmentCount, segmentDuration: fullSegmentDuration } = segmentation;
     const lastSegmentFrames = packets.outputFrames - framesPerSegment * (segmentCount - 1);
     const lastSegmentDuration = lastSegmentFrames / fps;
 
     const lines = [
       '#EXTM3U',
       `#EXT-X-VERSION:${HLS_VERSION}`,
+      '#EXT-X-INDEPENDENT-SEGMENTS',
       `#EXT-X-TARGETDURATION:${HLS_SEGMENT_DURATION}`,
       '#EXT-X-MEDIA-SEQUENCE:0',
       '#EXT-X-PLAYLIST-TYPE:VOD',
@@ -175,6 +189,19 @@ export class HlsService extends BaseService {
     return lines.join('\n');
   }
 
+  private prewarmVariant(assetId: string, sessionId: string, variantIndex: number, hintedSegment?: number) {
+    const session = this.sessions.get(sessionId);
+    if (session?.lastVariantIndex === variantIndex) {
+      return;
+    }
+
+    const nextSegment = session && session.lastRequestedSegment !== null ? session.lastRequestedSegment + 1 : undefined;
+    const segmentIndex = hintedSegment ?? nextSegment;
+    if (segmentIndex !== undefined) {
+      this.websocketRepository.serverSend('HlsSegmentRequest', { sessionId, assetId, variantIndex, segmentIndex });
+    }
+  }
+
   private getSegmentKey({ sessionId, variantIndex, segmentIndex }: ArgOf<'HlsSegmentResult'>) {
     return `${sessionId}:${variantIndex}:${segmentIndex}`;
   }

server/src/services/transcoding.service.spec.ts

@@ -381,8 +381,6 @@ describe(TranscodingService.name, () => {
       '50',
       '-keyint_min',
       '50',
-      '-crf',
-      '23',
       '-copyts',
       '-r',
       '50130000/2012441',
@@ -417,6 +415,8 @@ describe(TranscodingService.name, () => {
           'aac',
           '-preset',
           '12',
+          '-crf',
+          '35',
           '-svtav1-params',
           'hierarchical-levels=3:lookahead=0:enable-tf=0:mbr=4000k',
           '-hls_segment_filename',
@@ -436,6 +436,8 @@ describe(TranscodingService.name, () => {
           'hvc1',
           '-preset',
           'ultrafast',
+          '-crf',
+          '28',
           '-maxrate',
           '2500k',
           '-bufsize',
@@ -459,6 +461,8 @@ describe(TranscodingService.name, () => {
           'aac',
           '-preset',
           'ultrafast',
+          '-crf',
+          '23',
           '-maxrate',
           '2500k',
           '-bufsize',

server/src/services/transcoding.service.ts

@@ -5,6 +5,7 @@ import {
   HLS_BACKPRESSURE_PAUSE_SEGMENTS,
   HLS_BACKPRESSURE_RESUME_SEGMENTS,
   HLS_CLEANUP_INTERVAL_MS,
+  HLS_CRF,
   HLS_INACTIVITY_TIMEOUT_MS,
   HLS_LEASE_DURATION_MS,
   HLS_SEGMENT_DURATION,
@@ -221,6 +222,7 @@ export class TranscodingService extends BaseService {
           targetResolution: String(variant.resolution),
           maxBitrate: `${Math.round(variant.bitrate / 1000)}k`,
           gopSize: gop,
+          crf: HLS_CRF[variant.codec],
         },
         this.videoInterfaces,
         { strictGop: true, lowLatency: true },