Design Video Chapters - System Design | WinJob

Video Chapters Context and Platform Goals

Video chapters are navigation metadata: time-indexed markers on the progress bar, a chapter list, and shareable deep links. YouTube popularized auto chapters from speech; Spotify uses episode markers on long podcasts; Apple TV surfaces skip sections in educational content.

Quantified constraints

Assume 2B weekly plays, 800K new long-form uploads/day, median 14 minutes, target 6–10 chapters per video, p95 auto-chapter latency under 5 minutes after ASR completes.

Interviewers want you to separate read path (tiny JSON manifest at player init) from write path (ASR → segmentation → title generation → publish). Mention creator manual overrides and version binding to transcodeGeneration.

Reference implementation sketches

javaOne Dark Pro

1 public final class ChapterJobKey {
2   private final String videoId;
3   private final String transcodeVersion;
4   private final int modelVersion;
5   public String dedupeKey() {
6     return videoId + ":" + transcodeVersion + ":chapters-v" + modelVersion;
7   }
8 }

pythonOne Dark Pro

1 from dataclasses import dataclass
2  
3 @dataclass(frozen=True)
4 class ChapterBoundary:
5     start_ms: int
6     confidence: float
7     source: str
8  
9 def merge_boundaries(candidates: list[ChapterBoundary], min_gap_ms: int = 10_000) -> list[int]:
10     ordered = sorted(candidates, key=lambda c: c.start_ms)
11     starts: list[int] = []
12     last = -min_gap_ms
13     for c in ordered:
14         if c.start_ms - last >= min_gap_ms:
15             starts.append(c.start_ms)
16             last = c.start_ms
17     return starts

typescriptOne Dark Pro

1 interface ChapterMarker {
2   startMs: number;
3   title: string;
4   durationMs?: number;
5 }
6  
7 export function seekToChapter(
8   player: { seek: (ms: number) => void },
9   chapter: ChapterMarker,
10 ): void {
11   player.seek(chapter.startMs);
12 }

Failure and edge cases

For Video Chapters Context and Platform Goals, document what happens when ASR returns empty text, when creators delete a mid-video chapter leaving a gap, when transcodeVersion increments mid-playback, and when CDN serves an expired manifest during a viral spike. State explicit degraded behavior: hide markers rather than show wrong times.

Observability

Emit metrics chapter_job_latency_seconds, manifest_cache_hit_ratio, publish_conflict_total, and segmentation_confidence_histogram tagged by channel tier. Trace publish path with videoId and transcodeVersion on every span.

Why interviewers care

Video Chapters interviews reward crisp scope, explicit trade-offs, and failure stories—not generic microservice diagrams.

Interview checkpoint

Name one failure story for Video Chapters Context and Platform Goals that proves you understand real outages, not happy-path diagrams.

Key Highlights

•Chapters are time-indexed navigation metadata bound to a specific video version
•YouTube/Spotify/Apple use auto + manual chapter workflows at upload scale
•Player UX: scrubber markers, chapter list, deep links, and search snippets

1	public final class ChapterJobKey {
2	private final String videoId;
3	private final String transcodeVersion;
4	private final int modelVersion;
5	public String dedupeKey() {
6	return videoId + ":" + transcodeVersion + ":chapters-v" + modelVersion;
7	}
8	}

1	from dataclasses import dataclass
2
3	@dataclass(frozen=True)
4	class ChapterBoundary:
5	start_ms: int
6	confidence: float
7	source: str
8
9	def merge_boundaries(candidates: list[ChapterBoundary], min_gap_ms: int = 10_000) -> list[int]:
10	ordered = sorted(candidates, key=lambda c: c.start_ms)
11	starts: list[int] = []
12	last = -min_gap_ms
13	for c in ordered:
14	if c.start_ms - last >= min_gap_ms:
15	starts.append(c.start_ms)
16	last = c.start_ms
17	return starts

1	interface ChapterMarker {
2	startMs: number;
3	title: string;
4	durationMs?: number;
5	}
6
7	export function seekToChapter(
8	player: { seek: (ms: number) => void },
9	chapter: ChapterMarker,
10	): void {
11	player.seek(chapter.startMs);
12	}