Design Subtitles System - System Design | WinJob

Subtitles Context and Accessibility Mandate

Problem framing

Design a global subtitle platform for YouTube/Netflix/Rev scale: timed text that is legally required in many markets, toggled per language, and cheap to deliver at billions of fetches per day.

Design choices

Treat captions as first-class media assets, not player afterthoughts
Anchor on 500M catalog videos with ~40% having ≥1 published track
Target p95 track manifest load < 80ms and ±40ms player sync tolerance
Tier automation: ASR draft → MT locale → human QC by content risk

Deep dive

Personas: deaf/hard-of-hearing viewers needing SDH, international viewers needing translation, creators uploading community captions, compliance officers auditing WCAG/FCC coverage. Success = accurate timing, fast publish after upload, and CDN-friendly read path.

Operational detail

WCAG 2.1 expects captions for prerecorded video with few exceptions
FCC CVAA influences US streaming compliance timelines
YouTube popularized community contribution with owner publish gate
Netflix ships forced narrative (FN) separate from SDH tracks
Rev sells human transcription SLAs for premium accuracy
Soft subs enable language toggle without re-encoding mezzanine
Interviewers probe whether you separate ingest, edit, and delivery
Caption traffic is read-heavy; pipeline is write-bursty on upload
Accessibility defects create regulatory and brand risk, not just UX bugs
Frame-accurate timing matters for fast dialogue and music lyrics
Kids content may require stricter human review before auto-publish
Live captions differ from VOD—defer unless interviewer asks
Clarify OTT web/TV scope vs broadcast encoder burn-in early
State catalog scale before drawing microservices
Quantify daily caption fetches separately from video segment QPS
Call out immutable revisions for safe rollback after bad publish
Mention drift when audio master changes after caption publish
Position FinOps: storage per language revision vs CDN egress
Note search/indexing as optional extension, not MVP blocker
Success metric: percent videos with compliant captions per locale
Anti-goal: burning 40 languages into every mezzanine encode
Anti-goal: blocking play on ASR failure—degrade gracefully

javaOne Dark Pro

1 public final class CueWindow {
2   private final double startSec;
3   private final double endSec;
4   private final String text;
5   public boolean activeAt(double playheadSec) {
6     return playheadSec >= startSec && playheadSec < endSec;
7   }
8 }

pythonOne Dark Pro

1 from dataclasses import dataclass
2  
3 @dataclass(frozen=True)
4 class CaptionCue:
5     start_ms: int
6     end_ms: int
7     text: str
8  
9 def active_cue(cues: list[CaptionCue], t_ms: int) -> CaptionCue | None:
10     return next((c for c in cues if c.start_ms <= t_ms < c.end_ms), None)

typescriptOne Dark Pro

1 interface CaptionTrack {
2   videoId: string;
3   language: string;
4   revision: number;
5   vttUrl: string;
6 }
7  
8 export function pickTrack(
9   tracks: CaptionTrack[],
10   preferredLang: string,
11 ): CaptionTrack | undefined {
12   return tracks.find((t) => t.language === preferredLang) ?? tracks[0];
13 }

Interviewer positioning

Open with accessibility mandate and scale numbers before APIs; interviewers reward measurable NFRs early.

Why interviewers care

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

Interview checkpoint

Name one failure story for Subtitles Context and Accessibility Mandate that proves you understand real outages, not happy-path diagrams.

Key Highlights

•Sidecar WebVTT decouples languages from video encode
•Immutable revisions enable safe publish/rollback
•ASR+MT+human review tiered by content risk class

1	public final class CueWindow {
2	private final double startSec;
3	private final double endSec;
4	private final String text;
5	public boolean activeAt(double playheadSec) {
6	return playheadSec >= startSec && playheadSec < endSec;
7	}
8	}

1	from dataclasses import dataclass
2
3	@dataclass(frozen=True)
4	class CaptionCue:
5	start_ms: int
6	end_ms: int
7	text: str
8
9	def active_cue(cues: list[CaptionCue], t_ms: int) -> CaptionCue \| None:
10	return next((c for c in cues if c.start_ms <= t_ms < c.end_ms), None)

1	interface CaptionTrack {
2	videoId: string;
3	language: string;
4	revision: number;
5	vttUrl: string;
6	}
7
8	export function pickTrack(
9	tracks: CaptionTrack[],
10	preferredLang: string,
11	): CaptionTrack \| undefined {
12	return tracks.find((t) => t.language === preferredLang) ?? tracks[0];
13	}