Design Video Clipping Tool - System Design | WinJob

Video Clipping Context and Platform Goals

Scope a Twitch/YouTube/TikTok-class clipping product: trim VOD or live DVR, render a standalone asset, ship a viral share page. Anchor on 50M DAU, 4M clips/day, 42s average clip length, p95 time-to-playable under 25s for 60s VOD trims, and sub-8s instant clips from live buffers.

Problem framing

Scope a Twitch/YouTube/TikTok-class clipping product: trim VOD or live DVR, render a standalone asset, ship a viral share page.

Design choices

Treat clips as immutable published artifacts with pinned transcodeVersion.
Split synchronous trim validation from asynchronous GPU render.
Optimize share-page reads on CDN; isolate write-heavy render queues.
Support both VOD timeline trim and live instant-clip without duplicating storage paths.

Deep dive

Open by naming three user journeys: viewer clips a live moment, creator trims a VOD highlight, social crawler fetches OG metadata. Contrast Twitch Clip DVR, YouTube share-at-timestamp evolution, and TikTok cut-from-long-video export presets.

javaOne Dark Pro

1 public enum ClipState { DRAFT, QUEUED, RENDERING, READY, FAILED, REMOVED }
2 public record ClipRecord(String clipId, ClipState state, String manifestUri) {}

pythonOne Dark Pro

1 @dataclass(frozen=True)
2 class ClipJob:
3     source_id: str
4     transcode_version: str
5     preset: str
6  
7 def dedupe_key(job: ClipJob, start_ms: int, end_ms: int) -> str:
8     return f"{job.source_id}:{job.transcode_version}:{start_ms}-{end_ms}:{job.preset}"

typescriptOne Dark Pro

1 type ClipStatus = "draft" | "queued" | "rendering" | "ready" | "failed";
2  
3 export interface ClipManifest {
4   clipId: string;
5   transcodeVersion: string;
6   renditions: { height: number; playlistUrl: string }[];
7 }

Interviewer positioning

State headline SLIs before boxes: time-to-playable, clip page availability, render success rate, and cost per clip-minute.

Extended design notes

Design note 1: Weighted fair queuing isolates viral channel clip storms from default tenant render SLOs — applied in Video Clipping Context and Platform Goals (section 1).
Design note 2: Spot GPU workers need checkpointed ffmpeg jobs so preemption does not double-charge creators — applied in Video Clipping Context and Platform Goals (section 1).
Design note 3: Min 5s / max 60s clip guards prevent abuse loops that exhaust render pools — applied in Video Clipping Context and Platform Goals (section 1).
Design note 4: Embed CSP allowlists stop hotlinking clip players on unauthorized domains — applied in Video Clipping Context and Platform Goals (section 1).
Design note 5: Atomic publish swaps clip state to READY only when every rendition object exists in storage — applied in Video Clipping Context and Platform Goals (section 1).
Design note 6: 202 Accepted on create returns poll URL — never block HTTP thread on ffmpeg completion — applied in Video Clipping Context and Platform Goals (section 1).
Design note 7: Thumbnail burst on publish warms CDN via prefetch from social crawler user agents — applied in Video Clipping Context and Platform Goals (section 1).
Design note 8: Postgres holds clip metadata; object storage holds bytes — never stream video through API tier — applied in Video Clipping Context and Platform Goals (section 1).
Design note 9: DLQ plus reconciliation cron rescues stuck RENDERING rows after worker lease expiry — applied in Video Clipping Context and Platform Goals (section 1).
Design note 10: Cross-region clip replication follows read-heavy share spikes, not write-heavy trim traffic — applied in Video Clipping Context and Platform Goals (section 1).
Design note 11: FinOps tags attribute GPU-seconds per clip-minute to product lines and creator tiers — applied in Video Clipping Context and Platform Goals (section 1).
Design note 12: Player processing UI maps to clip.render.progress events over WebSocket or SSE — applied in Video Clipping Context and Platform Goals (section 1).
Design note 13: Virtual sub-manifests save storage but complicate parent GC — interviewers probe this trade — applied in Video Clipping Context and Platform Goals (section 1).
Design note 14: Burn-in watermark jobs always take full transcode path — budget separately in SLO math — applied in Video Clipping Context and Platform Goals (section 1).
Design note 15: Scene detection can propose clip boundaries but human confirm remains default for brand safety — applied in Video Clipping Context and Platform Goals (section 1).
Design note 16: Clip spam rate limits combine per-user, per-channel, and per-IP dimensions — applied in Video Clipping Context and Platform Goals (section 1).
Design note 17: Signed embed tokens bind clipId to parent channel and expiry for partner syndication — applied in Video Clipping Context and Platform Goals (section 1).
Design note 18: Twitch instant replay copies a rolling DVR segment window instead of re-reading hours of VOD — applied in Video Clipping Context and Platform Goals (section 1).
Design note 19: YouTube Clips bind share URLs to immutable child manifests so parent re-transcodes do not break playback — applied in Video Clipping Context and Platform Goals (section 1).
Design note 20: TikTok vertical presets re-encode landscape sources with safe-title overlays for mobile feeds — applied in Video Clipping Context and Platform Goals (section 1).
Design note 21: Partial remux copies GOP-aligned fMP4 segments and skips GPU when trim handles land on keyframes — applied in Video Clipping Context and Platform Goals (section 1).
Design note 22: Clip dedupe keys must include transcodeVersion or seekers drift after ladder replays — applied in Video Clipping Context and Platform Goals (section 1).
Design note 23: Share-page QPS dominates clip APIs — cache OG HTML and manifest at CDN with short purge hooks — applied in Video Clipping Context and Platform Goals (section 1).
Design note 24: Live clip latency budgets split between buffer finalize (ms) and GPU transcode queue (seconds) — applied in Video Clipping Context and Platform Goals (section 1).
Design note 25: Content-ID scans on clip audio run async and gate publish for rights-sensitive channels — applied in Video Clipping Context and Platform Goals (section 1).
Design note 26: Weighted fair queuing isolates viral channel clip storms from default tenant render SLOs — applied in Video Clipping Context and Platform Goals (section 1).
Design note 27: Spot GPU workers need checkpointed ffmpeg jobs so preemption does not double-charge creators — applied in Video Clipping Context and Platform Goals (section 1).
Design note 28: Min 5s / max 60s clip guards prevent abuse loops that exhaust render pools — applied in Video Clipping Context and Platform Goals (section 1).
Design note 29: Embed CSP allowlists stop hotlinking clip players on unauthorized domains — applied in Video Clipping Context and Platform Goals (section 1).
Design note 30: Atomic publish swaps clip state to READY only when every rendition object exists in storage — applied in Video Clipping Context and Platform Goals (section 1).
Design note 31: 202 Accepted on create returns poll URL — never block HTTP thread on ffmpeg completion — applied in Video Clipping Context and Platform Goals (section 1).
Design note 32: Thumbnail burst on publish warms CDN via prefetch from social crawler user agents — applied in Video Clipping Context and Platform Goals (section 1).
Design note 33: Postgres holds clip metadata; object storage holds bytes — never stream video through API tier — applied in Video Clipping Context and Platform Goals (section 1).
Design note 34: DLQ plus reconciliation cron rescues stuck RENDERING rows after worker lease expiry — applied in Video Clipping Context and Platform Goals (section 1).
Design note 35: Cross-region clip replication follows read-heavy share spikes, not write-heavy trim traffic — applied in Video Clipping Context and Platform Goals (section 1).
Design note 36: FinOps tags attribute GPU-seconds per clip-minute to product lines and creator tiers — applied in Video Clipping Context and Platform Goals (section 1).
Design note 37: Player processing UI maps to clip.render.progress events over WebSocket or SSE — applied in Video Clipping Context and Platform Goals (section 1).
Design note 38: Virtual sub-manifests save storage but complicate parent GC — interviewers probe this trade — applied in Video Clipping Context and Platform Goals (section 1).
Design note 39: Burn-in watermark jobs always take full transcode path — budget separately in SLO math — applied in Video Clipping Context and Platform Goals (section 1).
Design note 40: Scene detection can propose clip boundaries but human confirm remains default for brand safety — applied in Video Clipping Context and Platform Goals (section 1).
Design note 41: Clip spam rate limits combine per-user, per-channel, and per-IP dimensions — applied in Video Clipping Context and Platform Goals (section 1).
Design note 42: Signed embed tokens bind clipId to parent channel and expiry for partner syndication — applied in Video Clipping Context and Platform Goals (section 1).
Design note 43: Twitch instant replay copies a rolling DVR segment window instead of re-reading hours of VOD — applied in Video Clipping Context and Platform Goals (section 1).
Design note 44: YouTube Clips bind share URLs to immutable child manifests so parent re-transcodes do not break playback — applied in Video Clipping Context and Platform Goals (section 1).
Design note 45: TikTok vertical presets re-encode landscape sources with safe-title overlays for mobile feeds — applied in Video Clipping Context and Platform Goals (section 1).
Design note 46: Partial remux copies GOP-aligned fMP4 segments and skips GPU when trim handles land on keyframes — applied in Video Clipping Context and Platform Goals (section 1).
Design note 47: Clip dedupe keys must include transcodeVersion or seekers drift after ladder replays — applied in Video Clipping Context and Platform Goals (section 1).
Design note 48: Share-page QPS dominates clip APIs — cache OG HTML and manifest at CDN with short purge hooks — applied in Video Clipping Context and Platform Goals (section 1).
Design note 49: Live clip latency budgets split between buffer finalize (ms) and GPU transcode queue (seconds) — applied in Video Clipping Context and Platform Goals (section 1).
Design note 50: Content-ID scans on clip audio run async and gate publish for rights-sensitive channels — applied in Video Clipping Context and Platform Goals (section 1).
Design note 51: Weighted fair queuing isolates viral channel clip storms from default tenant render SLOs — applied in Video Clipping Context and Platform Goals (section 1).
Design note 52: Spot GPU workers need checkpointed ffmpeg jobs so preemption does not double-charge creators — applied in Video Clipping Context and Platform Goals (section 1).
Design note 53: Min 5s / max 60s clip guards prevent abuse loops that exhaust render pools — applied in Video Clipping Context and Platform Goals (section 1).
Design note 54: Embed CSP allowlists stop hotlinking clip players on unauthorized domains — applied in Video Clipping Context and Platform Goals (section 1).
Design note 55: Atomic publish swaps clip state to READY only when every rendition object exists in storage — applied in Video Clipping Context and Platform Goals (section 1).
Design note 56: 202 Accepted on create returns poll URL — never block HTTP thread on ffmpeg completion — applied in Video Clipping Context and Platform Goals (section 1).
Design note 57: Thumbnail burst on publish warms CDN via prefetch from social crawler user agents — applied in Video Clipping Context and Platform Goals (section 1).
Design note 58: Postgres holds clip metadata; object storage holds bytes — never stream video through API tier — applied in Video Clipping Context and Platform Goals (section 1).
Design note 59: DLQ plus reconciliation cron rescues stuck RENDERING rows after worker lease expiry — applied in Video Clipping Context and Platform Goals (section 1).
Design note 60: Cross-region clip replication follows read-heavy share spikes, not write-heavy trim traffic — applied in Video Clipping Context and Platform Goals (section 1).
Design note 61: FinOps tags attribute GPU-seconds per clip-minute to product lines and creator tiers — applied in Video Clipping Context and Platform Goals (section 1).
Design note 62: Player processing UI maps to clip.render.progress events over WebSocket or SSE — applied in Video Clipping Context and Platform Goals (section 1).
Design note 63: Virtual sub-manifests save storage but complicate parent GC — interviewers probe this trade — applied in Video Clipping Context and Platform Goals (section 1).
Design note 64: Burn-in watermark jobs always take full transcode path — budget separately in SLO math — applied in Video Clipping Context and Platform Goals (section 1).
Design note 65: Scene detection can propose clip boundaries but human confirm remains default for brand safety — applied in Video Clipping Context and Platform Goals (section 1).
Design note 66: Clip spam rate limits combine per-user, per-channel, and per-IP dimensions — applied in Video Clipping Context and Platform Goals (section 1).
Design note 67: Signed embed tokens bind clipId to parent channel and expiry for partner syndication — applied in Video Clipping Context and Platform Goals (section 1).
Design note 68: Twitch instant replay copies a rolling DVR segment window instead of re-reading hours of VOD — applied in Video Clipping Context and Platform Goals (section 1).
Design note 69: YouTube Clips bind share URLs to immutable child manifests so parent re-transcodes do not break playback — applied in Video Clipping Context and Platform Goals (section 1).
Design note 70: TikTok vertical presets re-encode landscape sources with safe-title overlays for mobile feeds — applied in Video Clipping Context and Platform Goals (section 1).
Design note 71: Partial remux copies GOP-aligned fMP4 segments and skips GPU when trim handles land on keyframes — applied in Video Clipping Context and Platform Goals (section 1).
Design note 72: Clip dedupe keys must include transcodeVersion or seekers drift after ladder replays — applied in Video Clipping Context and Platform Goals (section 1).
Design note 73: Share-page QPS dominates clip APIs — cache OG HTML and manifest at CDN with short purge hooks — applied in Video Clipping Context and Platform Goals (section 1).
Design note 74: Live clip latency budgets split between buffer finalize (ms) and GPU transcode queue (seconds) — applied in Video Clipping Context and Platform Goals (section 1).
Design note 75: Content-ID scans on clip audio run async and gate publish for rights-sensitive channels — applied in Video Clipping Context and Platform Goals (section 1).
Design note 76: Weighted fair queuing isolates viral channel clip storms from default tenant render SLOs — applied in Video Clipping Context and Platform Goals (section 1).
Design note 77: Spot GPU workers need checkpointed ffmpeg jobs so preemption does not double-charge creators — applied in Video Clipping Context and Platform Goals (section 1).
Design note 78: Min 5s / max 60s clip guards prevent abuse loops that exhaust render pools — applied in Video Clipping Context and Platform Goals (section 1).
Design note 79: Embed CSP allowlists stop hotlinking clip players on unauthorized domains — applied in Video Clipping Context and Platform Goals (section 1).
Design note 80: Atomic publish swaps clip state to READY only when every rendition object exists in storage — applied in Video Clipping Context and Platform Goals (section 1).
Design note 81: 202 Accepted on create returns poll URL — never block HTTP thread on ffmpeg completion — applied in Video Clipping Context and Platform Goals (section 1).
Design note 82: Thumbnail burst on publish warms CDN via prefetch from social crawler user agents — applied in Video Clipping Context and Platform Goals (section 1).
Design note 83: Postgres holds clip metadata; object storage holds bytes — never stream video through API tier — applied in Video Clipping Context and Platform Goals (section 1).
Design note 84: DLQ plus reconciliation cron rescues stuck RENDERING rows after worker lease expiry — applied in Video Clipping Context and Platform Goals (section 1).
Design note 85: Cross-region clip replication follows read-heavy share spikes, not write-heavy trim traffic — applied in Video Clipping Context and Platform Goals (section 1).
Design note 86: FinOps tags attribute GPU-seconds per clip-minute to product lines and creator tiers — applied in Video Clipping Context and Platform Goals (section 1).
Design note 87: Player processing UI maps to clip.render.progress events over WebSocket or SSE — applied in Video Clipping Context and Platform Goals (section 1).
Design note 88: Virtual sub-manifests save storage but complicate parent GC — interviewers probe this trade — applied in Video Clipping Context and Platform Goals (section 1).
Design note 89: Burn-in watermark jobs always take full transcode path — budget separately in SLO math — applied in Video Clipping Context and Platform Goals (section 1).
Design note 90: Scene detection can propose clip boundaries but human confirm remains default for brand safety — applied in Video Clipping Context and Platform Goals (section 1).
Design note 91: Clip spam rate limits combine per-user, per-channel, and per-IP dimensions — applied in Video Clipping Context and Platform Goals (section 1).
Design note 92: Signed embed tokens bind clipId to parent channel and expiry for partner syndication — applied in Video Clipping Context and Platform Goals (section 1).
Design note 93: Twitch instant replay copies a rolling DVR segment window instead of re-reading hours of VOD — applied in Video Clipping Context and Platform Goals (section 1).
Design note 94: YouTube Clips bind share URLs to immutable child manifests so parent re-transcodes do not break playback — applied in Video Clipping Context and Platform Goals (section 1).
Design note 95: TikTok vertical presets re-encode landscape sources with safe-title overlays for mobile feeds — applied in Video Clipping Context and Platform Goals (section 1).
Design note 96: Partial remux copies GOP-aligned fMP4 segments and skips GPU when trim handles land on keyframes — applied in Video Clipping Context and Platform Goals (section 1).
Design note 97: Clip dedupe keys must include transcodeVersion or seekers drift after ladder replays — applied in Video Clipping Context and Platform Goals (section 1).
Design note 98: Share-page QPS dominates clip APIs — cache OG HTML and manifest at CDN with short purge hooks — applied in Video Clipping Context and Platform Goals (section 1).
Design note 99: Live clip latency budgets split between buffer finalize (ms) and GPU transcode queue (seconds) — applied in Video Clipping Context and Platform Goals (section 1).
Design note 100: Content-ID scans on clip audio run async and gate publish for rights-sensitive channels — applied in Video Clipping Context and Platform Goals (section 1).
Design note 101: Weighted fair queuing isolates viral channel clip storms from default tenant render SLOs — applied in Video Clipping Context and Platform Goals (section 1).
Design note 102: Spot GPU workers need checkpointed ffmpeg jobs so preemption does not double-charge creators — applied in Video Clipping Context and Platform Goals (section 1).
Design note 103: Min 5s / max 60s clip guards prevent abuse loops that exhaust render pools — applied in Video Clipping Context and Platform Goals (section 1).
Design note 104: Embed CSP allowlists stop hotlinking clip players on unauthorized domains — applied in Video Clipping Context and Platform Goals (section 1).
Design note 105: Atomic publish swaps clip state to READY only when every rendition object exists in storage — applied in Video Clipping Context and Platform Goals (section 1).
Design note 106: 202 Accepted on create returns poll URL — never block HTTP thread on ffmpeg completion — applied in Video Clipping Context and Platform Goals (section 1).
Design note 107: Thumbnail burst on publish warms CDN via prefetch from social crawler user agents — applied in Video Clipping Context and Platform Goals (section 1).
Design note 108: Postgres holds clip metadata; object storage holds bytes — never stream video through API tier — applied in Video Clipping Context and Platform Goals (section 1).
Design note 109: DLQ plus reconciliation cron rescues stuck RENDERING rows after worker lease expiry — applied in Video Clipping Context and Platform Goals (section 1).
Design note 110: Cross-region clip replication follows read-heavy share spikes, not write-heavy trim traffic — applied in Video Clipping Context and Platform Goals (section 1).
Design note 111: FinOps tags attribute GPU-seconds per clip-minute to product lines and creator tiers — applied in Video Clipping Context and Platform Goals (section 1).
Design note 112: Player processing UI maps to clip.render.progress events over WebSocket or SSE — applied in Video Clipping Context and Platform Goals (section 1).
Design note 113: Virtual sub-manifests save storage but complicate parent GC — interviewers probe this trade — applied in Video Clipping Context and Platform Goals (section 1).
Design note 114: Burn-in watermark jobs always take full transcode path — budget separately in SLO math — applied in Video Clipping Context and Platform Goals (section 1).
Design note 115: Scene detection can propose clip boundaries but human confirm remains default for brand safety — applied in Video Clipping Context and Platform Goals (section 1).
Design note 116: Clip spam rate limits combine per-user, per-channel, and per-IP dimensions — applied in Video Clipping Context and Platform Goals (section 1).
Design note 117: Signed embed tokens bind clipId to parent channel and expiry for partner syndication — applied in Video Clipping Context and Platform Goals (section 1).
Design note 118: Twitch instant replay copies a rolling DVR segment window instead of re-reading hours of VOD — applied in Video Clipping Context and Platform Goals (section 1).
Design note 119: YouTube Clips bind share URLs to immutable child manifests so parent re-transcodes do not break playback — applied in Video Clipping Context and Platform Goals (section 1).
Design note 120: TikTok vertical presets re-encode landscape sources with safe-title overlays for mobile feeds — applied in Video Clipping Context and Platform Goals (section 1).

Video Clipping Context and Platform Goals — talking points

Clips are immutable child assets decoupled from parent lifecycle
Live instant-clip uses rolling DVR — not full VOD re-read
Share-page CDN cache is the billion-play read path

Why interviewers care

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

Interview checkpoint

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

Key Highlights

•Clips are immutable child assets decoupled from parent lifecycle
•Live instant-clip uses rolling DVR — not full VOD re-read
•Share-page CDN cache is the billion-play read path

1	public enum ClipState { DRAFT, QUEUED, RENDERING, READY, FAILED, REMOVED }
2	public record ClipRecord(String clipId, ClipState state, String manifestUri) {}

1	@dataclass(frozen=True)
2	class ClipJob:
3	source_id: str
4	transcode_version: str
5	preset: str
6
7	def dedupe_key(job: ClipJob, start_ms: int, end_ms: int) -> str:
8	return f"{job.source_id}:{job.transcode_version}:{start_ms}-{end_ms}:{job.preset}"

1	type ClipStatus = "draft" \| "queued" \| "rendering" \| "ready" \| "failed";
2
3	export interface ClipManifest {
4	clipId: string;
5	transcodeVersion: string;
6	renditions: { height: number; playlistUrl: string }[];
7	}