Video Delivery

Why Video Is Its Own Problem

 scale of bytes:
   webpage:        ~2MB
   5-min 1080p:    ~400MB      (200x a webpage!)
   1M concurrent viewers × 5Mbps = 5 Tbps egress

 no origin fleet serves this. video delivery is a CDN-native
 workload — the architecture exists to make one upload
 serve millions.

The Segmenting Foundation

 never stream a file; stream SEGMENTS (2-10s chunks):

 upload → transcode → segments → packaged
 master.m3u8 (playlist) + seg_001.ts, seg_002.ts, ...

 why segments win:
 - each is an independent cacheable URL
 - players fetch ahead (buffering tolerance)
 - adaptive switching between qualities mid-stream
 - CDN-friendly: immutable files, power-law popularity

Adaptive Bitrate Streaming (ABR)

 same content encoded at multiple bitrates:

 1080p @ 5Mbps ─┐
 720p  @ 2.5Mbs ─┼─► player picks per CURRENT bandwidth/device:
 480p  @ 1Mbps  ─┤     bandwidth drops → switch DOWN seamlessly
 360p  @ 0.6Mbps ┘     recovers       → switch UP

 user experience: quality flexes instead of buffering.
 the playlist (m3u8) describes the ladder; player logic
 measures throughput and chooses continuously.

The CDN Economics

 VOD (on-demand):
   segment popularity = power law → hit rates >99%
   origin sees ~1 fetch per segment per region, EVER
   1M viewers cost origin almost NOTHING extra

 LIVE:
   sliding window of recent segments, short TTLs (~segment duration)
   every viewer hits the same current segments simultaneously:
   1M viewers × same 3 segments = massive fan-out at edges,
   but STILL only ~1 origin fetch per segment (with shielding)
   
 live latency floor ≈ segment length + buffer:
   standard HLS: 15-30s    LL-HLS: 2-5s    WebRTC sub-second

Upload and Transcode Pipeline

 creator uploads → object storage
                     │ event
                     ▼
              transcode workers (queue-driven)
              - multiple resolutions/bitrate ladder
              - thumbnail/sprite generation  
              - packaging (HLS+DASH manifests)
                     │
                     ▼
            publish to origin/CDN, invalidate nothing
            (new unique paths)

 transcoding is compute-heavy batch work: queue-based
 scaling, minutes-long latency tolerated by design.

Interview Framing

“Design YouTube/video platform” scored skeleton: object storage for originals, queued transcoding to ABR ladders, segment-based packaging (HLS/DASH named), CDN delivery with shield math (“origin sees one fetch per segment”), player-side adaptive logic, and live-vs-VOD distinction (sliding window TTLs). Numbers matter here more than usual — quote the 5Tbps-style arithmetic to show you grasp why the architecture looks like it does.