Video Streaming Architecture: How HLS and DASH Deliver Video to Billions

Netflix's 17,000-Server CDN

Netflix delivers 15% of the world's internet traffic during peak hours. A 2-hour 4K movie is roughly 14GB. You cannot download 14GB before watching. You cannot stream it as one giant file because a 10-second WiFi hiccup would freeze everything. Video Streaming Architecture solves this challenge by breaking video into smaller segments and dynamically delivering the appropriate quality based on network conditions. Netflix solved this by using Open Connect, 17,000 dedicated CDN servers placed inside ISPs worldwide, serving pre-positioned video segments. The technique behind this is adaptive bitrate streaming.

How Adaptive Bitrate Streaming Works

Adaptive Bitrate Streaming Pipeline:

  [Original Video File - 14GB]
           |
  [Encoder/Transcoder]
           |
  Encode at multiple quality levels:
  +------+------+--------+---------+
  | 240p | 480p |  720p  |  1080p  | 4K 
  | 300  | 750  |  2.5   |   5     | 15  Mbps
  | kbps | kbps |  Mbps  |  Mbps   |
  +------+------+--------+---------+
           |
  Split each into 2-6 second segments:
  seg001.ts, seg002.ts, seg003.ts, ...
           |
  Generate manifest file (playlist):
  manifest.m3u8 (HLS) or manifest.mpd (DASH)
           |
  [CDN Edge Servers Worldwide]

  Total: 1 movie = 5 qualities x 1800 segments = 9000 files

The Adaptive Algorithm in Action

Player's Adaptive Algorithm:

  Time    Bandwidth   Buffer    Quality Selected   Experience
  ------  ---------   --------  ----------------   ----------
  0:00    10 Mbps     0s        720p (safe start)  Loading...
  0:05    10 Mbps     15s       1080p (upgrade!)   Crisp
  0:30    10 Mbps     20s       1080p              Smooth
  0:45    2 Mbps      12s       480p (WiFi drop!)  Slight blur
  1:00    2 Mbps      8s        480p               No stutter
  1:15    8 Mbps      6s        720p (recovering)  Getting better
  1:30    12 Mbps     18s       1080p (back!)      Crisp again

  Key insight: quality changes are SEAMLESS.
  Each segment is independently decodable.
  Switching from 1080p seg047 to 480p seg048
  is invisible to the viewer - no buffering wheel.

  Buffer target: 30 seconds ahead
  Quality rule: pick highest quality that downloads
  faster than real-time playback speed

HLS vs DASH

Feature         HLS (Apple)         DASH (MPEG)
--------------  ------------------  ------------------
Manifest        .m3u8 playlist      .mpd (XML)
Segment format  .ts or .fmp4        .m4s (fmp4)
Codecs          H.264, H.265, AV1   Any codec
DRM support     FairPlay             Widevine, PlayReady
Browser support All browsers         Chrome, Firefox, Edge
Latency         6-30 seconds         2-10 seconds
Live streaming  Yes (LL-HLS)         Yes (LL-DASH)
Standard        Apple proprietary    ISO standard

Netflix uses:  DASH on most devices, HLS on Apple devices
YouTube uses:  DASH primarily
Twitch uses:   HLS for broad compatibility

Live Streaming: The Latency Challenge

Live Streaming Pipeline:

  [Camera] -> [Encoder] -> [Origin Server] -> [CDN] -> [Player]
                |              |                        |
              encode         package                 decode
              ~1-2s          ~1-3s                   ~1-2s
                                    distribute
                                     ~1-5s

  Total glass-to-glass latency: 4-12 seconds

  Low-Latency Techniques:
  - Shorter segments (0.5s instead of 6s)
  - Chunked transfer encoding (send partial segments)
  - WebRTC for sub-second (Twitch experiments)

  Live vs VOD trade-offs:
  Feature       VOD              Live
  -----------   ---------------  ---------------
  Segments      Pre-encoded      Created in RT
  CDN warmup    Fast (popular)   Cold (new segs)
  Seeking       Any position     Edge only
  Latency       Zero concern     Critical metric

Interview Tip

When designing a video streaming system, say: 'I would use adaptive bitrate streaming (HLS for broad compatibility). The upload pipeline transcodes video into multiple quality levels (240p to 4K) and splits each into 2-6 second segments. A manifest file lists available qualities and segment URLs. The player's adaptive algorithm measures bandwidth and buffer level every few seconds, selecting the highest quality that downloads faster than playback speed. Segments are served from CDN edge nodes - once a popular video segment is cached at an edge, all local viewers get it without hitting origin. For live streaming, I would use LL-HLS with shorter segments to minimize latency to 3-5 seconds.'

Key Takeaway

HLS and DASH slice video into small segments at multiple quality levels. The player's adaptive algorithm picks quality based on real-time bandwidth measurement. CDNs cache the static segments globally. This combination enables smooth playback for any user on any connection speed.