Design Netflix Video Streaming

Netflix serves 300+ million subscribers across 190+ countries, delivering 94 billion hours of content in H2 2024 alone. Unlike user-generated video platforms (YouTube), Netflix is a consumption-first architecture—the challenge is not upload volume but delivering pre-encoded content with sub-second playback start times while optimizing for quality-per-bit across wildly different devices and network conditions. This design covers the Open Connect CDN, per-title/shot-based encoding pipeline, adaptive bitrate delivery, and the personalization systems that drive 80% of viewing hours.

Abstract

Netflix’s streaming architecture is shaped by three fundamental constraints:

1. Pre-known catalog enables proactive caching: Unlike YouTube’s real-time uploads, Netflix’s finite catalog (~7,500 titles in US) allows overnight proactive distribution to edge servers. This achieves 98%+ cache hit rates without origin egress during playback.

2. Quality perception varies by content: A documentary with static shots compresses efficiently at 2 Mbps; an action sequence needs 8+ Mbps for equivalent perceptual quality. Per-title and shot-based encoding optimizes each scene independently.

3. Device fragmentation demands multi-codec strategy: Smart TVs from 2015 support only H.264; modern browsers support AV1. Netflix encodes each title in multiple codecs (H.264, HEVC, VP9, AV1) to maximize quality-per-bit on each device class.

The core mechanisms:

• Open Connect CDN embeds Netflix appliances directly in ISP networks, serving 100% of video traffic from 1,000+ locations
• Per-title encoding analyzes content complexity and generates custom bitrate ladders, achieving 20% average bandwidth savings
• Shot-based (Dynamic Optimizer) encoding allocates bits per-scene, achieving 50% bitrate reduction for equivalent quality
• VMAF-driven quality control ensures perceptual quality targets (93+) rather than arbitrary bitrate thresholds
• Proactive fill pipeline pre-positions content during off-peak hours based on predicted popularity

Requirements

Functional Requirements

Non-Functional Requirements

Scale Estimation

Netflix-scale baseline (Q4 2024):

1
Subscribers: 301.63 million (global)
2
DAU estimate: ~150M (50% daily engagement)
3
Concurrent peak estimate: 20M+ streams
4

5
Viewing traffic:
6
- H2 2024: 94 billion hours watched
7
- Daily average: ~500M hours/day
8
- Peak concurrent: 20M streams × 8 Mbps average = 160 Tbps
9
- With 98% edge hit rate: 3.2 Tbps from origin (worst case)
10

11
Content library:
12
- US catalog: ~7,500 titles (3,800 movies, 1,800 TV shows)
13
- ~59% Netflix Originals
14
- Storage per title: 10-50 encoded variants × 2-8 GB each
15
- Total encoded storage: ~5-10 PB estimated
16

17
Open Connect scale:
18
- 1,000+ deployment locations
19
- 4,669+ servers mapped by researchers
20
- Single OCA: up to 400 Gbps serving capacity

Live streaming record (November 2024):

1
Tyson vs Paul fight: 65 million concurrent viewers
2
Demonstrates burst capacity handling for events

Design Paths

Path A: Third-Party CDN (Traditional)

Best when:

• Smaller catalog (< 1,000 titles)
• Limited geographic scope
• Variable/unpredictable traffic patterns
• Cost-sensitive at small scale

Architecture:

• Use commercial CDN (Akamai, CloudFront, Fastly)
• Origin servers in cloud provider
• CDN handles caching, routing, TLS termination

Trade-offs:

• ✅ No infrastructure investment
• ✅ Pay-per-GB pricing aligns with usage
• ✅ Instant global reach
• ❌ Per-GB costs prohibitive at scale (Netflix would pay billions)
• ❌ No ISP-level embedding (additional network hops)
• ❌ Less control over quality-of-experience
• ❌ Generic caching strategies, not content-aware

Real-world example: Most OTT services (Hulu, Peacock, Paramount+) rely on commercial CDNs. Acceptable when scale doesn’t justify custom infrastructure.

Path B: Custom CDN with ISP Embedding (Netflix Model)

Best when:

• Massive scale (100M+ subscribers)
• Predictable catalog enables proactive caching
• Quality differentiation is competitive advantage
• Long-term investment justifies infrastructure

Architecture:

• Custom appliances (OCAs) deployed in ISP facilities
• Proactive overnight content distribution
• No per-GB transit costs after hardware investment
• Direct peering eliminates intermediate hops

Trade-offs:

• ✅ 98%+ cache hit rates with proactive fill
• ✅ Sub-millisecond latency to end users
• ✅ Amortized cost far below commercial CDN at scale
• ✅ Complete control over caching and routing
• ❌ Massive upfront hardware investment
• ❌ ISP relationship management complexity
• ❌ Operational burden (hardware, software, support)
• ❌ Long deployment lead times for new regions

Real-world example: Netflix Open Connect handles 100% of Netflix traffic. Google has similar infrastructure (Google Global Cache) for YouTube.

Path Comparison

This Article’s Focus

This article focuses on Path B (Custom CDN) because:

1. Netflix scale justifies the infrastructure investment
2. The proactive fill model is unique to predictable catalogs
3. ISP embedding provides measurable quality differentiation
4. The architecture demonstrates principles applicable to any large-scale streaming service

High-Level Design

Component Overview

Traffic Split: AWS vs Open Connect

Netflix pioneered a clean separation:

This separation allows each infrastructure to optimize for its workload. AWS handles millions of API requests; Open Connect handles hundreds of terabits of video.

Playback Flow

1. Client requests playback → API Gateway authenticates, routes to Playback Service
2. Playback Service generates manifest → Lists available streams (codecs, resolutions, bitrates)
3. Steering Service selects OCAs → Returns ranked list based on proximity, load, health
4. Client acquires license → DRM key exchange with License Service
5. Client fetches segments from OCA → ABR algorithm selects quality
6. Playback begins → Continuous quality adaptation based on buffer and throughput

Control Plane vs Data Plane

Control plane (AWS):

• Authentication and authorization
• Manifest generation
• DRM license issuance
• Personalization and recommendations
• A/B testing
• Billing and account management

Data plane (Open Connect):

• Video segment delivery
• Cache management
• BGP routing announcements
• Health reporting

Open Connect CDN

Architecture

OCA Appliance Specifications

Netflix offers multiple OCA configurations to ISP partners:

Performance milestones:

• 2017: 100 Gbps TLS streaming from single server (FreeBSD, NGINX)
• 2021: 400 Gbps from single server with TLS offload to NIC

2021 configuration achieving 400 Gbps:

1
CPU: AMD EPYC 7502p (32 cores, Rome)
2
Memory: 256 GB DDR4-3200
3
Storage: 18 × 2TB Western Digital SN720 NVMe
4
Network: 2 × Mellanox ConnectX-6 Dx (PCIe 4.0 x16)
5
OS: FreeBSD
6
Web Server: NGINX
7
TLS: Offloaded to NIC (key enabler for 400 Gbps)

TLS offloading to the NIC was the critical optimization—without it, throughput capped at ~240 Gbps due to CPU-bound encryption.

Deployment Models

Embedded (Inside ISP Network):

• OCA deployed in ISP data center
• Zero transit cost for ISP
• Lowest latency to subscribers (1-5ms)
• ISP provides power, space, connectivity

Peering (Internet Exchange Point):

• OCA deployed at IXP (e.g., Equinix, DE-CIX)
• Serves multiple ISPs via peering
• Higher latency than embedded (10-50ms)
• Used for ISPs without embedded partnership

BGP Routing

OCAs announce routes via BGP to attract Netflix traffic:

1
Netflix ASN: 2906 (customer-facing)
2
Open Connect ASN: 40027 (OCA infrastructure)
3

4
Route announcement:
5
- OCA announces covering prefix (e.g., 45.57.0.0/17)
6
- ISP router prefers local OCA routes
7
- Traffic stays within ISP network

Route selection criteria (in order):

1. OCA availability and health
2. Route specificity (most specific prefix wins)
3. AS path length
4. MED (Multi-Exit Discriminator)
5. Geographic proximity

Fill Pipeline

Proactive caching philosophy:

Netflix doesn’t wait for cache misses. The Fill Pipeline predicts content demand and pre-positions files during off-peak hours.

Fill process:

1. Popularity computation: Daily prediction of what members will watch
2. Replica calculation: More popular content → more replicas across clusters
3. Manifest generation: Each OCA receives custom file list
4. Off-peak transfer: Nightly downloads during configurable windows
5. Verification: Checksum validation after transfer

Optimization milestone:

• Original: Title-level popularity rankings
• Improved: File-level popularity (specific resolution/codec combinations)
• Result: Same cache efficiency with 50% less storage

This works because not all files for a title are equally popular. 4K HDR files are only needed for capable devices; most streams use 720p or 1080p.

Cache Performance

At Netflix scale, a 1% improvement in cache hit rate eliminates terabytes of daily origin egress.

Video Encoding Pipeline

The Cosmos Platform

Netflix rebuilt its encoding pipeline on Cosmos (completed September 2023), replacing the previous Reloaded system.

Platform components:

Processing scale:

A single Stranger Things episode requires processing 900 shots through the pipeline.

Per-Title Encoding

Introduced in 2015, per-title encoding customizes bitrate ladders based on content complexity.

Why fixed ladders fail:

1
Fixed ladder (traditional):
2
720p: 3 Mbps
3
1080p: 5 Mbps
4
4K: 16 Mbps
5

6
Reality:
7
- Documentary (low motion): VMAF 93 at 720p/1.5 Mbps
8
- Action movie (high motion): VMAF 93 at 720p/4 Mbps
9

10
Result: Fixed ladder wastes 2.5 Mbps on documentaries, under-serves action movies

Per-title methodology:

1. Encode source to hundreds of resolution/bitrate combinations
2. Compute VMAF score for each combination
3. Plot quality vs. bitrate curve (convex hull analysis)
4. Select optimal points on the curve as the custom ladder

Results:

• Average: 20% bandwidth savings without quality loss
• Some titles: 50%+ reduction (low-complexity content)

Shot-Based Encoding (Dynamic Optimizer)

Per-title improved on fixed ladders, but applied uniform bitrate across an entire title. Shot-based encoding optimizes each scene independently.

How it works:

1. Shot detection: Identify scene boundaries (cuts, fades, motion discontinuities)
2. Complexity analysis: Compute per-shot complexity (motion, texture, film grain)
3. Bitrate allocation: Allocate more bits to complex shots, fewer to simple ones
4. Encoding: IDR frames aligned with shot boundaries
5. Assembly: Concatenate variable-bitrate shots

Results:

Example scene:

1
Scene 1: Dialogue in dim room (low complexity)
2
  - Fixed: 5 Mbps, VMAF 96
3
  - Dynamic: 2 Mbps, VMAF 95
4
  - Savings: 3 Mbps, negligible quality loss
5

6
Scene 2: Car chase with explosions (high complexity)
7
  - Fixed: 5 Mbps, VMAF 82 (quality loss)
8
  - Dynamic: 9 Mbps, VMAF 93
9
  - Improvement: 11 VMAF points
10

11
Result: Same average bitrate, but quality is optimized per-scene

Codec Strategy

AV1 adoption (as of 2024):

Netflix reported AV1 now powers ~30% of streaming traffic, becoming their primary format for capable devices. On TVs specifically, AV1 delivers 1/3 less bandwidth than H.264/HEVC for equivalent quality.

Encoding decision tree:

1
if device.supports(AV1):
2
    serve AV1
3
elif device.supports(VP9):
4
    serve VP9
5
elif device.supports(HEVC) and device.is_apple:
6
    serve HEVC
7
else:
8
    serve H.264

VMAF Quality Metric

Video Multimethod Assessment Fusion (VMAF) is Netflix’s open-source perceptual quality metric, developed with USC, IPI/LS2N Nantes, and UT Austin LIVE lab.

Components:

• Visual Information Fidelity (VIF)
• Detail Loss Metric (DLM)
• Mean Co-Located Pixel Difference (MCPD)
• Fused via SVM regression

Score interpretation:

Why VMAF replaced PSNR:

• PSNR measures pixel-level difference, not human perception
• Film grain scores poorly on PSNR but looks natural to viewers
• Blurring scores well on PSNR but looks bad to viewers
• VMAF correlates with Mean Opinion Score (MOS) from human evaluations

Netflix replaced PSNR with VMAF in June 2016.

Adaptive Bitrate Streaming

Streaming Protocols

Netflix uses CMAF (Common Media Application Format) for packaging, which allows generating both HLS and DASH manifests from the same encoded segments.

Manifest Generation

When a client requests playback, the Playback Service generates a device-specific manifest:

Manifest contents:

• Available bitrate ladder (device-appropriate)
• Codec options (filtered by device capability)
• Audio track options (languages, formats)
• Subtitle tracks
• DRM information (license server URL, key IDs)
• CDN URLs (OCA endpoints)

Example HLS variant playlist (simplified):

1
#EXTM3U
2
#EXT-X-VERSION:7
3

4
#EXT-X-STREAM-INF:BANDWIDTH=12000000,RESOLUTION=3840x2160,CODECS="av01.0.13M.08"
5
4k-av1/playlist.m3u8
6

7
#EXT-X-STREAM-INF:BANDWIDTH=8000000,RESOLUTION=1920x1080,CODECS="av01.0.08M.08"
8
1080p-av1/playlist.m3u8
9

10
#EXT-X-STREAM-INF:BANDWIDTH=3000000,RESOLUTION=1280x720,CODECS="av01.0.04M.08"
11
720p-av1/playlist.m3u8
12

13
#EXT-X-STREAM-INF:BANDWIDTH=1000000,RESOLUTION=854x480,CODECS="av01.0.01M.08"
14
480p-av1/playlist.m3u8

ABR Algorithm

Netflix’s ABR algorithm balances multiple objectives:

Input signals:

• Buffer level (seconds of content buffered)
• Throughput history (recent segment download speeds)
• Device constraints (memory, CPU, battery)
• Network type (WiFi vs cellular)

Algorithm approach (hybrid):

1
Throughput estimation:
2
- Exponential weighted moving average of recent downloads
3
- Conservative safety margin (70% of estimated)
4

5
Buffer-based adjustment:
6
- Low buffer (< 10s): Prioritize stability, drop quality
7
- Target buffer (30s): Normal operation
8
- High buffer (> 45s): Allow quality increase
9

10
Quality selection:
11
safe_bitrate = throughput_estimate × 0.7
12
buffer_factor = buffer_level / target_buffer
13

14
if rebuffering_recent:
15
    selected = lowest_quality
16
elif buffer_factor < 0.5:
17
    selected = quality_below_safe_bitrate
18
else:
19
    selected = highest_quality_below_safe_bitrate

Startup behavior:

1. Start at conservative quality (720p or below)
2. Fetch first few segments quickly to build buffer
3. Ramp up quality as buffer exceeds 10 seconds
4. Reach target quality within 15-30 seconds

Quality switch constraints:

• Minimum dwell time between switches: 10 seconds (prevents oscillation)
• Maximum drop per switch: 2 quality levels
• Emergency drop: Any level if buffer < 5 seconds

Rebuffering Prevention

Netflix targets < 0.1% rebuffering ratio (industry-leading).

Mechanisms:

DRM and License Management

Multi-DRM Strategy

Netflix uses three DRM systems to cover all devices:

All three use CENC (Common Encryption) for the encrypted content, so Netflix encrypts once and decrypts with any compatible DRM.

License Acquisition Flow

Security layers:

• Device attestation: CDM proves it’s a legitimate device
• Entitlement check: License server verifies subscription
• Key rotation: Periodic re-keying for live content
• Output protection: HDCP enforcement for HD/4K

Offline Download

Download architecture:

1. Client requests download license (more permissive duration)
2. Segments downloaded and encrypted to device storage
3. Offline license has expiration (typically 7-30 days)
4. Playback requires periodic license renewal when online

Challenges:

• Storage estimation before download
• Codec/bitrate changes require re-download
• License expiration handling
• Device storage management

Personalization System

Recommendation Architecture

Netflix recommendations drive 75-80% of viewing hours. The system is estimated to save over $1 billion annually by reducing subscriber churn.

Algorithm Types

Personalized Artwork

Netflix dynamically selects thumbnail artwork per user:

How it works:

• Each title has multiple artwork variants
• Contextual bandit model predicts which variant will maximize click-through
• If Netflix knows you like Uma Thurman, Pulp Fiction shows Uma’s image
• If you like action, the same movie shows an action scene

Technical implementation:

• Pre-computed personalized artwork at render time
• Cached results for fast serving
• Labels components in artwork (actors, genres, moods)
• A/B testing validates selections

A/B Testing Scale

Every product change goes through rigorous A/B testing:

• Users may be in multiple concurrent tests
• Tests run for weeks to measure long-term retention
• Causal inference prevents confounding
• Multi-objective optimization balances engagement vs. satisfaction

Frontend Considerations

Client Architecture

Netflix clients (TV, mobile, web) share common patterns:

Core responsibilities:

Playback Start Optimization

Target: < 2 seconds to first frame

Techniques:

• DNS prefetch for OCA domains
• TLS session resumption
• Manifest pre-fetch while browsing
• License acquisition in parallel with segment fetch
• Initial segment at lower quality for fast decode

Offline Playback

Mobile apps support offline downloads with:

• Storage estimation: Show size before download
• Partial downloads: Resume interrupted downloads
• Smart downloads: Auto-download next episodes
• License management: Handle expiration gracefully
• Quality selection: User chooses download quality

Mobile-optimized encoding:

Netflix introduced AVCHi-Mobile and VP9-Mobile encodes specifically optimized for mobile screens and storage constraints.

Infrastructure

AWS Architecture

Netflix runs all non-video services on AWS:

EVCache scale (distributed cache):

1
Clusters: 200 Memcached clusters
2
Instances: 22,000 server instances
3
Operations: 400 million ops/second
4
Items: 2 trillion cached items
5
Storage: 14.3 petabytes total

Resilience (Chaos Engineering)

Netflix pioneered chaos engineering to ensure resilience:

Philosophy:

Systems should be resilient to failure by design. Random production failures ensure this resilience is real, not theoretical.

Global Architecture

Conclusion

Designing Netflix-scale streaming requires different optimizations than user-generated video platforms:

Key architectural decisions:

1. Custom CDN (Open Connect) eliminates transit costs and reduces latency by embedding in ISP networks
2. Proactive fill pipeline achieves 98%+ cache hit rates by predicting demand and pre-positioning content
3. Per-title and shot-based encoding saves 30-50% bandwidth by optimizing each scene independently
4. Multi-codec strategy (H.264 → VP9 → AV1) maximizes quality-per-bit for each device class
5. VMAF-driven quality control ensures perceptual quality targets rather than arbitrary bitrate thresholds
6. Personalization at scale drives 80% of viewing, justifying massive recommendation infrastructure investment

What this design optimizes for:

• Sub-2-second playback start
• Industry-leading rebuffering ratio (< 0.1%)
• Bandwidth efficiency (best quality per bit)
• Global scale (300M+ subscribers, 190+ countries)

What this design sacrifices:

• Upfront infrastructure investment ($100M+ in Open Connect)
• Operational complexity (ISP relationships, custom hardware)
• Encoding compute costs (per-title/shot analysis is expensive)

When to choose this design:

• Premium streaming services where quality is a differentiator
• Scale justifies custom infrastructure investment (> 100M subscribers)
• Predictable catalog enables proactive caching (not user-generated)

Appendix

Prerequisites

• CDN architecture: edge caching, origin shield concepts
• Video encoding: codecs, containers, bitrates, transcoding
• Streaming protocols: HLS, DASH fundamentals
• DRM: encryption, license servers, CDM integration
• Distributed systems: microservices, caching, eventual consistency

Terminology

Summary

• Netflix separates control plane (AWS) from data plane (Open Connect) for optimal scaling
• Open Connect embeds cache servers in 1,000+ ISP locations, serving 100% of video traffic
• Per-title encoding analyzes content complexity to generate custom bitrate ladders (20% savings)
• Shot-based encoding allocates bits per-scene for 30-50% additional efficiency
• AV1 codec now powers 30% of streaming, delivering 50%+ bandwidth savings over H.264
• Proactive fill pipeline pre-positions content overnight, achieving 98%+ cache hit rates
• Personalization (recommendations, artwork) drives 80% of viewing hours

References

• Netflix Open Connect - Official CDN documentation
• Netflix Open Connect Appliances - OCA specifications
• Netflix Tech Blog - Per-Title Encode Optimization - Per-title encoding methodology
• Netflix Tech Blog - Dynamic Optimizer - Shot-based encoding
• Netflix Tech Blog - Serving 100 Gbps - OCA performance optimization
• FreeBSD 400Gbps Presentation - 400 Gbps achievement details
• Netflix Tech Blog - Rebuilding Video Processing Pipeline - Cosmos platform architecture
• Netflix Tech Blog - The Netflix Cosmos Platform - Pipeline components
• Netflix VMAF GitHub - Quality metric source and documentation
• Netflix Tech Blog - HDR Dynamic Optimization - HDR encoding optimization
• Netflix EVCache GitHub - Distributed caching system
• Netflix Research - Recommendations - Recommendation system overview
• HLS Specification (RFC 8216) - HTTP Live Streaming standard
• DASH Specification (ISO/IEC 23009-1) - MPEG-DASH standard