How Netflix Serves 300 Million Users

Netflix accounts for 15% of all downstream internet traffic worldwide. More than YouTube. More than every social media platform combined. When 300 million subscribers press Play, the request touches DNS, an API gateway, microservices, a custom CDN, multiple databases, and an event pipeline processing 2 trillion events per day. And it all happens in under 2 seconds.

This page is the capstone of the system design series. Every concept from the previous pages — DNS, load balancing, CDN, caching, databases, queues, replication, consistent hashing, circuit breakers — appears here, applied to a real system at massive scale. Netflix isn't just using these concepts. They invented several of them. Chaos Monkey, EVCache, Zuul, and Open Connect all came from Netflix engineering.

But most architecture diagrams only show the user-facing flow: you press Play, video arrives. They skip the other half — how content gets from a studio's editing suite to 18,000+ servers inside ISPs worldwide. This page covers both sides: the request journey AND the content pipeline. Plus the estimation math that system design interviews actually test.

One click. Seven systems. Under 2 seconds.

When you press Play, your device doesn't download the entire movie. It receives a manifest — a list of URLs for 4-second video chunks at various quality levels. Your device then fetches chunks one at a time from the nearest Open Connect Appliance, adapting quality per chunk based on your current bandwidth.

This is why Netflix never fully buffers. It only needs to be a few chunks ahead. If your WiFi drops, it plays the next chunk at lower quality rather than buffering. When bandwidth recovers, quality climbs back up. The user barely notices. This adaptive approach is why Netflix can stream 4K globally — they don't need consistent 25 Mbps, just enough bandwidth to stay a few chunks ahead.

The investment in AV1 codec is strategic. AV1 delivers the same visual quality as H.264 at 30% less bandwidth. At Netflix's scale (500+ petabytes per day), 30% less bandwidth means hundreds of millions of dollars in CDN and ISP costs saved annually. This is why Netflix co-founded the Alliance for Open Media to develop AV1.

The 7-year migration that changed everything

In 2007, Netflix was a monolith running in its own data center. A single database corruption incident in August 2008 took down DVD shipping for 3 days. Engineers couldn't isolate the problem because everything was entangled. That was the turning point.

The migration to AWS and microservices took 7 years (2008-2015). They didn't do a big-bang rewrite. They strangled the monolith — new features went into microservices, while they gradually extracted existing functionality. By 2015, the last monolith service was retired. Today, ~1,000 microservices communicate via gRPC and Kafka, each owned by a small team, each deployed independently 100+ times per day.

In 2025, Netflix expanded its modernization through a zero-configuration service mesh built on Envoy proxies, enabling unified resilience, routing, and observability without application-level libraries. They also adopted GraphQL Federation for more efficient API composition across hundreds of services.

Why Netflix builds its own hardware

Open Connect Appliances (OCAs) are custom-built by Netflix. Each box packs 100-200 TB of SSD storage, optimized for sequential read throughput. Netflix gives these servers to ISPs for free. The deal is straightforward: ISPs get 95% reduction in upstream bandwidth costs (a Stranger Things binge stays on their local network), and Netflix gets single-hop delivery to users.

The economics are decisive. At 15% of global internet traffic, renting CDN capacity from Akamai at market rates would cost billions per year. Building custom hardware costs millions. The back-of-envelope math made the decision obvious — build, don't rent. Today, 18,000+ OCAs sit inside 6,000+ ISPs across 190+ countries. Nearly all video traffic stays within the ISP's network.

Content distribution is intelligent: not every OCA gets every title. Netflix's control plane predicts which titles will be popular in each region and pre-caches them during off-peak hours. A global hit like Squid Game goes everywhere. A niche documentary caches primarily in regions where it's likely to be watched — but also in cities with relevant diaspora communities.

Chaos engineering is a business decision, not a technical one

Netflix's revenue is $39+ billion per year. At 300 million subscribers paying an average of $11/month, one hour of global downtime costs roughly $2 million in direct revenue — plus immeasurable brand damage. The 2008 DVD outage lasted 3 days. That kind of failure at today's scale would be catastrophic.

The Chaos Monkey investment paid off definitively on Christmas Eve 2012. AWS Elastic Load Balancing had a major outage. Services across the internet went down — Reddit, Heroku, and numerous others. Netflix stayed up. Their circuit breakers detected the ELB failures, stopped routing through affected paths, and continued streaming from healthy instances.

Netflix now processes 38 million QoE (Quality of Experience) events per second during live events like the Jake Paul vs. Mike Tyson fight, which streamed to 60+ million concurrent viewers. Their observability platform Atlas processes 17 billion metrics and 700 billion distributed traces daily on 1.5 petabytes of log data.

Why back-of-envelope matters in interviews and architecture

System design interviews at Google, Meta, Amazon, and Netflix all include estimation questions. Not because they want exact numbers — they want to see you think about scale. Can you reason about whether a single server suffices, or whether you need a distributed system? Can you identify the bottleneck before building anything?

The calculator above shows each formula and its derivation. The key insight: start with users, derive everything else. Subscribers → DAU → concurrent streams → bandwidth → storage → API calls → events. Each step multiplies. A 2x increase in subscribers doesn't just mean 2x servers — it means 2x bandwidth, 2x storage, 2x events, and potentially 4x peak load if the growth concentrates in one timezone.