
Scaling for 50 Million Concurrent Viewers: How Hotstar Mastered Tsunami Traffic on AWS
Discover how Disney+ Hotstar engineers highly resilient infrastructure to survive sudden "Tsunami Traffic" during IPL matches. This deep dive explores their custom predictive autoscaling, Datacenter Abstraction across Amazon EKS, and the AWS Elemental video pipeline that guarantees sub-10 second latency for over 50 million concurrent fans.
Scaling for 50 Million Concurrent Viewers: How Hotstar Mastered Tsunami Traffic on AWS during IPL
Tags:
#AWS #EKS #CloudArchitecture #DevOps #Hotstar #AWSBuilderCenter #SystemDesign Executive Summary
Streaming live sports to tens of millions of fans simultaneously is the ultimate stress test for any cloud infrastructure. For Disney+ Hotstar, broadcasting the Indian Premier League (IPL) isn't just about managing high traffic; it's about surviving sudden, catastrophic surges in concurrency known as "Tsunami Traffic."
During the IPL, Hotstar routinely breaks global streaming records, peaking at over 50 million concurrent viewers. This case study breaks down how Hotstar engineered a highly resilient, globally distributed architecture on AWS. By bypassing standard autoscaling limits and leveraging Amazon EKS, EC2 Spot Instances, the AWS Elemental Suite, and a custom Datacenter Abstraction layer, they achieved sub-10 second latency at a scale previously thought impossible.
1. The Engineering Challenge: "Tsunami Traffic" & The Dhoni Effect
The traffic pattern during a high-stakes IPL match is entirely unpredictable. A star player like M.S. Dhoni walking onto the pitch can cause a concurrency spike from 5 million to 25 million viewers in a matter of seconds.
Conversely, when a star gets out, millions of fans immediately exit the live stream and flock back to the homepage or personalized content recommendations. This sudden drop in the video stream triggers a massive, instantaneous shockwave of read/write requests across auxiliary microservices—often referred to as the "Thundering Herd" problem.
The Problem with Standard Autoscaling:
Traditional CPU/Memory-based scaling relies on standard metrics (like Amazon CloudWatch). By the time these metrics are collected, aggregated (typically 1-5 minutes), and evaluated to trigger an autoscaling group (ASG) to spin up new EC2 instances, the traffic spike has already overwhelmed the system.
2. The Architectural Deep-Dive on AWS
To survive these Tsunamis, Hotstar discarded traditional scaling rules and built custom abstractions tailored for extreme burst traffic.
Custom Predictive Autoscaling
Instead of reactive scaling, Hotstar built a predictive autoscaling engine. This engine monitors real-time request velocity rather than infrastructure metrics. If the rate of incoming HTTP requests accelerates past a predefined threshold, the system preemptively provisions massive compute capacity before the CPU utilization even registers a spike.
Datacenter Abstraction across Amazon EKS
Kubernetes has inherent limits on cluster size, API server throughput, and VPC IP address availability. To handle 50M+ users, Hotstar could not rely on a single Amazon EKS cluster. Instead, they implemented a Datacenter Abstraction layer.
The Datacenter Abstraction layer treats multiple independent EKS clusters in a region as one massive, unified pool of compute. This allows for seamless traffic distribution and completely bypasses individual cluster limits and IP exhaustion issues.
Microservices & Isolated Blast Radius
To prevent the "Dhoni drop" from bringing down the entire platform, Hotstar rigorously separated concerns. Playback, user profile, recommendations, and analytics are isolated microservices. If the homepage recommendation engine crashes under load, the video playback microservice remains entirely unaffected.
3. The Broadcast-Grade Video Pipeline
Ingesting, encoding, and distributing live video at this scale requires a pipeline that guarantees zero packet loss and ultra-low latency. Hotstar utilizes a potent mix of AWS services to manage the load cost-effectively.
- Compute (Amazon EKS & EC2 Spot): The backbone of the platform. Hotstar leverages EC2 Spot Instances heavily. By intelligently distributing workloads across multiple instance families and availability zones, they achieve massive scale while optimizing costs, falling back to On-Demand instances only when Spot capacity drops.
- Video Pipeline (AWS Elemental Suite): The live feed ingestion relies on AWS Elemental to guarantee broadcast-grade reliability.
4. Real-Time Analytics & Telemetry
To power their predictive autoscaling and real-time dashboards, Hotstar needs to process billions of events per minute.
They built a massive event-streaming architecture using Apache Kafka (managed via Amazon MSK) and Apache Flink (via Amazon EMR). Every play, pause, buffer, and heartbeat from the client apps is streamed into Kafka. Flink processes these streams in real-time to detect anomalies, update live viewership counters, and trigger the predictive autoscaling engines.
5. Caching & State Management
Databases are often the first bottleneck during a traffic tsunami. Hotstar minimizes direct database hits by implementing aggressive, multi-layered caching.
- Amazon ElastiCache (Redis): A sub-millisecond Redis layer manages leaderboards, real-time player stats, and stream metadata.
- NoSQL Datastores: For persistent state that requires high write-throughput (like user session tracking), they utilize highly scalable NoSQL solutions (like DynamoDB or ScyllaDB) rather than traditional relational databases.
6. Behind the Scenes: AWS Networking & Traffic Routing
Handling a 50 million user tsunami requires extreme optimization at the networking layer, long before traffic hits the compute nodes.
- Amazon Route 53 & DNS Routing: Hotstar uses Route 53 to globally manage DNS. During an IPL match, they heavily utilize weighted routing and latency-based routing policies to evenly distribute the initial surge of traffic across their multi-region setup, preventing any single AWS region from taking the full brunt of a spike.
- Load Balancer Pre-Warming: While standard AWS Application Load Balancers (ALB) scale elastically, they need time to scale up their underlying nodes. To prevent dropped TCP handshakes during an instant tsunami, Hotstar works closely with AWS enterprise support to "pre-warm" their load balancers to their peak expected capacity before the match even begins.
- Overcoming IP Exhaustion: When tens of thousands of Kubernetes pods spin up in minutes, standard Amazon VPC CNI setups can rapidly exhaust available IP addresses within a subnet. Hotstar manages this by utilizing secondary CIDR blocks and custom networking configurations, ensuring that IP address scarcity never throttles their predictive scaling engine.
7. Project HULK: Chaos Engineering at Scale
You cannot test for a 50 million user spike by waiting for match day. Hotstar built Project HULK, an internal chaos engineering and load testing framework.
HULK runs on AWS and simulates synthetic tsunamis, generating millions of artificial requests per second from distributed AWS regions. It intentionally terminates EC2 nodes, introduces network latency, and takes down entire availability zones to ensure the architecture heals itself before real viewers are ever affected.
8. Key Metrics & Business Impact
The results of this robust AWS architecture are staggering:
- Peak Concurrency: Handled over 50 million concurrent streams smoothly without degradation.
- Latency: Maintained sub-10 second glass-to-glass latency, ensuring viewers don't get spoilers from their neighbors or social media.
- Reliability: Achieved 99.99%+ uptime during critical match moments.
- Cost Efficiency: Saved millions of dollars by aggressively utilizing Amazon EC2 Spot Instances during peak loads.
9. Key Takeaways for Cloud Builders
For cloud architects and DevOps engineers building on AWS, the Hotstar IPL case study offers critical lessons:
- Don't rely solely on reactive metrics: For unpredictable workloads, build predictive scaling based on request velocity and business metrics.
- Abstract your boundaries: Don't let single-cluster Kubernetes limits dictate your scale. Abstract multiple EKS clusters into unified computing pools.
- Embrace Chaos Engineering: If you don't break your infrastructure in testing, your users will break it in production. Test at true scale using tools like Project HULK.
- Leverage Spot Instances smartly: Massive compute doesn't have to mean massive bills. Design a resilient architecture that can handle instance interruptions gracefully.
- Isolate your blast radius: Ensure your heavy read/write auxiliary services (like recommendations) cannot impact your core business pipeline (video streaming).
Enjoyed reading this content? Let the author know!
Your likes, comments, shares, and saves help creators reach more builders.
Loading recommendations
Loading article