How Hotstar Handles Crores of Live Viewers

India vs New Zealand, 2019 World Cup semi-final. At the peak, 25.3 million people were watching live on Hotstar at the same time, a world record then.

Live cricket is one of the hardest traffic problems in tech. A web series is watched over weeks. A match is watched by everyone at the same minute, and the traffic jumps whenever something big happens on the field.

This post explains how Hotstar handled it, using what its own engineers presented at AWS re:Invent 2019 ("Scaling hotstar.com for 25 million concurrent viewers").


🎯 Quick Answer (30-Second Read)

  • Peak: 25.3M concurrent viewers, 10 Tbps+ bandwidth, 1M+ requests per second.
  • The real problem was speed: traffic grew by over 1 million users per minute, but a new server took about a minute to boot.
  • So they didn't rely on autoscaling. They pre-warmed infrastructure before the match.
  • They scaled on concurrency and request count, not CPU.
  • Project HULK load-tested a fake match before the real one.
  • Panic mode: keep critical services up and switch off non-critical ones under extreme load.

The Scale in Numbers

25.3M
peak concurrent viewers
10 Tbps+
peak bandwidth
1M+
peak requests per second
10B+
clickstream messages

How fast the records climbed (from Hotstar's slides):

xychart-beta title "Peak concurrent viewers on Hotstar (millions)" x-axis ["IPL 2018 final", "IPL 2019 final", "WC 2019 IND vs NZ"] y-axis "Millions" 0 --> 30 bar [10.3, 18.6, 25.3]

The Real Problem: A Race Against Time

Hotstar's engineers called it a race against time:

  • Traffic can grow by more than 1 million users per minute, often triggered by push notifications ("India batting!").
  • A new application server takes about a minute to boot.
  • They needed a 90-second reaction time, using fully baked machine images so new servers start ready to serve.

If you wait for servers to look busy before adding more, the wave has already hit.


Why They Didn't Rely on Autoscaling

Hotstar's slides list why standard cloud autoscaling didn't fit live cricket:

  • Insufficient capacity errors: the cloud region may not have enough of a machine type at that moment
  • One instance type per autoscaling group, which limits flexibility
  • Step sizes too small for a sudden flood
  • Retries and exponential backoff that slow scaling down
  • Availability-zone imbalances ("game of zones")
Key takeaway: autoscaling reacts to load that has already arrived. For predictable, massive spikes, you have to prepare before the spike.

Their Battle-Tested Scaling Strategy

  1. Pre-warm infrastructure before match time. Capacity is in place before the first ball.
  2. Scale on request count or platform concurrency, not CPU. Concurrency shows the real demand earlier.
  3. Proactive, automated scale-up based on concurrency.
  4. EC2 Spot Fleet with diversified instance types, so one unavailable machine type doesn't block scaling.

For how traffic gets spread across those servers, see Load Balancer Explained.


Project HULK: A Fake Match Before the Real One

Before big games, Hotstar ran "Game Day" exercises with an internal tool called Project HULK:

HULK capability Scale
Load generation 108,000 CPUs
Memory 216 TB RAM
Network 200 Gbps outbound
Geography 8 regions

It ran load tests, performance and "tsunami" tests, chaos engineering, and traffic patterns modelled with machine learning.

What they looked for: each system's breaking point, the "death wave", bottlenecks and choke points, hidden patterns, and failures at the network and application level. The "chaos" they simulated included push notifications, tsunami traffic, delayed scale-up, bandwidth limits, higher latency and CDN network failures.


Panic Mode: Fail Gracefully

Even with all that preparation, Hotstar planned for overload with panic mode:

  • P0 services must always stay up, the core of watching the match
  • Graceful degradation: turn off non-critical services to protect the critical path
stateDiagram-v2 [*] --> Normal Normal --> PanicMode: load beyond safe capacity PanicMode --> Normal: load back under control state PanicMode { [*] --> P0_up P0_up: Critical services stay up NonCritical_off: Non-critical features switched off P0_up --> NonCritical_off }

Their stated lesson: "Okay to degrade gracefully." Users forgive a missing feature far more readily than a frozen stream.


What Every Engineering Team Can Learn

Hotstar's own key takeaways were:

  • Prepare for failures
  • Understand your user journey
  • It's okay to degrade gracefully

Those apply far beyond cricket: product launches, sales days, exam results, anything where traffic arrives all at once.


The Numbers Keep Growing

The 2019 record didn't last. In the 2023 IPL final, JioCinema reported over 3.2 crore (32M) concurrent viewers. In the 2025 Champions Trophy final, JioHotstar reported a peak of 61.2 million. The same principles apply at that scale: prepare early, test the breaking points, and protect the core experience.


Frequently Asked Questions

How many people watched the 2019 India vs New Zealand match on Hotstar?

Hotstar reported a peak of 25.3 million concurrent viewers during the 2019 World Cup semi-final between India and New Zealand. That was a world record for live streaming concurrency at the time.

Why didn't Hotstar use autoscaling?

Traffic grew by over a million users per minute, while new servers took about a minute to boot. Hotstar's engineers also cited capacity errors, step sizes and availability-zone issues. So they pre-warmed infrastructure before matches and scaled proactively on concurrency.

What is Project HULK?

Project HULK was Hotstar's load-testing and game-day platform. It generated traffic using 108,000 CPUs and 216 TB of RAM across 8 regions to run load, tsunami and chaos tests, finding breaking points before real matches.


Conclusion

Hotstar handled 25.3 million concurrent viewers not with a single clever trick, but with preparation: pre-warmed capacity, scaling on concurrency instead of CPU, rehearsing the match with Project HULK, and a panic mode that sacrifices extras to keep the stream alive.


Sources

Last verified: 14 October 2026.


Related reads: Load Balancer Explained · Why Apps Crash During High Traffic · Rate Limiter Explained · Why Netflix Rarely Goes Down

I write about how India's tech gets built. Follow on LinkedIn or X, or invite me to speak.