When JioHotstar streamed the 2023 Indian Premier League cricket finals, the platform drew a then-record 32 million concurrent viewers — but the engineering team had built and tested systems capable of handling 50 million to 75 million. Ashutosh Agrawal, who worked on the backend streaming platform since 2018, shared the engineering strategy behind that event at P99 CONF, detailing how his team prepares for high-uncertainty live events where video delivery can't fail. The talk offers lessons for any engineering team facing traffic spikes, from live sports to Black Friday sales.

The team's approach centers on massive over-provisioning: they routinely build capacity for roughly double what they anticipate, creating headroom to absorb unexpected surges without requests stacking up. Authentication systems face their biggest test not during the final match itself but on opening day, when huge numbers of new users or those who haven't logged in for a full year all try to authenticate within the first five to ten minutes. That's where peak concurrency hits hardest, requiring preemptive scaling of critical systems and rigorous load testing of partners. Any partner failing to meet latency service-level agreements gets cut off during the event, with traffic instantly rerouted to the next partner in the hierarchy.

Agrawal described the team's strict latency budgets: any in-house system with no third-party calls must operate within 10 to 15 milliseconds purely server-side for simple lookups, while more complex cases get 40 to 50 milliseconds. "You can sacrifice every other system and component in the line, but anything in the line of video delivery has to be a golden path — it cannot fail," he explained. The team uses a 200-point checklist covering everything from database versions and maintenance windows to timeout configurations between load balancers, services, application servers, and CDNs — most checkpoints stemming from past incidents. One critical lesson came from a service that lacked proper keep-alive timeout configuration: during peak traffic it discarded connections instead of reusing them, causing the load balancer to crash under request load.

Agrawal closed with three recommendations for engineers preparing for high-uncertainty events: first, establish clear principles and get organization-wide alignment on priorities, since conflicts create decision-making bottlenecks during critical moments; second, challenge systems and reject mediocre performance, since computing power has improved massively and server-side latencies shouldn't sit in the hundreds of milliseconds; third, identify the critical path and keep it solid, removing anything that isn't a true top-priority dependency from the synchronous path to the user experience. That might mean saying no to features or re-architecting pieces, but it's the most important step. For platforms where viewer experience depends on split-second synchronization, the margin for error shrinks to nothing — and the difference between flawless delivery and frustrated users comes down to how rigorously teams stress-test their assumptions before the event begins.