All posts
Engineering·21 April 2026·8 min read
What IPL 2026 taught us about concurrency at scale
Peak 184k concurrent viewers on one stream. Here's what broke, what held, and the four ops patterns that kept us live.
AS
Abhishek Sharma
Head of Product
The night that broke our dashboards
IPL 2026 Opening Night. MI vs CSK at Wankhede. Our realtime concurrency graph hit 184,220 at 21:30 and stayed there for four hours. This post is what we learned from keeping that graph flat.
What broke first
The ingest path held fine — Kafka ingestion scaled linearly as we'd rehearsed. What broke was our Grafana dashboard query layer, which had been reading directly off the events table instead of the 1-minute rollup.
The four patterns that held
- Redis sorted sets for live concurrency. Not a MySQL query. Ever.
- Per-ISP ABR ladders. Jio fibre ≠ Airtel ≠ ACT. One ladder doesn't fit.
- Circuit breakers on the payment webhook. Razorpay got slow. We degraded.
- Alerts with runbook URLs attached. Nobody had to think at 10pm.
Tagsconcurrencyscaleiplredis