Six million concurrent viewers, and a grid of Gatling generators
The company where I enjoyed my time most as a performance engineer. Flood.io, before Tricentis bought it, and the test that made it famous.
Flood's model: upload, scale, go
Flood competed with BlazeMeter and Loadero, and its model was simple and powerful: you upload your scripts, Gatling, Selenium, Playwright, JMeter, and you run them with however many instances you need.
A single instance, an M5 XLarge with 4 CPUs and 16 GB of RAM, could launch 500 JMeter users, sometimes a thousand depending on the logic. The tooling was far more optimized for load than anything I had used before.
The Hotstar test
For the Cricket India Final, they reserved multiple AWS sites, A through D depending on region, with 500 to 1000 load generators per site just for the load side. And that is only the generators for the test itself — not the infrastructure that handled the actual streaming, which scaled by device type because every renderer behaves differently.
No controller. Every node is a generator.
Gatling ran the show, and I loved the architecture: no controller box, every node was a load generator, everything aggregated back into Flood. The granularity was coarse, fifteen second buckets, but it was enough to understand what was happening at that scale.
The human side
My job was coaching. Clients whose scripts failed, developers, architects, and performance engineers learning how to get a good result out of the tool. Helping people run tests that actually meant something.