The platform held an ordinary day, but on a popular match the bets arrived in an avalanche and settlement queued up. We wrote a load generator, reproduced that peak hour on a test bench and saw the culprit: settlement ran in a single stream. Once we split it across independent lanes, the rate grew almost fifteenfold, and the fix went to production.
First we built our own load generator: a fleet of up to 274 virtual players behaving like real ones — placing bets, topping up, browsing history.
The ceiling was found using Little’s law: bottlenecks moved from one node to the next until an avalanche of more than 100 000 bets on a single live event exposed the main one — settlement ran through a single handler.
After partitioning by the hash of the identifier, the rate grew from 290 to 4360 bets per minute, and the fix reached production. The runs were checked for more than speed: a financial audit across 123 543 bets found zero double payouts and zero lost credits.