Festival bus search: hold the surge

The story

Dussehra is two weeks away and a bus platform's route search is about to triple. Travellers compare operators, hold a seat while they pay, and a held seat must disappear from every search at once. The availability database is shared with the operators' reporting tools, so it cannot simply be made huge.

The target

• Load: 2,800 requests a second for 8 seconds, nine in ten of them searches, with a few very hot routes.
• Latency: p99 within 100 ms.
• Budget: $3,400 a month at this load. Half the budget or less earns full cost marks; one and a half times it earns none.
• Faults: three 6-second drills, each with its fault landing 2 seconds in. The search service runs three times slower for 3 seconds, one service instance crashes, and the cache fails over to its replica. A drill passes when under 1% of requests fail and p99 stays within 200 ms.

What is scored

Throughput 25%, latency 30%, cost 15% and resilience 30%, out of 1,000,000. Every entrant runs the same arrival stream; its seed is published when the week closes so anyone can replay a score.

How to enter

Choose Open in Sandbox. The canvas starts from the palette and starter design this challenge is built on. Build, run it against this scenario, then enter. You get five scored entries this week.