The problem
The studio had built the game in nine months and had never run a load test. Marketing promised 6× the expected concurrency on launch day — matchmaking, presence and economy services that had to hold without a single incident ticket.
What we did
We built the backend as independent Go services behind a gateway: matchmaking as a reservation system, presence on Redis streams, and a currency-ledger economy that's auditable to the cent. Kafka carries every game event into ClickHouse for product analytics.
Launch readiness was treated as an engineering discipline: load-tested to 150% of the promised spike, a load-shedding path that degrades gracefully (queue the player, never drop them), and a runbook rehearsed the week before with the studio on the call.
Impact
Launch day absorbed the spike with zero incidents. The telemetry pipeline then told product exactly where players quit — matchmaking p95 held at 28ms, and the studio shipped its first live-ops event on time.