Back to work

Global launch readiness and live operations

Just before launch, I caught issues with a parallel AI audit, removed waste on the hot path found in load testing, and ported a predecessor project's kill switch, which turns off features without redeploying. After launch, I support operations with maintenance scenarios and reviews, CS commands and a live data correction process.

Situation

A simultaneous global launch in five languages was set for July 30, 2026, and feature changes kept coming until just before launch. After launch, live updates and regular maintenance repeat, and because the server is event-sourced, live data cannot be edited in bulk in the DB.

Decisions

Launch risk audit: Starting nine days before launch, I ran three rounds of audits with AI agents in parallel, one per area, then checked the main findings again against the actual code and assigned severities. I corrected false positives using commit history and the actual code, and left a check rule that prevents the same misjudgment.

Load testing: Together with AI, I analyzed traces from a staging load test a colleague ran and confirmed the bottleneck was waiting on the DB, not CPU. I cleaned up waste on the hot path, for example by cutting a settings round trip repeated on every query down to once at connection initialization.

Kill switch: I ported the runtime flags and RPC blocking from a predecessor project, adapted to our server's structure, and added them before launch so a feature with a problem can be turned off without redeploying.

Maintenance: Before maintenance, I first write a scenario with the order of work and the checkpoints, and proceed after a teammate reviews it. I documented migrations that depend on deploy order in a guide and a maintenance checklist.

CS and data corrections: I extended the CS commands so the operations team could fix paid equipment, event currencies and pass states, which the existing tools could not fix. Live data is corrected with logic migrations that run when a player connects, and I wrote five versions myself, for reward refunds, content transfers and retroactive application of new criteria.

Cutting back: I had AI go as far as implementing a change to reduce pod memory, but judged its complexity greater than the gain and closed it myself. A warmup warning fired on every boot, and the cause was a loop I had added. I removed the loop because the warmer already warms the same path when it loads user state in an earlier step.

Result

I fixed two reward-handling issues found in the launch risk audit before launch. The game launched globally on July 30, 2026, and live updates and maintenance continue.