Leaderboard performance problems rarely announce themselves as “data structure issues.” They surface instead as CPU spikes, tail-latency explosions, and on-call alerts that refuse to quiet down. That’s exactly what we encountered: a slice-based leaderboard implementation that initially appeared perfectly reasonable, but began to collapse once the system surpassed the 10,000-user mark and update workloads started behaving like O(N²).
This article walks through what broke, how profiling made the root cause undeniable, and how the issue was fixed by replacing the original approach with an indexed skip list augmented with span counters and a hash-based identity layer. The redesign reduced critical operations to O(log N), stabilized memory usage, and pushed update latency below one millisecond.