Tagged articles

controlled experiment

1 articles · Page 1 of 1
Huolala Tech
Huolala Tech
Oct 10, 2026 · Backend Development

How Flame Graphs Cut 17% CPU at 200k QPS: Logging Stack Traces & Map Scans

At 200k QPS, async-profiler flame graphs revealed that synchronous stack-trace capture on every log line and full-table scans of a fee-item cache consumed 25% of CPU; removing the stack trace and indexing the cache by code dropped single-node CPU from 65% to 48% with no RT regression, validated by controlled ablation experiments and production rollout.

CPU optimizationConcurrentHashMapJava performance
0 likes · 13 min read
How Flame Graphs Cut 17% CPU at 200k QPS: Logging Stack Traces & Map Scans