Debugging a MongoDB Performance Regression with perf, Flamegraphs, and gdb (4.2.1 vs 4.0.13)
I recently chased down a write-heavy performance regression that showed up when running the same workload on MongoDB 4.2.1 versus MongoDB 4.0.13. The punchline: the “slow” version spent a surprising amount of CPU time rebuilding WiredTiger row-store keys during eviction-driven reconciliation. With a targeted fix to key handling, the regression disappeared. This post walks through the debugging approach end-to-end: how I reproduced the issue, what the flamegraphs revealed, how I used gdb to confirm the execution path, and why the root cause points to a specific key-buffer reset. ...