-
Type:
Task
-
Resolution: Fixed
-
Priority:
Major - P3
-
Affects Version/s: None
-
Component/s: None
-
None
-
DB Integration & Observability
-
Fully Compatible
-
None
-
None
-
None
-
None
-
None
-
None
-
None
Goal
Catch regressions in the cost of tracing for untraced requests, at the default sampling level, and for traced requests.
Benchmark support (prototype on calvin.nguyen/otel-perf-gate and otel-cheap-fixes)
- otel::traces::BenchmarkTracing modes: kOff, kUntraced, kDefaultSampling (documented factor), kTraced (export on the request thread), and background-export modes that match production. Export-thread CPU and dropped spans are reported separately.
- Fixtures: crud_bm BM_FIND_ONE/BM_INSERT_ONE for every mode; BM_SEP_PING router/shard untraced/traced.
- GenerateAndCheckPerfResults relative_to/max_ratio support for cross-fixture ratios.
Remaining
- Thresholds in performance_thresholds.yml (owned by the perf team) based on CI instruction counts: untraced = off; default sampling < 2%; traced < 10% (background-export mode).
- Use ping only as a fixed per-span cost gate, not a percentage target.
Measurement notes
- Compare modes within one binary: code layout alone moved baselines by up to 4 us between builds.
- On ARM64, use frame-pointer builds and perf for profiles; callgrind's call graph is unreliable on optimized builds (identical code folding).