[nhachicha]

JAVA-6303

Benchmark harness for the OpenTelemetry spec's "Performance Implications" section (mongodb/specifications#1986, DRIVERS-3620). Draft: for discussion of method and the Java configuration mapping before this is wired into Evergreen.

What it adds

• OtelBenchmarkSuite (driver-benchmarks/.../otel/): runs DriverBench "Small doc insertOne" and "Find one by ID", sync and Reactive Streams, under one tracing configuration per JVM, selected by -Dorg.mongodb.benchmarks.otel.config:
  • off: no ObservabilitySettings (baseline)
  • api-only: ObservationRegistry.create() with no handlers; Micrometer short-circuits to a no-op observation, so the driver runs its instrumentation code but nothing records
  • sdk-ratio: OTel SdkTracerProvider with traceIdRatioBased(0.01), bridged via DefaultTracingObservationHandler, no exporter or processor
  • sdk-always: same with alwaysOn()
• InstrumentedBenchmarkRunner: wall time, process CPU time and GC time per iteration; reports MB/s, wall/op, CPU/op, CPU-excluding-GC/op.
• -Dorg.mongodb.benchmarks.otel.verify=true installs a counting span processor and reports spans per op (2.00 under sdk-always, 0.02 under sdk-ratio), for wiring checks only.
• otel-bench.py: one JVM per configuration, configurations interleaved and order rotated per repetition, writes REPORT.md/summary.json; accepts a pre-OTel worktree for the spec's one-time comparison.
• otel-compare-langs.py, otel-compare-runs.py: cross-language and before/after tables.

New benchmark-only deps in the catalog: opentelemetry-bom 1.53.0, opentelemetry-sdk-trace, micrometer-tracing-bridge-otel. Used only by driver-benchmarks; the driver itself is unchanged.

Results (Apple M4 Max, mongod 8.2.6, Temurin 17, 5 interleaved reps, medians)

Loss vs off / added CPU µs per op (excl. GC):

API	Task	api-only	sdk-ratio 1%	sdk-always	off vs pre-OTel commit
sync	Small doc insertOne	+4.2% / +1.6	+9.6% / +4.5	+10.8% / +5.3	+0.8%
sync	Find one by ID	+5.4% / +2.2	+12.2% / +5.6	+13.3% / +6.7	−1.1%
async	Small doc insertOne	+3.1% / +1.7	+5.5% / +4.3	+6.3% / +5.5	−0.4%
async	Find one by ID	+2.6% / +0.7	+6.3% / +4.5	+7.5% / +5.5	+0.6%

Pre-OTel commit is e1bf5f57b1 (parent of #1695). A follow-up draft gates attribute building on the sampling decision and roughly halves the added CPU per op.

Usage

./gradlew :driver-benchmarks:runOtel \
  -Dorg.mongodb.benchmarks.data=<driver-performance-test-data>/ \
  -Dorg.mongodb.benchmarks.output=/tmp/out.json \
  -Dorg.mongodb.benchmarks.otel.config=sdk-ratio

python3 -I driver-benchmarks/otel-bench.py --data <data> --out <dir> --reps 5 [--pre-otel-worktree <path>]

Open points

• Should run as an Evergreen perf variant; these numbers are from a laptop.
• Async tasks block per operation (per-op overhead, not pipelined throughput).
• api-only mapping for a Micrometer driver is a judgement call; raised on the spec PR.