Skip to content

performance.md

Latest commit

 

History

History
62 lines (41 loc) · 2.71 KB

performance.md

File metadata and controls

62 lines (41 loc) · 2.71 KB

Performance Workflow

qs_rust now includes the same high-level snapshot scenarios used by the sibling ports:

  • Encode deep snapshot:
    • depths: 2000, 5000, 12000
    • iterations: 20, 20, 8
  • Decode snapshot:
    • C1: count=100, comma=false, utf8=false, value_len=8, iterations=120
    • C2: count=1000, comma=false, utf8=false, value_len=40, iterations=16
    • C3: count=1000, comma=true, utf8=true, value_len=40, iterations=16

The harness lives in src/bin/qs_perf/main.rs.

Run locally

cargo run --release --bin qs_perf
cargo run --release --bin qs_perf -- --scenario encode --format json

Useful flags:

  • --warmups N
  • --samples N
  • --max-encode-depth N
  • --output /path/to/file

Baselines

Baseline file locations:

These files now contain real captured measurements and serve as the checked-in Rust self-baselines for the current harness shape. Refresh them only from a normal interactive host shell after confirming that qs_perf starts cleanly and emits stable JSON.

If the old local startup failure mode reappears in an agent-driven shell, treat it as an execution-environment problem rather than a trustworthy perf result. In particular, if cargo test -- --list, cargo run --release --bin qs_perf ..., or a direct launch like ./target/debug/qs_perf --help stalls after Cargo prints Running ..., do not recapture baselines from that environment.

Compare command

python3 scripts/capture_perf_baselines.py --scenario all
python3 scripts/compare_perf_baseline.py --scenario all

If a baseline file is still marked pending, the compare script exits with a clear message instead of pretending there is a meaningful threshold to enforce. If qs_perf times out, the compare script now exits with an explicit timeout error instead of hanging indefinitely.

Cross-port snapshots

Use the local orchestration script to run the Rust, Python, Dart, Kotlin, Swift, and C# snapshot harnesses and normalize the output into checked-in artifacts:

python3 scripts/cross_port_perf.py

Artifact locations:

The checked-in snapshot is informative only and should be refreshed on the same machine when you want a new comparison point. If any sibling harness fails, times out, or changes output format, prefer an explicitly partial snapshot over silently dropping that language.