Files
wickra/.github/workflows/bench.yml
T
kingchenc aacb9280f1 Honest tiered cross-library benchmark + streaming/batch perf (#186)
## Summary

An honest, tiered cross-library benchmark — and the optimization pass it triggered.

### Performance (wickra-core, outputs unchanged)
Profiling against the other Rust TA crates exposed real inefficiencies. Each
benchmarked indicator is now **5–79% faster** in both streaming and batch:

- **SMA, Bollinger**: flat `Box<[f64]>` ring buffers replace `VecDeque` (−69…79%).
- **RSI**: `100·ag/(ag+al)` collapses three divisions into one; Wilder smoothing
  hoists `1/period` out of the hot path (−46%).
- **ATR**: reciprocal hoisted (−42%).
- **EMA/RSI/ATR**: per-tick `Option<f64>` hot state → bare `f64` + ready flag.

Net result vs `kand`: Wickra now wins **RSI, Bollinger and ATR** (streaming), and
ties `ta-rs` on SMA — up from losing every indicator 1.5–6× before.

### Benchmark harness
New `crates/wickra-bench` (publish=false): a Criterion benchmark comparing Wickra
against `kand`, `ta-rs` and `yata` on an identical BTCUSDT candle series, in
streaming and batch modes. Peer APIs were verified against their source, not
guessed. Wired into the nightly `cross-library-bench` workflow as a separate job.

### Honest README
The benchmark section is rewritten into three layered tables (Rust core vs Rust
crates; Python vs the Python ecosystem) that **show the losses as well as the
wins**. The "only library that combines…" claim is gone; the new framing is
breadth + multi-language reach + the deliberate safety trade-off that costs raw
speed. Added an origin/why-slower rationale and a star CTA.

### Python benchmark
Added `tulipy` runners and expanded per-tick streaming coverage to SMA/EMA/RSI/
MACD/Bollinger. `bench.in`/`bench.txt` now lock `TA-Lib` + `tulipy` (hash-pinned);
`pandas-ta` stays out (it requires Python ≥ 3.12, the bench runs on 3.11).

### Notes
- TA-Lib/tulipy numbers in the README Python table are marked ⧗ — they are
  produced by the CI Linux job (C extensions don't build cleanly on every
  desktop), not measured locally.
- The matching `wickra-docs` prose update is committed separately and will be
  pushed with the release, per the docs-don't-lead-the-registries rule.

Verified locally: `cargo fmt`, `cargo test --workspace --all-features` (3413 core
+ bindings), `cargo clippy --workspace --all-targets --all-features -D warnings`,
Node build + 498 tests, and pytest all green.
2026-06-06 20:57:31 +02:00

146 lines
5.8 KiB
YAML
Raw Blame History

This file contains ambiguous Unicode characters
This file contains Unicode characters that might be confused with other characters. If you think that this is intentional, you can safely ignore this warning. Use the Escape button to reveal them.
name: Cross-library benchmark
# Audit finding R10: previously the cross-library benchmark ran on every push
# and every pull-request to `main`, adding 510 minutes of build + bench time
# per CI run with no consumer of the resulting artefact. The benchmark is now
# scheduled (nightly at 03:00 UTC) and on-demand via `workflow_dispatch`. The
# CI pipeline proper (.github/workflows/ci.yml) still verifies build / tests /
# lints on every push and pull-request.
on:
schedule:
# Nightly at 03:00 UTC. Pick a slot well away from common European /
# American working-hours pushes to keep this off the critical path.
- cron: "0 3 * * *"
workflow_dispatch:
inputs:
size:
description: "Number of bars per indicator (default 20000)"
required: false
default: "20000"
iterations:
description: "Batch iterations per indicator (default 10)"
required: false
default: "10"
# Least-privilege default for the auto-injected GITHUB_TOKEN. The single job
# only builds and uploads an artifact (upload-artifact uses the artifact
# storage API, not the contents scope), so it never needs repo write (OpenSSF
# Scorecard: Token-Permissions).
permissions:
contents: read
env:
CARGO_TERM_COLOR: always
# Network-flake resilience: retry transient registry/DNS failures at the tool
# level so a blip fetching crates.io / PyPI inside any build step (cargo,
# maturin, pip) retries automatically instead of failing the job. Cargo treats
# "couldn't resolve host" / connect / timeout as spurious and retries with
# backoff; 10 attempts ride out a transient DNS blip on a runner.
CARGO_NET_RETRY: "10"
CARGO_NET_GIT_FETCH_WITH_CLI: "true"
npm_config_fetch_retries: "5"
npm_config_fetch_retry_maxtimeout: "120000"
PIP_RETRIES: "5"
PIP_DEFAULT_TIMEOUT: "120"
jobs:
cross-library-bench:
name: Cross-library benchmark report
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
with:
persist-credentials: false
- uses: dtolnay/rust-toolchain@29eef336d9b2848a0b548edc03f92a220660cdb8 # stable branch, 2026-03-27
- uses: Swatinem/rust-cache@e18b497796c12c097a38f9edb9d0641fb99eee32 # v2
continue-on-error: true # cache is an optimisation; never block on a stuck/slow restore
timeout-minutes: 6
- name: Set up Python
id: setup_python
continue-on-error: true
uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405 # v6.2.0
with:
python-version: "3.11"
cache: pip
cache-dependency-path: .github/requirements/bench.txt
- name: Wait before Python retry
if: steps.setup_python.outcome == 'failure'
shell: bash
run: |
echo "::warning::setup-python failed (likely CDN flake), waiting 30s before retry..."
sleep 30
- name: Set up Python (retry)
if: steps.setup_python.outcome == 'failure'
uses: actions/setup-python@a309ff8b426b58ec0e2a45f0f869d46889d02405 # v6.2.0
with:
python-version: "3.11"
cache: pip
cache-dependency-path: .github/requirements/bench.txt
- name: Install Python deps + peer libs
run: |
python -m pip install --upgrade pip
# Hash-locked deps (OpenSSF Scorecard PinnedDependencies). bench.yml
# runs on a single Python version (3.11), so one lock file suffices.
python -m pip install --require-hashes -r .github/requirements/bench.txt
- name: Build Wickra wheel
working-directory: bindings/python
run: maturin build --release --out dist
- name: Install Wickra wheel
working-directory: bindings/python
run: python -m pip install --find-links dist --force-reinstall wickra
- name: Run cross-library benchmark
working-directory: bindings/python
# workflow_dispatch inputs are untrusted; pass them through the
# environment and quote them rather than interpolating into the shell
# command (OpenSSF Scorecard: Dangerous-Workflow).
env:
BENCH_SIZE: ${{ github.event.inputs.size || '20000' }}
BENCH_ITERATIONS: ${{ github.event.inputs.iterations || '10' }}
run: |
python -m benchmarks.compare_libraries \
--size "$BENCH_SIZE" \
--iterations "$BENCH_ITERATIONS" \
--streaming-window 5000 --streaming-iterations 2 \
| tee benchmark.txt
- name: Upload report
uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7.0.1
with:
name: cross-library-bench
path: bindings/python/benchmark.txt
rust-cross-bench:
name: Rust cross-library benchmark report
runs-on: ubuntu-latest
steps:
- uses: actions/checkout@de0fac2e4500dabe0009e67214ff5f5447ce83dd # v6.0.2
with:
persist-credentials: false
- uses: dtolnay/rust-toolchain@29eef336d9b2848a0b548edc03f92a220660cdb8 # stable branch, 2026-03-27
- uses: Swatinem/rust-cache@e18b497796c12c097a38f9edb9d0641fb99eee32 # v2
continue-on-error: true # cache is an optimisation; never block on a stuck/slow restore
timeout-minutes: 6
# Wickra vs the other Rust TA crates (kand, ta-rs, yata) on an identical
# candle series — the like-for-like engine comparison with no binding
# overhead. Streaming + batch, in crates/wickra-bench/benches/cross_lib.rs.
- name: Run Rust cross-library benchmark
run: cargo bench -p wickra-bench --bench cross_lib | tee rust_cross_bench.txt
- name: Upload Rust report
uses: actions/upload-artifact@043fb46d1a93c77aae656e7c1c64a875d1fc6a0a # v7.0.1
with:
name: rust-cross-bench
path: rust_cross_bench.txt