blackbox-bench

A lightweight Python microbenchmarking library with a Rust core (PyO3 + maturin). Designed as a "criterion for Python" — minimal harness overhead, statistically rigorous output, and a small CLI that drops into CI.

import blackbox_bench

bench = blackbox_bench.Bench()

@bench.benchmark
def hash_1kb():
    import hashlib
    hashlib.sha256(b"x" * 1024).digest()

bench.run()
bench.report()                # human-readable table
bench.report(format="json", path="results.json")

blackbox-bench results
─────────────────────────────────────────────────────────────────────────────────
Name        Mean   Median  StdDev    Min    Max  Ops/sec          CI 95%  Outliers
─────────────────────────────────────────────────────────────────────────────────
hash_1kb  1.4 µs   1.4 µs  12.0 ns 1.4 µs 1.5 µs  710,221  [1.4 µs, 1.4 µs]        3
─────────────────────────────────────────────────────────────────────────────────

• Per-batch timing with auto-batch-sizing — sub-microsecond functions are batched until each sample takes ≥5 µs, well above the timer resolution floor.
• Per-iteration overhead measurement and subtraction — the harness probes itself at startup so reported nanoseconds attribute time to your code, not blackbox-bench.
• Bootstrap 95% confidence intervals for the mean, plus Tukey or MAD outlier detection.
• blackbox_bench.black_box(value) — opaque pass-through that the optimiser can't see through.
• bench.iter_batched(setup, routine) — setup runs once per sample (untimed); routine runs the timed batch.
• throughput= — @bench.benchmark(throughput=1024) reports MB/s alongside ops/sec.
• params=[...] — parameterised benchmarks; one result per parameter.
• Opt-in HDR histograms — Bench(histogram=True) attaches a percentile-queryable HdrHistogram to each result.
• blackbox_bench.compare(baseline_json, current_json) — classifies each row as unchanged / regressed / improved / new / removed using CI overlap, not raw change_pct.
• Four reporters — table, JSON, self-contained HTML (with inline SVG sparklines), JUnit-compatible XML (with a raw alternative).
• blackbox-bench run --profile — wraps each benchmark in py-spy and emits SVG flamegraphs alongside the results.

pip install blackbox-bench

Pre-built abi3 wheels are published for cpython 3.10 / 3.11 / 3.12 / 3.13 on linux (x86_64 + aarch64), macOS (x86_64 + aarch64), and windows x86_64. Building from source requires a Rust toolchain.

Optional extras:

pip install blackbox-bench[profile]   # adds py-spy for `blackbox-bench run --profile`

import blackbox_bench

bench = blackbox_bench.Bench(
    warmup=5,
    target_time_ns=1_000_000_000,
    outlier_method="tukey",        # or "mad" / "none"
    overhead_subtract=True,
    histogram=False,
)

@bench.benchmark
def quick(): sum(range(100))

@bench.benchmark(name="hash_kb", throughput=1024, params=[10, 100, 1000])
def hashing(n):
    import hashlib
    hashlib.sha256(b"x" * n).digest()

results = bench.run()
bench.report(format="html", path="results.html")

import random

@bench.benchmark
def sort_random():
    return bench.iter_batched(
        setup=lambda: random.sample(range(1_000), 1_000),
        routine=lambda xs: sorted(xs),
    )

setup runs once per sample and isn't timed; routine is the timed call inside the batch.

with bench.measure("payload_build"):
    payload = build_huge_payload()

bench.run()  # appends measured contexts to results

# bench_hashing.py
import blackbox_bench

@blackbox_bench.benchmark
def sha256_1kb():
    import hashlib
    hashlib.sha256(b"x" * 1024).digest()

blackbox-bench run bench_hashing.py --warmup 5 --iterations 100
blackbox-bench run benches/ --format html --output report.html
blackbox-bench run benches/ --save baseline.json
blackbox-bench compare baseline.json current.json   # CI-classified diff

import json
report = blackbox_bench.compare(
    open("baseline.json").read(),
    open("current.json").read(),
)
for row in report.rows:
    print(row.name, row.classification, row.change_pct)

ComparisonReport.format("xml") emits JUnit with <failure> on regressed rows — drop the file into Jenkins/GitHub Actions test reporters.

The harness has to stay small relative to the user's function:

• Tight sampling loop — Instant::now() and the per-batch call dispatch are raw FFI (PyObject_CallNoArgs / PyObject_CallObject), skipping PyO3's higher-level wrappers inside the timed window.
• GIL released during stats — bootstrap CI, Tukey/MAD, and the histogram run inside py.detach(...) so other Python threads aren't blocked while blackbox-bench crunches its samples.
• Reused scratch buffers — median, tukey, mad, and bootstrap_ci_mean share two Vecs owned by the Runner; a 100-benchmark suite still allocates only once for the lot.

The criterion benchmarks at benches/rust_internals.rs measure these primitives directly. benches/bench_dogfood.py measures the assembled harness end-to-end (the empty-pass benchmark should report ~0–1 ns after overhead subtraction).

See MIGRATION.md. Most v0.1.0 code only needs an import swap; the one surface that changed without alias is Bench.report(json_output=True) → Bench.report(format="json"). For exact v0.1.0 semantics:

from blackbox_bench.legacy import Bench, benchmark, BenchmarkResult

The legacy shim is removed in v1.1.

MIT — see LICENSE.