Crate subms_stats 

subms-stats

Latency-distribution statistics for measurement-heavy workloads. Pure functions on &[u64] sample arrays + an ergonomic SubMsSamples facade.

This is the statistics primitive the subms perf harness uses internally, published as its own crate so you can use it without pulling in the bench machinery. Bring your own Vec<u64> of nanosecond readings (from any source) and compose the analyses you need.

Feature flags

The crate ships a tiny core that’s always on (percentile, mean, stddev) plus optional features that you opt into via Cargo:

• histogram (default) - log2-spaced CDF buckets
• jitter (default) - per-window variance score
• tail (default) - CTE, Hill estimator, fatness ratio
• robust (default) - IQR, MAD, CoV, skewness, kurtosis
• compare (default) - KS statistic, Cohen’s d
• bootstrap (default) - bootstrap CIs for percentiles

Everything is on by default. For a minimal build:

[dependencies]
subms-stats = { version = "0.5", default-features = false }

Then pick only the features you need:

subms-stats = { version = "0.5", default-features = false, features = ["histogram", "tail"] }

Quickstart

use subms_stats::SubMsSamples;

let raw = vec![100u64, 200, 150, 300, 250, 175, 125, 400];
let s = SubMsSamples::new(&raw);
let p99 = s.p99();
let _ = p99;

Re-exports

pub use percentiles::mean;

pub use percentiles::percentile;

pub use percentiles::percentile_sweep;

pub use percentiles::stddev;

pub use samples::SubMsSamples;

pub use bootstrap::bootstrap_percentile_ci;

pub use compare::cohens_d;

pub use compare::ks_statistic;

pub use histogram::cdf_buckets;

pub use jitter::jitter_score;

pub use robust::coefficient_of_variation;

pub use robust::iqr;

pub use robust::kurtosis;

pub use robust::median_absolute_deviation;

pub use robust::skewness;

pub use tail::conditional_tail_expectation;

pub use tail::hill_tail_index;

pub use tail::tail_fatness_ratio;

Modules

bootstrap

Bootstrap confidence intervals. Behind the bootstrap Cargo feature (on by default).

compare

Distribution comparison: “is candidate slower than baseline?” Behind the compare Cargo feature (on by default).

histogram

Log2-spaced CDF buckets. Behind the histogram Cargo feature (on by default).

jitter

Jitter score: was the measurement environment stable across the run? Behind the jitter Cargo feature (on by default).

percentiles

Core percentiles + first-moment statistics. Always available; no Cargo feature gate.

robust

Robust statistics: low sensitivity to outliers, useful when the raw distribution has a heavy tail (which latency typically does). Behind the robust Cargo feature (on by default).

samples

Ergonomic wrapper around a borrowed &[u64] slice of nanosecond latency samples. Method-chained API for the common statistics.

tail

Tail analysis: the bit institutional users actually care about. Behind the tail Cargo feature (on by default).