Module latency_histogram 

Fixed-bucket latency histogram with lock-free inserts + a global registry keyed by (fn, path) tag for emit_timing integration.

Records microsecond latencies into log-scale buckets (powers of 2). 30 buckets covers 1µs to ~1 hour, enough for our actual workload (S3 GET 10-100ms, Metal decode 1-2s, daemon RPC <µs to ms). p50/p90/p99/p99.9 derived via bucket-boundary interpolation.

Why fixed buckets vs HDR or T-digest

• HDR (hdrhistogram crate): more accurate but heavy dep and each instance is ~100KB. We want one histogram per (fn, path) tag, easily 20+ tags → 2MB+ per daemon. Overkill for alpha.
• T-digest: better accuracy at the tails but requires a merge operation that’s not lock-free.
• Fixed log-scale buckets: cheap (240 bytes per histogram), lock-free, perfect-enough percentile accuracy for our alpha “is the tail blowing up?” question.

Each bucket boundary = 2^bucket_idx microseconds. Bucket N covers [2^N, 2^(N+1)) µs. Bucket 0 = 1-2µs; bucket 20 = 1.05s to 2.1s; bucket 30 = ~17min to ~34min.

Use

use wombatkv_node::latency_histogram::LatencyHistogram;
let h = LatencyHistogram::new();
h.record_us(123);
h.record_us(456);
let snap = h.snapshot();
assert!(snap.total_count() == 2);

Structs

HistogramSnapshot

Point-in-time snapshot of a LatencyHistogram.

LatencyHistogram

Lock-free fixed-bucket latency histogram.

Constants

HISTOGRAM_BUCKETS

Number of log-scale buckets. 30 covers 1µs to ~17min.

Functions

emit_snapshot_jsonl

Emit a single MyelonInstr line per (tag, snapshot), the same shape as embed::emit_timing events so existing log consumers can pick it up without schema changes.

record_global

Record a latency observation against the global registry under tag. Creates the histogram on first use. Read-lock-fast path for repeat tags; write-lock only on first-use insert.

reset_all

Reset every histogram in the registry. Useful for periodic-emit windows where each window’s stats are reported then zeroed.

snapshot_all

Snapshot every histogram in the registry.