Module shared_hyper_log_log 

SharedHyperLogLog - cross-process probabilistic distinct-count estimator.

2^p AtomicU8 registers, each storing the maximum-observed rank (leading-zero count + 1) of items hashed to that register. Estimate via harmonic mean with bias correction.

Why this is safe

Each insert is exactly one fetch_max on one AtomicU8. No CAS loops, no Drop guards, no spin waits. The cache-line contention is bounded to one register per insert.

Accuracy

Standard error ~= 1.04 / sqrt(m) where m = 2^p.

p	m	std err	size
8	256	6.5%	256B
10	1024	3.3%	1 KB
12	4096	1.6%	4 KB
14	16384	0.8%	16 KB
16	65536	0.4%	64 KB

Encoding

Hash item to u64 h. Register index = top p bits of h. Rank = (leading_zeros of (h << p) | (1 << (63-p))) + 1, clamped to 64. (The OR ensures rank is bounded even when low bits are 0.)

Structs

HLLHeader

SharedHyperLogLog

Enums

HLLError

Constants

HLL_MAGIC

MAX_PRECISION

Maximum precision (larger = more memory).

MIN_PRECISION

Minimum precision (smaller = less memory but worse accuracy).

Functions

hll_file_size