Skip to content

cloudflare/cardinality-estimator

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

71 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

cardinality-estimator

build docs.rs crates.io License

cardinality-estimator is a Rust crate designed to estimate the number of distinct elements in a stream or dataset in an efficient manner. This library uses HyperLogLog++ with an optimized low memory footprint and high accuracy approach, suitable for large-scale data analysis tasks. We're using cardinality-estimator for large-scale machine learning, computing cardinality features across multiple dimensions of the request.

Overview

Our cardinality-estimator is highly efficient in terms of memory usage, latency, and accuracy. This is achieved by leveraging a combination of unique data structure design, efficient algorithms, and HyperLogLog++ for high cardinality ranges.

Getting Started

To use cardinality-estimator, add it to your Cargo.toml under [dependencies]:

[dependencies]
cardinality-estimator = "1.0.0"

Then, import cardinality-estimator in your Rust program:

use cardinality_estimator::CardinalityEstimator;

let mut estimator = CardinalityEstimator::<12, 6>::new();
estimator.insert("test");
let estimate = estimator.estimate();

println!("estimate = {}", estimate);

Please refer to our examples and benchmarks in the repository for more complex scenarios.

Low memory footprint

The cardinality-estimator achieves low memory footprint by leveraging an efficient data storage format. The data is stored in three different representations - Small, Array, and HyperLogLog - depending on the cardinality range. For instance, for a cardinality of 0 to 2, only 8 bytes of stack memory and 0 bytes of heap memory are used.

Low latency

The crate offers low latency by using auto-vectorization for slice operations via compiler hints to use SIMD instructions. The number of zero registers and registers' harmonic sum are stored and updated dynamically as more data is inserted, resulting in fast estimate operations.

High accuracy

The cardinality-estimator achieves high accuracy by using precise counting for small cardinality ranges and HyperLogLog++ with LogLog-Beta bias correction for larger ranges. This provides expected error rates as low as 0.02% for large cardinalities.

Benchmarks

To run benchmarks you first need to install cargo-criterion binary:

cargo install cargo-criterion

Then benchmarks with output format JSON to save results for further analysis:

make bench

We've benchmarked cardinality-estimator against several other crates in the ecosystem:

Please note, that hyperloglog and probabilistic-collections crates have bug in calculation of precision p based on provided probability:

  • incorrect formula: p = (1.04 / error_probability).powi(2).ln().ceil() as usize;
  • corrected formula: p = (1.04 / error_probability).powi(2).log2().ceil() as usize;

We're continuously working to make cardinality-estimator the fastest, lightest, and most accurate tool for cardinality estimation in Rust.

Benchmarks presented below are executed on Linux laptop with 13th Gen Intel(R) Core(TM) i7-13800H processor and compiler flags set to RUSTFLAGS=-C target-cpu=native.

Memory usage

Cardinality Estimators Memory Usage

Table below compares memory usage of different cardinality estimators. The number in each cell represents stack memory bytes / heap memory bytes / heap memory blocks at each measured cardinality.

Our cardinality-estimator achieves the lowest stack and heap memory allocations across all different cardinalities.

Note, that hyperloglogplus implementation has particularly high memory usage especially for cardinalities above 256.

cardinality cardinality_estimator amadeus_streaming probabilistic_collections hyperloglog hyperloglogplus
0 8 / 0 / 0 48 / 4096 / 1 128 / 4096 / 1 120 / 4464 / 2 160 / 0 / 0
1 8 / 0 / 0 48 / 4096 / 1 128 / 4096 / 1 120 / 4096 / 1 160 / 36 / 1
2 8 / 0 / 0 48 / 4096 / 1 128 / 4096 / 1 120 / 4096 / 1 160 / 36 / 1
4 8 / 16 / 1 48 / 4096 / 1 128 / 4096 / 1 120 / 4096 / 1 160 / 92 / 2
8 8 / 48 / 2 48 / 4096 / 1 128 / 4096 / 1 120 / 4096 / 1 160 / 188 / 3
16 8 / 112 / 3 48 / 4096 / 1 128 / 4096 / 1 120 / 4096 / 1 160 / 364 / 4
32 8 / 240 / 4 48 / 4096 / 1 128 / 4096 / 1 120 / 4096 / 1 160 / 700 / 5
64 8 / 496 / 5 48 / 4096 / 1 128 / 4096 / 1 120 / 4096 / 1 160 / 1400 / 13
128 8 / 1008 / 6 48 / 4096 / 1 128 / 4096 / 1 120 / 4096 / 1 160 / 3261 / 23
256 8 / 4092 / 7 48 / 4096 / 1 128 / 4096 / 1 120 / 4096 / 1 160 / 10361 / 43
512 8 / 4092 / 7 48 / 4096 / 1 128 / 4096 / 1 120 / 4096 / 1 160 / 38295 / 83
1024 8 / 4092 / 7 48 / 4096 / 1 128 / 4096 / 1 120 / 4096 / 1 160 / 146816 / 163
2048 8 / 4092 / 7 48 / 4096 / 1 128 / 4096 / 1 120 / 4096 / 1 160 / 207711 / 194
4096 8 / 4092 / 7 48 / 4096 / 1 128 / 4096 / 1 120 / 4096 / 1 160 / 207711 / 194
8192 8 / 4092 / 7 48 / 4096 / 1 128 / 4096 / 1 120 / 4096 / 1 160 / 207711 / 194
16384 8 / 4092 / 7 48 / 4096 / 1 128 / 4096 / 1 120 / 4096 / 1 160 / 207711 / 194
32768 8 / 4092 / 7 48 / 4096 / 1 128 / 4096 / 1 120 / 4096 / 1 160 / 207711 / 194
65536 8 / 4092 / 7 48 / 4096 / 1 128 / 4096 / 1 120 / 4096 / 1 160 / 207711 / 194
131072 8 / 4092 / 7 48 / 4096 / 1 128 / 4096 / 1 120 / 4096 / 1 160 / 207711 / 194
262144 8 / 4092 / 7 48 / 4096 / 1 128 / 4096 / 1 120 / 4096 / 1 160 / 207711 / 194
524288 8 / 4092 / 7 48 / 4096 / 1 128 / 4096 / 1 120 / 4096 / 1 160 / 207711 / 194
1048576 8 / 4092 / 7 48 / 4096 / 1 128 / 4096 / 1 120 / 4096 / 1 160 / 207711 / 194

Insert performance

Cardinality Estimators Insert Time

Table below represents insert time in nanoseconds per element.

Our cardinality-estimator demonstrates the lowest insert time for most of the cardinalities.

cardinality cardinality-estimator amadeus-streaming probabilistic-collections hyperloglog hyperloglogplus
0 0.64 88.12 70.19 82.69 17.45
1 2.42 91.5 80.2 131.86 60.65
2 2.21 44.3 45.34 81.48 34.96
4 6.9 25.59 24.85 54.38 36.22
8 7.27 15.62 17.92 43.54 35.55
16 6.99 12.15 14.44 37.24 33.4
32 7.9 9.6 12.78 34.23 32.49
64 10.14 8.97 11.86 32.55 39.04
128 15.47 8.52 11.49 31.76 48.37
256 13.42 8.01 11.24 31.44 65.58
512 9.92 8.1 11.11 31.34 100.25
1024 8.32 8.14 12.52 31.73 171.71
2048 7.31 7.92 12.52 32.03 120.71
4096 7.11 8.01 11.04 32.73 63.5
8192 8.81 8.02 10.97 33.08 37.36
16384 8.08 8.01 11.03 32.75 22.24
32768 6.55 7.96 11.01 32.37 13.3
65536 5.35 7.96 10.96 31.95 8.41
131072 4.48 7.9 10.97 31.71 5.71
262144 3.91 7.95 10.95 31.52 4.26
524288 3.58 7.64 10.95 31.47 3.47
1048576 3.35 7.95 10.95 31.47 3.04

Estimate performance

Cardinality Estimators Estimate Time

Table below represents estimate time in nanoseconds per call.

Our cardinality-estimator shows the lowest estimate time for most of the cardinalities, especially smaller cardinalities up to 128.

Note, that amadeus-streaming implementation is also quite effective at estimate operation, however it has higher memory usage as indicated by table above. Implementations probabilistic-collections, hyperloglogplus and hyperloglogplus have much higher estimate time, especially for higher cardinalities.

cardinality cardinality-estimator amadeus-streaming probabilistic-collections hyperloglog hyperloglogplus
0 0.18 7.9 15576.4 125.03 24.89
1 0.18 9.19 15619.8 134.3 64.62
2 0.18 9.18 15615.5 134.4 70.51
4 0.18 9.2 15642.7 134.01 89.16
8 0.18 9.19 15611.1 134.41 132.0
16 0.18 9.19 15621.6 134.39 211.4
32 0.18 9.19 15637.1 130.58 357.55
64 0.18 9.19 15626 130.26 619.95
128 0.18 9.18 15640.8 130.33 1134.12
256 11.31 9.09 15668 133.5 2205.7
512 11.3 9.09 15652 129.58 4334.05
1024 11.31 9.09 15687.1 129.79 8392.59
2048 11.28 9.11 15680.4 129.8 8.08
4096 11.29 38.63 15803.4 129.49 4342.07
8192 11.28 38.98 23285 129.51 4345.7
16384 11.29 38.17 26950.7 132.96 4341.9
32768 6.02 10.86 31168 7674.3 4334.98
65536 6.05 4.1 33123.8 40986.4 4327.48
131072 6.02 4.1 33772.4 42113.7 4327.29
262144 6.02 4.11 34711.7 43587 4329.63
524288 6.02 4.1 36091.2 43582.8 4327.8
1048576 6.02 4.11 37877.1 45055.3 4327.37

Error rate

Cardinality Estimators Error Rate

Table below represents average absolute relative error across 100 runs of estimator on random elements at given cardinality.

Our cardinality-estimator performs on par well with amadeus-streaming and hyperloglog estimators, but has especially smaller low error rate for cardinalities up to 128.

Note, that probabilistic-collections implementation seems to have bug in its estimation operation for cardinalities >=32768.

cardinality cardinality_estimator amadeus_streaming probabilistic_collections hyperloglog hyperloglogplus
0 0.0000 0.0000 0.0000 0.0000 0.0000
1 0.0000 0.0000 0.0000 0.0000 0.0000
2 0.0000 0.0000 0.0000 0.0000 0.0000
4 0.0000 0.0000 0.0000 0.0000 0.0000
8 0.0000 0.0000 0.0000 0.0000 0.0000
16 0.0000 0.0019 0.0013 0.0025 0.0000
32 0.0000 0.0041 0.0031 0.0041 0.0000
64 0.0000 0.0066 0.0086 0.0078 0.0000
128 0.0000 0.0123 0.0116 0.0140 0.0000
256 0.0080 0.0097 0.0094 0.0084 0.0000
512 0.0088 0.0100 0.0087 0.0090 0.0000
1024 0.0080 0.0094 0.0101 0.0095 0.0000
2048 0.0092 0.0093 0.0090 0.0107 0.0100
4096 0.0099 0.0108 0.0113 0.0114 0.0103
8192 0.0096 0.0095 0.0131 0.0126 0.0109
16384 0.0116 0.0107 0.0204 0.0229 0.0117
32768 0.0125 0.0109 1.46e14 0.0437 0.0116
65536 0.0132 0.0133 2.81e14 0.0143 0.0118
131072 0.0116 0.0121 1.41e14 0.0128 0.0127
262144 0.0137 0.0144 7.04e13 0.0122 0.0116
524288 0.0138 0.0136 3.52e13 0.0116 0.0121
1048576 0.0113 0.0124 1.76e13 0.0141 0.0110
mean 0.0064 0.0078 3.14e13 0.0101 0.0052