flowstats

Streaming algorithms for Rust.

Overview

flowstats provides space-efficient probabilistic data structures for processing data streams. These algorithms trade a small amount of accuracy for dramatic reductions in memory usage, making them ideal for:

• Real-time analytics - Count distinct users, track popular items
• Monitoring systems - Compute percentiles, detect anomalies
• Data pipelines - Aggregate statistics across distributed workers
• Resource-constrained environments - Process unbounded streams with fixed memory

Additionally:

• Most structures are mergeable for distributed computation (except Space-Saving)
• It supports no_std environments, see no_std support section below.

Features

Installation

[dependencies]
flowstats = "0.1"

Or with specific features:

[dependencies]
flowstats = { version = "0.1", default-features = false, features = ["cardinality", "frequency"] }

Quick Start

Count Distinct Elements (HyperLogLog)

use flowstats::HyperLogLog;
use flowstats::traits::CardinalitySketch;

let mut hll = HyperLogLog::new(14); // ~0.8% error, 16KB memory

for user_id in user_events {
    hll.insert(&user_id);
}

println!("Unique users: {}", hll.estimate()); // ≈ actual count ± 0.8%

Estimate Percentiles (t-digest)

use flowstats::TDigest;
use flowstats::traits::QuantileSketch;

let mut digest = TDigest::new(100.0);

for latency in request_latencies {
    digest.add(latency);
}

println!("p50: {:?}", digest.quantile(0.5));
println!("p99: {:?}", digest.quantile(0.99));

Find Top-K Items (Space-Saving)

use flowstats::SpaceSaving;

let mut top_k = SpaceSaving::new(100); // Track top 100

for page in page_views {
    top_k.add(page);
}

for (page, count) in top_k.top_k(10) {
    println!("{}: {} views", page, count);
}

Test Set Membership (Bloom Filter)

use flowstats::BloomFilter;

let mut seen = BloomFilter::new(100_000, 0.01); // 1% false positive rate

for item in stream {
    if seen.contains(item.as_bytes()) {
        // Probably seen before (might be false positive)
    } else {
        // Definitely new
        seen.insert(item.as_bytes());
    }
}

Estimate Frequencies (Count-Min Sketch)

use flowstats::CountMinSketch;

let mut cms = CountMinSketch::new(0.001, 0.01); // ε=0.1%, δ=1%

for event in events {
    cms.add(event.as_bytes());
}

let freq = cms.estimate(b"login");
println!("Login events: ≤ {}", freq); // Upper bound on true count

Random Sampling (Reservoir)

use flowstats::ReservoirSampler;

let mut sampler = ReservoirSampler::new(1000); // Keep 1000 samples

for record in unlimited_stream {
    sampler.add(record);
}

// Each record had equal probability of being sampled
let sample = sampler.sample();

Running Statistics

use flowstats::RunningStats;

let mut stats = RunningStats::new();

for value in measurements {
    stats.add(value);
}

println!("Mean: {}", stats.mean());
println!("Stddev: {}", stats.stddev());
println!("Min: {:?}, Max: {:?}", stats.min(), stats.max());

Distributed Aggregation

All sketches support merging for distributed computation:

use flowstats::HyperLogLog;
use flowstats::traits::{CardinalitySketch, Sketch};

// Worker 1
let mut hll1 = HyperLogLog::new(14);
for id in shard_1 { hll1.insert(&id); }

// Worker 2
let mut hll2 = HyperLogLog::new(14);
for id in shard_2 { hll2.insert(&id); }

// Coordinator
hll1.merge(&hll2).unwrap();
println!("Total unique: {}", hll1.estimate());

Algorithm Selection Guide

*t-digest accuracy depends on compression parameter and data distribution.

Feature Flags

[features]
default = ["std", "cardinality", "frequency", "quantiles"]
full = ["cardinality", "frequency", "quantiles", "membership", "sampling", "statistics"]

# Algorithm families
cardinality = []   # HyperLogLog
frequency = []     # Count-Min Sketch, Space-Saving
quantiles = []     # t-digest
membership = []    # Bloom Filter
sampling = []      # Reservoir Sampler
statistics = []    # RunningStats

# Optional
std = []           # Use standard library (disable for no_std)
serde = ["dep:serde"]  # Serialization support

no_std Support

flowstats supports no_std environments that have the alloc crate.

Cargo.toml:

[dependencies]
flowstats = { version = "0.1", default-features = false, features = ["cardinality"] }

Usage:

#![no_std]
extern crate alloc;

use alloc::format;
use flowstats::HyperLogLog;
use flowstats::traits::CardinalitySketch;

// You must provide a global allocator for your target

fn count_distinct() -> f64 {
    let mut hll = HyperLogLog::new(10); // Use lower precision for memory-constrained environments
    
    for i in 0..100u32 {
        hll.insert(&format!("{}", i));
    }
    
    hll.estimate()
}

Requirements:

• Target must support the alloc crate (heap allocation)
• You must provide a global allocator (#[global_allocator])
• libm is included automatically for floating-point math

Parallel Processing

Sketches are not internally synchronized, but they're designed for parallel workloads via the merge pattern:

use std::thread;
use flowstats::HyperLogLog;
use flowstats::traits::{CardinalitySketch, Sketch};

// One sketch per thread, merge at the end
let handles: Vec<_> = data_shards.into_iter().map(|shard| {
    thread::spawn(move || {
        let mut hll = HyperLogLog::new(14);
        for item in shard {
            hll.insert(&item);
        }
        hll
    })
}).collect();

// Merge results
let mut combined = HyperLogLog::new(14);
for handle in handles {
    combined.merge(&handle.join().unwrap()).unwrap();
}

For concurrent access to a single sketch, wrap in Arc<Mutex<_>>.

Error Handling

• NaN values: Silently ignored in t-digest and RunningStats
• Merge errors: Return MergeError for incompatible configurations
• Serde validation: Deserialize validates invariants

Implementation Notes

• Hashing: Uses xxHash (xxh3) for speed. Deterministic within a process but not across versions.
• MSRV: Minimum supported Rust version is 1.75.0
• Stability: Pre-1.0 - API may change between minor versions

Roadmap

• Cuckoo Filter (better than Bloom for deletions)
• Theta Sketch (set operations on cardinalities)
• MinHash (Jaccard similarity)
• DDSketch (relative error quantiles)
• KLL Sketch (alternative quantiles)
• SIMD optimizations

License

Licensed under either of:

• Apache License, Version 2.0 (LICENSE-APACHE or http://www.apache.org/licenses/LICENSE-2.0)
• MIT license (LICENSE-MIT or http://opensource.org/licenses/MIT)

at your option.

References

• HyperLogLog: the analysis of a near-optimal cardinality estimation algorithm
• Computing Extremely Accurate Quantiles Using t-Digests
• An Improved Data Stream Summary: The Count-Min Sketch
• Efficient Computation of Frequent and Top-k Elements in Data Streams