flink-rs

Stream processing in Rust, with a Flink-inspired DataStream API.

flink-rs helps you build typed streaming pipelines and explore familiar DataStream patterns in Rust. The workspace includes a local execution engine, time-based windows, and runnable workloads from streaming benchmarks.

Get started

Run the complete Nexmark query suite with a deterministic auction stream:

cargo run --release -p flink-rs-nexmark -- --query all --events 100000

Each query prints its input count, output count, elapsed time, throughput, and checksum. See the workload guide for more ways to explore the project.

Build a stream

use flink_rs_streaming::ExecutionEnvironment;

fn main() -> Result<(), Box<dyn std::error::Error>> {
    let environment = ExecutionEnvironment::new();
    let result = environment
        .from_collection([1, 2, 3, 4])
        .filter(|value| value % 2 == 0)
        .map(|value| value * 10)
        .collect();

    environment.execute()?;
    assert_eq!(result.values()?, vec![20, 40]);
    Ok(())
}

What you can explore

• Typed pipelines: sources, maps, filters, flat maps, unions, and sinks.
• Keyed streams: per-key reductions and rich user functions with lifecycle hooks.
• Time-based processing: event-time and processing-time windows, sessions, triggers, allowed lateness, and late-data side outputs.
• Local execution: bounded and polling sources with watermarks, timers, and validated execution plans.
• Runnable workloads: Nexmark, Yahoo Streaming Benchmark, DSPBench, DEBS 2014 and 2022, Linear Road, Theodolite, TPCx-IoT-style, and Flink JMH ports.

Explore the API, run the workloads, and follow upcoming work on the roadmap.

Local benchmark snapshot

A Criterion run on Apple Silicon (100,000 records per case, 2026-11-11):

Workload	Million records/s
Global windows	3.76
Tumbling windows	3.70
Sliding windows	0.86
Session windows	1.18
Tuple key-by	11.23
Array key-by	8.88

Run the full local suite with:

cargo bench -p flink-rs-flink-jmh --bench flink_jmh_port

Set FLINK_RS_JMH_FILTER to a workload name, such as tumblingWindow, to focus on one case. See the benchmark guide for full-size options and details about the measurement setup.

Project

• DataStream API guide
• Workload guide
• Roadmap
• Source code

Licensed under MIT or Apache-2.0.