tarwyn

tarwyn is a NetworkTables 4.1 server for FRC robots, written in Rust. AdvantageScope and other NT4 tools connect to it unchanged, and it comes with clients for Rust, Java, Python and C++.

cargo run -p tarwyn_server

use tarwyn_client::Client;

let client = Client::connect("10.4.88.2");
client.subscribe("test", |data| println!("{data:?}"));
client.start();
client.send_bool("test", true);

Values and control go over a WebSocket on port 5810, and telemetry over UDP 5809. There is no authentication, so use --bind 127.0.0.1 to keep the server local. cargo doc --workspace --open builds the API docs.

Latency

Time from a publish to the subscriber's callback on one machine, with default settings:

Rate	tarwyn	ntcore
500 Hz	21 µs	104 µs
50 Hz	43 µs	142 µs

The full results are in bench/RESULTS.md, and bench/BENCHMARK.md explains how they were taken.

Most of that time is thread wakeups. Two flags spend CPU to avoid them, and the client has the same settings as Config::predict and Config::busy_poll. Windows ignores both.

Flag	Default	Does	Costs
--predict <MICROS>	200	wakes the reader just before the next periodic message	~7% of a core at 500 Hz
--busy-poll <MICROS>	0	spins after every message	a whole core

Requirements

	Needs
Server	64-bit Linux (glibc 2.35+), macOS or Windows
Rust client	Rust 1.88+
Java client	JDK 25+ with --enable-native-access
Python client	Python 3.11 to 3.14
C++ client	C++23 and the WPILib 2027 wpimath headers

Releases are built for linux-x86_64, linux-aarch64, windows-x86_64 and darwin-arm64. The roboRIO, musl and 32-bit targets are not supported.

Clients

The C++ header (bindings/cpp/tarwyn.hpp) and the Java library both wrap the C ABI in bindings/c, and Python uses a PyO3 module in bindings/python.

cargo build -p tarwyn-c --release                 # libtarwyn and tarwyn.h
(cd bindings/java && ./gradlew build)             # the jar
(cd bindings/python && uv run --group dev pytest) # the wheel and its tests
cmake -S bindings/cpp -B build/cpp -DWPILIB_INCLUDE_DIR=<wpimath headers>
cmake --build build/cpp && ctest --test-dir build/cpp
cargo xtask package <target> <platform>           # a release zip

Logging

The client can mirror published values to a wpilog file, which AdvantageScope, Elastic and the DataLogTool can open:

client.log_to("/home/lvuser/match.wpilog")?;
let path = client.log_to_drive("match.wpilog")?;  // first writable USB mount

Writing the log never blocks a publish. log_dropped() counts records that did not fit in the queue, and logging_healthy() reports whether writes still succeed.

Channel names that start with TARWYN_INTERNAL are reserved for the server.