Robot Operating System (ROS)

MCAP writes got a lot faster this year, and we can now show it with numbers that compare across every language. A single lidar or high-resolution camera can produce dozens to hundreds of megabytes a second. If the logger can't keep up, the robot drops data or burns CPU it needs for autonomy. Every training dataset and every bug investigation starts as a recording, so the recorder has to keep pace with the sensors. MCAP is the open-source log format we created at Foxglove, and it is the default recording format for ROS 2. Until this year, the C++, Rust, Go, Python, and TypeScript libraries each had their own micro-benchmarks, so nobody could say whether a change made things faster or how the languages compared. We built a shared harness that feeds byte-identical data to every library, including a simulated robot recording with five channels from a 200 Hz IMU up to 512 KB camera frames. Each run checksums its input and counts what it reads back, so a fast result from dropped messages gets caught. Since January, C++ LZ4 writes are about 8x faster on the fixed-payload benchmark, and Rust fixed-payload write time is down by more than half. Batched, SIMD-accelerated CRCs cut chunked write time about 46% in C++ and 31% in Go. The Python ROS 2 decoder uses 90% less memory on nested messages. Most of this arrives with an upgrade, and the post covers where the new compression defaults trade file size against write speed. Several of these changes came from the community, including Cody Smith's zstd context reuse in Rust and Sebastian Lipponer's buffer pre-allocation in C++. Next, we're exploring a faster Python library that wraps the Rust core. Chris Lalancette wrote up the full results: foxglove.dev/blog/measuring-…