Running Safely with Scissors: Lock-free Concurrency for Maximum Throughput
Rust comes with exceptional concurrency primitives, but even a RWLock can be a bottleneck when data changes within microseconds. Under high load such bottlenecks pose a system-wide risk as blocked writers build backpressure and lead to resource exhaustion. In such cases, designing around lock-free, append-only data structures becomes critical for maximizing read and write throughput by guaranteeing that readers and writers never block each other.
This talk explains these principles through a case study: a metrics database backing Shopify's monitoring system, ingesting millions of samples per second per process. We will explore how to leverage composite atomic watermarks and segmented arenas for efficient memory management, as well as generation-tagged bitmaps for sparse validity tracking. These primitives will come together into a system where readers and writers operate uninterrupted without tapping the brakes.
Most importantly, we show how the type system can tame this complexity and enforce single-writer, many-reader semantics at compile time. By safely isolating unsafe code we can expose an ergonomic API and prevent abstraction leaks to the rest of the codebase.
