Memory ordering in practice (not the spec)

The Rust memory model docs are precise and cold. On a hot path you need a smaller question: what synchronization does this atomic actually guarantee?

This is part three of the lock-free series — after ABA/tagging and reclamation.

The three you use daily

Relaxed

owns: atomicity only

counters, statisticsno cross-thread ordering

Cheapest. Don't use for publishing a pointer others will dereference.

Acquire / Release

owns: publish-subscribe

Release store after writing dataAcquire load before reading it

The default pattern for lock-free handoff.

SeqCst

owns: global order

rare in app codedebugging heisenbugs

Strongest, slowest. Reach for it when weaker orders fail and you need proof.

A handoff that works

handoff.rs

1data.write(payload);2ready.store(true, Ordering::Release);34if ready.load(Ordering::Acquire) {5    let v = data.read();6}

🦀

Ferris' hot tip

If you're mixing atomics and mutexes on the same data, stop — pick one story. Hybrid "mostly lock-free" code is where ordering bugs hide for months.

When to read the spec

Before you ship a general-purpose concurrent crate. Before you argue on Twitter. For application queues and metrics, Acquire/Release + crossbeam gets you home.