When is std::memory_order_relaxed actually safe to use?

Relaxed atomic operations guarantee atomicity (no data races on the single variable itself) but provide zero synchronization or ordering constraints relative to other memory accesses.

// Safe for independent counters
std::atomic<uint64_t> g_totalPacketCount{0};
void OnPacketReceived() {
    g_totalPacketCount.fetch_add(1, std::memory_order_relaxed);
}

Because other threads only read the counter for telemetry (without relying on it to guard other data), relaxed is 100% safe and saves CPU pipeline synchronization cycles on ARM and RISC-V.

Use relaxed for metrics, telemetry, and statistics. Use acquire-release for flags that synchronize shared buffers.

Clear rule of thumb that prevents over-synchronization in high-throughput engines.