Concurrent rendering means rendering is interruptible. It isn't parallelism: React still runs on the main thread, one piece at a time. The difference is that a render no longer has to finish once started. React can:
- pause it every few milliseconds so the browser can handle events and paint;
- resume it later where it left off;
- abandon it and start over when a more urgent update arrives, or keep several versions of the UI in preparation.
That's safe because rendering has no visible effects. React computes the new tree off screen and touches the DOM only in the commit, which is still synchronous and uninterruptible. So the user sees either the old UI or the complete new one. The price is that components must be pure: a render may run several times, or never commit.
Priorities decide what gets interrupted. Each update gets a priority (a lane). Discrete input (click, keydown, input) is the most urgent, then continuous input (mousemove, scroll), then default updates (timers, network callbacks). Those render without yielding, as before. Transitions, useDeferredValue re-renders, Suspense retries and hidden <Activity> content are non-urgent: React renders them in slices, and a click or keystroke arriving meanwhile is handled first, after which the stale work restarts.
It's opt-in per update. React 18 dropped the idea of a "concurrent mode" for the whole app. Every createRoot root can render concurrently, but by default updates behave like React 17. Concurrent rendering is used only where you use a concurrent feature, usually through a router or framework that wraps navigations in startTransition.
The user-visible result: typing and clicking stay responsive while a large, non-urgent update is prepared in the background, and intermediate states that were superseded are never shown.