Replacing a running service with a new version all at once is risky: if the new version is broken, everyone is affected immediately. Staged approaches reduce that risk. One keeps the old and new versions side by side and switches traffic over only once the new one is confirmed healthy, with an instant switch back if not.
Another releases the new version to a small slice of users first, watching closely before widening it. Both share a principle: expose a change to as few people as possible until evidence shows it is safe, so a mistake is caught small rather than large.