well, you usually put a load balancer and multiple instances of your service to handle individual server failures. In a basic no-lb case, your single server fails, you restart it and move on (local failure). In a load balancer case, your lb introduces its own global risks e.g. the load balancer can itself fail, which you can restart, but the load balancer can have a bug and stop handling sticky sessions when your servers are relying on it, and now you have a much harder to track brown-out event that is affecting every one of your users for a longer time, it's hard to diagnose, might end up with hard to fix data issues and transactions, and restarting the whole might not be enough.

So yeah, there is no fault tolerance if the timeframe is large enough, there are just less events, with much higher costs. It's a tradeoff.

The cynical in me thinks that the one advantage of these complex CYA systems is that when systems fail catastrophically like CrowdStrike did, we can all "outsource" the blame to them.