How does enterprise rejection work, such as google.com?

Question

We have several fedora systems that are configured to work on the Internet, FTP, and email. We would like to reflect these services so that we can provide almost 100% reliability for our users. I am a pretty experienced Linux administrator, but I don't have much experience with backup systems.

What is the best way to do this? How do google and amazon do this? Google.com resolves multiple IP addresses, but if my local desktop computer caches one of the IP addresses that are not available, I'm going to get a communication error message. How do they prevent this?

If one of their servers goes down, how is it automatically redirected to another system if the end user never knows about it?

I understand that there are fault-tolerant devices, but they are only intended to abandon the system itself, and not for a complete network.

Let's say we have the worst case scenario, for example, my main system becomes inaccessible. What are the main components that Linux systems use to provide this capability?

I am looking for concepts or approaches, not answers such as "check outstack". What are the actual components that make up the solution? What needs to be done to realize this opportunity?

Answer 1

The published IP addresses of Google or Amazon are not really their servers, but load balancers. The balancer dynamically redirects incoming client connections to the actual servers, and it can be hundreds or thousands of them behind a single IP address.

You may ask, what if the load balancer drops? Well, they usually work with clustering - at least 2, and all have the same virtual IP address. If one is down, the virtual IP is still working because another member of this load balancing cluster is still fulfilling this request.

If the cluster load balancer drops, they automatically change their DNS so that this virtual IP address is no longer advertised to DNS clients. To do this quickly, the typical TTL on these IP addresses is set quite low, a few minutes or less (for google.com it is 5 minutes).

Of course, reality is a little more complicated than that, but it should give you a rough idea.