Art of Codeartofcode
Lesson 7 of 15

4. DNS

1
The picture to keep in your head
2
4.1 Why do we need names at all?
3
4.2 What are the parts of a domain name?
4
4.3 What happens, step by step, when a name is looked up?
5
4.4 What kinds of DNS records are there?
6
4.5 What is a TTL, and why does DNS cache so much?
7
4.6 How can DNS spread users across servers or send them to the nearest one?
8
4.7 Why do DNS changes "take time to propagate"?
9
4.8 What happens when DNS fails?
10
Memory card and where the analogy breaks
11
Practice: five levels
12
Interview questions
13
Solutions
14
Sources
15
Code lab
4. DNS

4.6 How can DNS spread users across servers or send them to the nearest one?

3 min read+10 XP

In your phone's contacts: A big company gives one helpline number, but the call is connected to the nearest branch. And if one branch is closed, you're sent to another.

In the system: Because the authoritative server chooses the answer, DNS can do simple traffic steering.

Round-robin DNS: give one name several A records. Different users get the list in a different order and mostly connect to the first.

  • Simple and free.
  • But DNS isn't a load balancer. It doesn't know which server is busy or dead, and caches keep handing out a dead server's IP until the TTL runs out. Clients also don't spread perfectly.

GeoDNS (geographic routing): the authoritative server answers based on where the question comes from.

GeoDNS sending a Mumbai user and a Frankfurt user to different regions.

Figure: one name, two answers. Each user is sent to the region closest to them.

  • The "where" is usually the resolver's location, not the user's. A user in Mumbai using a resolver in Singapore may be routed as if they were in Singapore. An extension called EDNS Client Subnet lets resolvers pass along part of the user's network address to fix this, where supported.

Health-checked DNS failover: managed DNS services can check your servers and stop handing out the IP of one that's down. With a short TTL, most users move within minutes.

In practice: DNS picks the region (or the nearest set of load balancers), and real load balancers (Chapter 7) spread traffic inside the region. Many large sites also use anycast, where the same IP address is announced from many locations and the network itself delivers you to the nearest one. CDNs (Chapter 26) rely heavily on both.

Memory code: "DNS picks the city; the load balancer picks the counter." And DNS can't see a dead server until the TTL runs out.