DNS Load Balancing and High Availability

Round-robin DNS is easy to set up, and it improves both performance and availability.

The Limits of Round-Robin DNS

Round-robin DNS does not know how busy each server is, or where the user is. It only rotates through the list. This causes several problems.

• Uneven load. A large resolver can serve thousands of users. It may cache one answer and give the same first address to all of them until the TTL expires. So one server can receive far more traffic than the others.
• Distance. A user in Madrid may receive the address of a server in Asia, while a closer server has little traffic.
• Failed servers still get traffic. DNS does not check whether a server is healthy. If one server crashes, its address is still returned in answers. Even after someone removes it, cached copies keep sending users there until their TTL expires. Some clients try the next address after a failure, but many users still see errors.

Smarter DNS Answers

Many DNS services go beyond simple rotation. They choose each answer with more information.

• Health-checked DNS. The DNS service checks each server regularly, for example with an HTTP request every 30 seconds. When a server fails its checks, its address is removed from the answers.
• Weighted DNS. Each address gets a share of the answers. For example, 90 percent of answers point to the current version of a service and 10 percent to a new version.
• Location-based DNS. The DNS server returns the address of the region closest to the user. It usually judges this from the location of the user's resolver. Some resolvers also pass part of the user's IP address, so the answer matches the user's location more closely.

All of these are still affected by caching. A change in the answer reaches users only after the cached answers expire. So services that use DNS for failover usually set short TTLs, like 30 to 60 seconds.

Geographically Distributed DNS Servers

DNS servers themselves can also be spread across many locations, instead of sitting in one data center.

This gives two benefits.

• Faster resolution. Users get answers from a DNS server near them, so each lookup takes less time.
• Redundancy. One location can become unreachable because of a server failure or a network outage. Users can then still be served by DNS servers in other locations.

Every large DNS provider works this way. So do the root servers, which the next section explains.

Anycast Routing

Anycast lets multiple servers in different places share the same IP address.

Here is how it works. Each location tells the internet's routers that it can reach that IP address. Routers share this information with a protocol called BGP (Border Gateway Protocol), which is how networks tell each other which addresses they can reach. When a query is sent to the shared address, the network delivers it to the nearest location. Routers judge this by factors like network distance and availability.

The user does not choose, and the DNS operator does not choose. The network's routing chooses.

Anycast gives three benefits at once.

• Load balancing. Queries are spread across many servers, so no single server becomes a bottleneck.
• Lower latency. Each user reaches a nearby server, which reduces resolution time.
• High availability. If one location fails, it stops announcing the address. The network then automatically sends queries to the next closest location that shares the same IP address. No one has to change any configuration.

Anycast suits DNS very well because a DNS query is usually one short question and one short answer. The 13 root server clusters use anycast, which is how each one can be made of many servers around the world. Public resolvers like 1.1.1.1 and 8.8.8.8 also use it.

The difference from round-robin DNS is important. Round-robin gives out different addresses and does not consider where users are. Anycast gives out one address, and the network delivers each request to a nearby location.

CDNs and DNS

A Content Delivery Network (CDN) is a network of distributed servers that cache and deliver web content to users based on their location. It improves performance, reliability, and security by spreading load across many servers and serving each user from a nearby server.

DNS is how a CDN sends each user to a nearby server. Here are the steps.

1. The website points its name to the CDN with a CNAME record. For example, www.example.com becomes an alias for example.cdnprovider.net.
2. A user's resolver looks up www.example.com and follows the CNAME to the CDN's name.
3. The CDN's own DNS server chooses the best edge server for this user, based on location, server load, and other factors.
4. It returns that edge server's IP address, and the user's browser connects there.

So the routing decision happens during name resolution, before any content is sent. CDNs usually use short TTLs, so they can send users to a different edge server quickly when one is busy or has failed. The What is CDN? lesson covers CDNs in more detail.

Where DNS Load Balancing Stops

DNS is useful for spreading load, but it has limits.

• Caching delays every change. Resolvers, operating systems, and browsers keep answers until the TTL expires, and sometimes longer.
• No control over single requests. DNS picks an address for a lookup, not for each request. One cached answer can be used for thousands of requests.
• Limited knowledge of servers. DNS does not see how busy each server is right now.

So large systems usually combine two layers. DNS, often with anycast or location-based answers, picks the region. Inside each region, a load balancer spreads requests across healthy servers, and it reacts to failures in seconds. The load balancing chapter explains load balancers.

Key Takeaways

• DNS can choose which IP address to return, so it can spread load and avoid failed servers before traffic reaches them.
• Round-robin DNS rotates through several addresses for one name. It ignores each server's real load and the user's location, so traffic can be uneven.
• Health-checked, weighted, and location-based DNS choose answers with more information, but every change waits for cached answers to expire.
• Geographically distributed DNS servers give nearby users faster answers and keep working if one location fails.
• Anycast lets many servers share one IP address. The network routes each query to the nearest server, and it fails over to the next closest one automatically.
• A CDN uses its own DNS server to return the address of the best edge server for each user.
• Large systems use DNS to pick a region, and load balancers to spread requests inside it.

DNS gives every user an address, and the choice of that address is the first place a system can spread load and avoid failures. That completes the DNS chapter. Next is a Flashcards Review of the key terms, followed by the Chapter Assessment.

Practice Questions

Try each question first, then open the answer.

1. example.com uses round-robin DNS with three A records, and one of the three servers crashes. What happens to users, and why is round-robin DNS alone risky here?

<details> <summary>Show answer</summary>

Some users are still sent to the crashed server. Round-robin DNS does not check server health, so it keeps returning the failed address. Even after someone removes the address, cached copies keep sending users there until their TTL expires. Some clients retry another address, but many users see errors. Health-checked DNS or a load balancer avoids this.

</details>

2. A site uses round-robin DNS, and every query gets the addresses in a different order. Why can one server still receive much more traffic than the others?

<details> <summary>Show answer</summary>

Because answers are cached and shared. A large resolver can serve thousands of users. It may cache one answer and give the same first address to all of them until the TTL expires. Round-robin DNS also ignores how busy each server is, and some requests need much more work than others.

</details>

3. A DNS service uses anycast, with the same IP address announced from sites in Tokyo, Frankfurt, and Virginia. The Tokyo site goes offline. What happens to users in Tokyo?

<details> <summary>Show answer</summary>

Their queries automatically reach the next nearest site. When the Tokyo site goes offline, it stops announcing the shared address. The network then routes queries for that address to another site, like Virginia or Frankfurt. Users do not change any settings, and the DNS operator does not change any records. Answers may be slower because the site is farther away.

</details>

4. A website uses a CDN. How does DNS send a user in Madrid to an edge server in Madrid, instead of a server on another continent?

<details> <summary>Show answer</summary>

The CDN's own DNS server chooses the edge server. The website's name is a CNAME alias for the CDN's host name. When the user's resolver looks it up, the CDN's DNS server sees where the query comes from. It returns the IP address of a nearby edge server in Madrid, and the browser connects there.

</details>

5. A team wants to send 10 percent of its users to a new version of a service. Which DNS technique fits, and what is its main drawback?

<details> <summary>Show answer</summary>

Weighted DNS. It returns the new version's address in about 10 percent of answers. The drawback is caching. One resolver's cached answer may serve many users, so the real split is only approximate. Changing the weights also takes effect only after the TTL expires. A load balancer can split traffic per request, much more precisely.

</details>