DNS: how a request finds your server

Interviewers ask: "How do you move the site to a new server with no downtime?" Or: "How do far-away users reach a near server?" Both start with DNS.

A weak answer says "DNS gives the address". A strong answer says how long the old address stays in use, and how to make that time short.

Think about a phone book. A helper looks up a number for you and remembers it. If the helper does not know a number, the helper asks the office that owns the book.

Two kinds of DNS server turn a host name into an IP address (lesson 7). A resolver looks names up for you and keeps each answer for a while. An authoritative server holds the official list for your domain.

Records

A record is one line in that list: a name, a type, a value and a TTL. (IN means the internet. Ignore it.)

results.example.in.   300   IN   A   203.0.113.10

A name that has a CNAME cannot have any other record. The plain domain example.in always has NS records, so it cannot be a CNAME.

TTL

TTL (time to live) is the number of seconds that a resolver may keep an answer and reuse it. The record above has a TTL of 300 seconds.

• A long TTL means fewer lookups. The cost: a change reaches users slowly, because old answers stay in use.
• A short TTL means a change spreads fast. The cost: more lookups.

Nobody can take back an answer already given. Some programs keep answers longer, so wait a little longer (our own advice).

Geo routing

With geo routing, the authoritative server picks its answer by where the asker is. Askers in India get a Mumbai server. Askers in Germany get a Frankfurt server. Round trips (lesson 7) get shorter.

The limit: the server sees the address of the resolver, not of the user. A user of a far-away resolver may be placed wrongly. EDNS Client Subnet helps: the resolver shares part of the user's address.

The site results.example.in moves to a new server. Assumptions (made up to teach with): TTL 3,600 seconds.

Step 1: change at once. A resolver that stored the answer at 09:59:59 keeps it until 10:59:59, so the last users move at about 11:00. This answers the predict box. For up to 60 minutes, some users reach the old one.

Step 2: a better plan.

1. At least 3,600 seconds before the move, lower the TTL to 60 seconds.
2. Change the address. Keep the old server running.
3. After a few minutes, switch the old server off and raise the TTL.

Old answers now last at most 60 seconds.

Step 3: the price. Assumption: 5,000 resolvers ask for this name, once per TTL. At TTL 3,600 that is 5,000 / 3,600 = 1.4 questions per second. At TTL 60 it is 5,000 / 60 = 83 (60 times more).

Step 4: geo routing. Assumption: from India, the round trip is 30 ms to Mumbai and 150 ms to Frankfurt. A first request needs 2 round trips and 20 ms of work (lesson 7), not counting the DNS lookup.

• Sent to Mumbai: 2 x 30 + 20 = 80 ms.
• Sent to Frankfurt by mistake: 2 x 150 + 20 = 320 ms (4 times slower).

• A short TTL solves slow change. It costs more lookups. I would use it only around a planned move.
• Geo routing solves long round trips. It costs extra records and some wrong answers. I would change it if many users are placed wrongly.
• DNS solves finding a server by name. It costs slow change: some programs keep old answers longer than the TTL. I would not rely on DNS alone for a switch that must finish in seconds.

"I change the record and everyone sees the new server at once." Stored answers live until the TTL ends: up to 60 minutes here.

"I lower the TTL just before the move." Resolvers hold the old answer for the old TTL. Lower it one old TTL earlier.

1. What are a TTL and a CNAME? What does geo routing do?
2. A TTL is 1,800 seconds. You change the address at 09:00. (a) When do all resolvers that follow the TTL use it? (b) 20,000 resolvers ask once per TTL. How many questions per second at TTL 1,800, and at TTL 30?
3. A team moves on Saturday at 02:00. The TTL is 86,400 seconds. When must they lower it?