DNS TTLs are a budget, not a setting
A long TTL is a loan. It buys you fewer lookups and faster first connections today; you repay it the moment you need to move a record and discover that half the resolvers in the world will keep handing out the old address for another hour.
The usual advice — lower the TTL a day before a migration — works, but only if you lower it before anyone caches the old value. That is the part people skip. The record has to be served with the short TTL for at least one full old-TTL period, or you are just waiting.
A workable policy
- Steady state: 3600 seconds for anything I might need to move, 300 for records that front a service I care about.
- Before a change: drop to 60, wait one old TTL, then change.
- After the change: keep 60 for a day, then raise it back.
Sixty seconds is not about fast failover; it is about being able to undo a mistake in the same coffee. Anything shorter mostly adds query volume without making you safer.
What TTL does not fix
Resolvers are allowed to ignore your TTL. Some do, especially under load, and some client operating systems cache addresses for their own reasons. If you need a record to move instantly, you do not need a smaller TTL — you need a name that was never cached, or an indirection you control.
The time to discover your TTL policy is wrong is not during the incident.
One practical habit: after any DNS change, query the record from two or three public resolvers and from the machine that matters. If they disagree, you are still inside the old TTL window, and no amount of refreshing will help.