DNS

One-liner: Hierarchical, cached, eventually consistent directory mapping names to records; also the coarsest layer of load balancing and failover.

Say this#

SAY THIS

Recursive resolver — the client-side resolver (ISP, 8.8.8.8, 1.1.1.1) that does the lookups and caches.
Root → TLD → authoritative — the delegation chain: . → .com → ns1.example.com.
Authoritative nameserver — holds the zone (Route 53, Cloudflare DNS, NS1, bind).
Record types — A (IPv4), AAAA (IPv6), CNAME (alias), NS, MX, TXT, SRV, CAA; ALIAS/ANAME at the apex.
TTL — how long resolvers cache; low TTL = faster failover, more queries.
Negative caching — NXDOMAIN cached per SOA minimum TTL.
GeoDNS / latency-based routing — answer differs by client location (Route 53 latency/geo policies).
Weighted / failover routing — DNS-level traffic split and health-checked failover.
Anycast — one IP, many locations; BGP picks nearest; how root servers and CDNs work.
DNS propagation — misnomer: it's cache expiry, not push.
DoH / DoT — DNS over HTTPS / TLS.
Split-horizon DNS — internal vs external answers.

How it works#

TEXT

browser cache -> OS cache -> recursive resolver (cached?)    -> root (.)  -> "ask .com"   -> .com TLD  -> "ask ns1.example.com"   -> authoritative -> A 203.0.113.10, TTL 60browser cache -> OS cache -> recursive resolver (cached?) 
   -> root (.)  -> "ask .com"
   -> .com TLD  -> "ask ns1.example.com"
   -> authoritative -> A 203.0.113.10, TTL 60

Client asks recursive resolver; cache hit returns immediately.
Miss: resolver walks root → TLD → authoritative (roots/TLDs are heavily cached so usually one hop).
Authoritative may answer based on resolver location (GeoDNS) or health (failover).
Result cached for TTL at every layer; some resolvers/OSes clamp or ignore TTLs.

As a load balancer: return multiple A records (round-robin), weighted answers, or geo-nearest. Coarse because of caching: stale IPs survive for TTL after failover, and clients pick records unevenly.

Trade-offs#

Choice	Pro	Con
Low TTL (30-60 s)	Fast failover and migrations	More resolver queries, higher authoritative load/cost
High TTL (hours)	Fewer lookups, resilient to DNS outages	Slow to change
DNS round-robin LB	Free, global	No health awareness at client; caching skews distribution
GeoDNS	Route users to nearest region	Resolver location ≠ user location (ECS helps)
Anycast	Fastest, DDoS-absorbing	Needs BGP control; TCP sessions can break on route changes
CNAME vs ALIAS at apex	CNAME simple	CNAME not allowed at zone apex; use ALIAS/ANAME

Numbers#

Uncached full resolution: ~20-120 ms; cached: <1 ms.
Typical TTLs: 60-300 s for app endpoints, 3600+ s for stable records.
13 root server identities (A–M), each anycast across hundreds of sites.
UDP DNS message limit 512 bytes classic, ~1232-4096 with EDNS0; falls back to TCP when larger.

Say this#

How it works#

Trade-offs#

Numbers#

In the interview#