The line is fine, the firewall is green, ping works; but the user can’t open any site. In the field, DNS is most often behind this picture. In most organizations, DNS is seen as a service that just runs on the domain controller; yet branch access, internet egress, SaaS performance and email delivery are tied to the same chain. Below we look at the topic from the network team’s perspective: the internal resolution chain, split-horizon, hosting the external domain, TTL and DNSSEC decisions, and separating layers with dig and nslookup during a fault. Day-to-day operation of Active Directory DNS (scavenging, zone replication) is out of scope.
Resolver and authoritative server: two separate roles
There are two different roles in DNS. Most design mistakes come from loading these two roles onto the same server without thinking.
| Role | Function | Whom does it serve? |
|---|---|---|
| Recursive resolver | Follows the query to the end on behalf of the client and caches the result | Internal clients only |
| Authoritative server | Gives definitive answers for its own zone, doesn’t recurse | The internet or the internal network, depending on the zone |
| Forwarder | Doesn’t resolve the query itself, forwards it to another resolver | The resolver’s upstream |
If an internet-facing server both answers authoritatively for your external zone and does recursive resolution for everyone, you have an open resolver. This server is used as a reflector in DNS amplification attacks, the cache poisoning surface grows, and one day an abuse report arrives from the carrier. The rule is simple: an internet-facing server gives only authoritative answers, and recursion is open only to queries from the internal network. In BIND, the recursion and allow-recursion settings determine this; in Windows DNS, you disable recursion on the server hosting the external zone. A dig @public-ip example.org test from outside shows this gap in a minute.
Internal resolution: forwarders, conditional forwarding, root hints
In a typical setup, clients receive the internal DNS servers, mostly domain controllers, via DHCP. For names not in its own zone, the internal server uses one of two ways:
- Root hints: The server does resolution itself, starting from the root servers. There’s little external dependency; but the server must be able to go out to the internet directly on UDP and TCP 53. Forgetting TCP 53 closed produces intermittent errors with large responses and DNSSEC.
- Forwarder: Queries are forwarded to the servers of the carrier or a resolver service you choose. A shared cache and central filtering become easier; in return, you depend on the reachability and policy of the upstream server.
A mistake you may encounter in dual-carrier setups is configuring only one carrier’s resolver as the forwarder. Carrier resolvers mostly respond only to queries from their own IP blocks. When the primary line goes down and traffic moves to the second carrier, the forwarder goes silent and the complaint “there’s internet but no site opens” starts. Failover worked. Name resolution didn’t. That’s why you should definitely add a DNS resolution test to the test steps in the dual internet line and failover design article.
Conditional forwarding sends queries for a specific zone to specific servers. It’s used for a partner’s internal zone in a company merger, a supplier domain behind a site-to-site tunnel or a cloud provider’s private DNS zone. When the tunnel is down, the target servers can’t be reached and the client waits as long as the timeout for every query; the user reports this as “the application froze.” For forwarder and zone settings on the AD DNS side, see the Active Directory DNS configuration guide on the systems side.
The complementary control on the network side is the egress rule. If you allow clients to go directly to port 53 on the internet, internal DNS policy (conditional forwarding, split-horizon, filtering) is easily bypassed, and during a fault you can’t tell which client used which resolver. Open DNS egress on the firewall only to the internal resolvers.
Split-horizon DNS design
Split-horizon (split-brain) DNS is the same name giving a different answer depending on who asks. For example, portal.example.com returns the public IP on the firewall (203.0.113.20) when asked from the internet and the server’s private IP (10.10.20.15) when asked from the internal network. The internal user goes directly to the server without going around through the firewall. The ways to reach the same result with NAT, and their limits, are in the enterprise NAT design article under the hairpin heading.
There are three ways to implement it:
- Full-copy internal zone: A separate zone for
example.comis created in internal DNS. This is the method that causes the most problems: because the internal zone is the only authority for this name, if records in the external zone such aswww,mailandautodiscoveraren’t added internally too, internal users can’t resolve them. Every external change has to be made in two places, and one day someone forgets. - Single-record (pinpoint) zone: Only a zone named
portal.example.comis created in internal DNS, and an A record at the zone’s own name (@) points to the private IP. Every other name resolves normally from the internet. This is the method with the lowest maintenance effort. - Views: On servers such as BIND, a single server serves different zone files based on the query’s source address. Internal and external records are managed in the same place; but a wrongly matched source network leaks internal addresses to the internet. ACLs should be reviewed often.
If you’re building from scratch, choose a subdomain of a domain you own (such as corp.example.com) for the internal Active Directory domain; the need for split-horizon is minimized. .local is reserved for multicast DNS, and unregistered made-up suffixes may later become a real top-level domain. Both turn out costly later.
Authoritative DNS hosting for the external domain
Where the external zone sits can be the weakest link in website and email availability. RFC 2182 recommends that a zone’s authoritative servers be on different networks and in geographically separate locations. Two servers in the same cabinet don’t meet this recommendation.
| Option | Advantage | Limitation |
|---|---|---|
| The registrar’s DNS | Requires no setup | API, record type and availability support vary |
| Managed DNS service (anycast) | Distributed infrastructure, API, health-checked responses | Account security (MFA, permissions) becomes critical |
| Your own servers | Full control | Location and carrier diversity, patching and DDoS resilience are on you |
| Hidden primary + secondary service | Records are managed internally, secondaries answer the internet | Requires AXFR/IXFR and TSIG configuration |
In the hidden primary model, the zone file is kept on a server on the internal network and doesn’t appear in the NS records; only secondary servers updated by zone transfer answer the internet. Restrict transfers with a TSIG key and write the allowed addresses explicitly. Whichever model you choose, protect the registrar account and the DNS management account separately with MFA. Someone who gets into these accounts can redirect all your web and email traffic wherever they want. A password shouldn’t be the only defense.
TTL strategy and change plan
TTL determines how long a record is kept in resolver caches. A long TTL reduces query load and resolution latency; a short TTL lets changes propagate quickly. Rather than using a very short TTL permanently, it’s healthier to plan the change:
- Note the current TTL value of the record to be changed (example: 86400 seconds).
- Lower the TTL (example: 300 seconds) and wait at least as long as the old TTL; otherwise the old value stays in caches for a long time.
- Change the record; verify it both from the authoritative servers and from several different resolvers.
- Once there’s no chance of rolling back, raise the TTL to its normal value.
What causes trouble on migration night is usually skipping step 2: the TTL is lowered not a day before but five minutes before, and the old IP keeps being returned to some users for another day. Take negative caching into account too: a resolver that queries a name not yet created keeps the NXDOMAIN response in cache for a while according to the values in the SOA record (RFC 2308). Querying “just to see if it exists” before adding a new record leads to getting a “not found” response for a while after the record is added. Calculate these times separately for MX and NS changes.
DNSSEC basics
DNSSEC verifies with digital signatures that a response wasn’t altered in transit (RFC 4033–4035). It has two sides:
- Signing (authoritative side): The zone is signed with keys, an RRSIG is created for each record set; a digest of the key (the DS record) is published in the parent zone through the registrar.
- Validation (resolver side): The resolver checks the signature chain starting from the root. If the signature is invalid, SERVFAIL is returned to the client.
For an enterprise network, the low-cost first step is enabling validation on internal resolvers. Signing the external zone, on the other hand, requires an operating process: a key rollover schedule, updating the DS record in the right order when changing DNS providers, and monitoring. If you sign the zone with a new key while the DS record points to the old key, your domain becomes unreachable for every validating resolver. It doesn’t look like the fault is yours. But it is. If you use split-horizon, because the internal view will remain unsigned, make sure the internal resolver hosts this zone authoritatively or that an exception is defined (domain-insecure in Unbound, validate-except in BIND).
Resolution latency at branches
If branch clients go over the WAN to the DNS server at headquarters for every query, every new connection is slowed by the tunnel latency. When the WAN goes down, nobody can resolve names even if the branch’s local internet works. Options by branch size: placing a domain controller or caching resolver at the branch, pointing the branch firewall’s DNS proxy feature at the internal resolvers, or at least giving clients two different central resolvers. Write this decision into the template within the standard branch network design; don’t have the same debate again every time a branch opens.
At branches with local internet breakout, there’s a second effect too. CDN and SaaS providers mostly choose the response based on the location of the resolver making the query. If traffic exits at the branch while the DNS query exits at headquarters, the user may be directed to an endpoint near headquarters, and the complaint “SaaS is slow at the branch” comes. Check that the DNS egress point and the traffic egress point match.
Verification with dig and nslookup
During a fault, the first job is to separate which layer the problem is in: the authoritative server, the internal resolver, delegation or DNSSEC? The commands below make this separation in a few minutes. In the examples, 10.10.0.10 is the internal resolver and ns1.example.net is the external authoritative server.
# Non-recursive query to the authoritative server: the value at the record’s source
dig @ns1.example.net portal.example.com A +norecurse
# What does the internal resolver return? (split-horizon check)
dig @10.10.0.10 portal.example.com A +short
# Delegation chain from the root
dig portal.example.com A +trace
# SOA serial number: are the secondary servers up to date?
dig @ns1.example.net example.com SOA +short
dig @ns2.example.net example.com SOA +short
# DNSSEC validation: look for the "ad" flag in the response
dig @10.10.0.10 example.com A +dnssec
# Equivalents on the Windows side
nslookup -type=SOA example.com 10.10.0.10
Resolve-DnsName -Name portal.example.com -Type A -Server 10.10.0.10
When reading the output: if SOA serial numbers differ between servers, zone transfer isn’t working. If in the +trace output the NS records in the parent zone don’t match your zone’s own NS records, the delegation is inconsistent; if a listed server doesn’t answer authoritatively for the zone (no aa flag, or REFUSED), there’s a “lame delegation.” If the validating resolver returns SERVFAIL but the query succeeds with +cd (checking disabled), the problem is most likely in the DNSSEC chain. Query the authoritative server first, then the resolver.
Design checklist
- Recursion is disabled on internet-facing servers; internal resolvers are open only to internal networks.
- The forwarder is reachable from both internet lines; DNS resolution is checked in the failover test.
- A split-horizon method has been chosen; how external record changes will be reflected internally is written down.
- The external zone has at least two authoritative servers on different networks; the registrar and DNS accounts are protected with MFA.
- The TTL-lowering step has been added to the change procedure.
- If DNSSEC is used, the key rollover and provider change procedure is written down.
- Clients’ direct DNS egress to the internet is restricted.
If you want to address DNS architecture together with the line structure and branch design, see our enterprise network consulting and internet line management pages. The primary sources for the protocol’s basic concepts are RFC 2182 and RFC 4033.
Frequently asked questions
Is using .local as an internal domain a problem?
Yes, we don’t recommend it. .local is reserved for multicast DNS by RFC 6762; macOS, Linux and some printers try to resolve this suffix with mDNS first, resulting in delays or resolution errors. A publicly trusted certificate can’t be obtained for this name either. Use a subdomain of a domain you own (for example corp.example.com).
Split-horizon DNS or NAT hairpin?
Most organizations prefer split-horizon: internal users go directly to the server with its private IP, with no extra translation or unnecessary traffic on the firewall. Hairpin is a practical alternative where internal DNS can’t be modified or the number of records is very small. Whichever you choose, put in writing which record is kept where.
Is it right to use the carrier’s DNS server as a forwarder?
It works, but in a dual-line setup the carrier resolver mostly responds only to queries from its own IP block; when the line changes, resolution stops. Doing your own recursive resolution with root hints, or choosing a resolver service reachable from both lines with known logging and filtering policies, is more predictable.
Is DNSSEC also needed on the internal network?
The real gain is the internal resolver validating signed zones on the internet; enabling this is low-cost. Signing internal zones requires key management and usually gives limited benefit. If you sign your external domain, test that the internal view in split-horizon isn’t flagged as bogus by the validating resolver.