Moving DHCP, DNS and Active Directory is not dragging three VMs from one column to another. Everything leans on everything else. DNS finds the domain, DHCP hands out the right settings and plenty of apps simply assume that time, names and authentication already work.
- Dependenciesnames, addresses, time and authentication
- Pilotnew services on two sites
- Cutoverone scope at a time
- Evidencefailure tests and rollback
First find out what depends on what
Before creating a new server, I want the dependency map:
- DNS zones, forwarders and conditional resolution
- domain controllers and roles that are actually used
- Active Directory subnets and site links
- DHCP scopes, reservations, options and relays
- time sources
- devices with hard-coded addresses
- applications still pointing to an IP instead of a name.
Hidden dependencies are often boring: a switch with two DNS addresses hard-coded, a forgotten DHCP relay, a camera that accepts one time server, or a script expecting the old server name.
I also separate authoritative service from cache or relay. Moving a local DNS resolver is not the same job as moving a domain-integrated zone.
Two pilot sites do not make everything redundant by magic
Services on two sites provide options, but resilience still depends on network paths and shared dependencies. Two VMs on different sites may fail together when they rely on the same storage, link or management service.
For Active Directory and DNS, examine:
- replication between controllers
- declared sites and subnets
- DNS servers handed to clients
- resolution during loss of one site or the MPLS
- roles that remain elsewhere and whether they are required during the failure.
For DHCP, the design depends on the product and network: native failover, split scopes or another supported mechanism. Relays still need to reach the right servers and leases need to remain coherent. Copying a configuration database without testing renewal gives us a nice screenshot, not a reliable service.
Migrate in small scopes
I cut the migration into four chunks:
- Build the new services without removing the old ones.
- Test with a clearly identified pilot VLAN or site.
- Cut over clients through a controlled relay, option or configuration change.
- Observe before disabling the old path.
Every stage needs a precise rollback. “Put it back” is not a plan if nobody knows which options, routes and records changed.
DHCP details
For every scope, check network, mask, gateway, DNS, lease duration, reservations, exclusions and special options. Test a new lease, a renewal and a device restart.
Lease duration affects the cutover. A very long lease delays new options. A very short lease adds traffic and exposes DHCP failure faster. The duration may be reduced before migration, then returned to a normal value after stabilisation.
DHCP relay is critical in a segmented network. If the request crosses a filtered zone, rules must allow the exact path without opening the entire zone towards the servers.
DNS and Active Directory details
Active Directory relies heavily on DNS. Clients need to find domain service records, reach an appropriate controller and keep time consistent for authentication.
I test more than nslookup:
- useful forward and reverse resolution
- domain-controller discovery
- sign-in with one site available and then unavailable
- replication and appearance of expected changes
- external resolution through forwarders
- behaviour of devices that are not domain members.
A warm cache may hide a failure. Tests need a fresh client or a safely cleared cache, not random disruption on production systems.
Failures to play before the real failure
Normal operation is not enough. I prepare controlled tests:
| Scenario | What I want to observe |
|---|---|
| one server lost | failover without manual client configuration |
| one pilot site lost | continuity from the second site within the planned scope |
| MPLS lost | known local behaviour even if some functions are unavailable |
| external datacenter unavailable | no hidden dependency for migrated services |
| link restored | recovery without duplicates, conflict or stuck replication |
Two servers drawn side by side are not resilience yet. It becomes real when we cut one piece and degraded mode does exactly what we expected.
When to retire the old service
I remove the old path after an observation period and clear criteria: no expected clients remain, logs are quiet, dependencies are handled, backups are verified and rollback is either no longer needed or replaced by a recovery procedure.
The final documentation must show the normal path, degraded path, owners and verification commands. Otherwise, six months later the network “works” but nobody remembers why.