The data centre migration checklist

Most migration plans are built from documentation. Most documentation is wrong. That single sentence explains the majority of migrations that overrun, and the fix is unglamorous: go and look at the hall before you plan anything.

Phase 1: audit what is actually there

Not what the CMDB says. What is physically installed, cabled and drawing power. A useful audit captures:

• Make, model and serial number of every device, with its rack and U position
• Every cable, both ends, including cross-connects leaving the cabinet
• Measured power draw per circuit and per cabinet, not nameplate ratings
• Photographs of every cabinet front and rear
• Anything installed that nobody can identify or claim ownership of

Phase 2: map the dependencies

Physical inventory tells you what exists. It does not tell you what breaks when you unplug it. Dependency mapping is the part that determines your move sequence, and it needs input from application owners rather than infrastructure alone.

• Which services depend on which hosts, and in which direction
• Which systems are genuinely resilient across sites and which only look it
• Shared infrastructure: authentication, DNS, NTP, monitoring, backup targets
• Licence servers and anything else tied to a MAC address or an IP that will change
• External dependencies: carrier circuits, cross-connects, third-party links

The systems that look resilient but are not deserve particular attention. Active-active across two sites is genuinely movable with no downtime. Active-passive where the passive node has never actually been tested under load is not, whatever the architecture diagram says.

Phase 3: group the estate into waves

A move wave is a set of equipment that can move together within one change window without breaking anything left behind. Sizing waves is a judgement call between two failure modes: waves too large to fit their window, and so many small waves that the project takes a year.

1. 1Start with the systems that have no inbound dependencies and move outward
2. 2Keep tightly coupled systems in the same wave rather than splitting them across windows
3. 3Size each wave against the real duration of shutdown, unrack, transport, rack and verify, not the optimistic one
4. 4Leave shared infrastructure (authentication, DNS, monitoring) until either everything depending on it has moved, or build it new at the destination first
5. 5Put something low-risk in wave one, because wave one is where you discover what your estimates were wrong about

Phase 4: give every wave a rollback

A rollback point is a defined state you can return to if the wave does not come up cleanly, and a defined moment at which you decide. Both halves matter. A rollback you can technically perform but have no time left to execute is not a rollback.

• Define the go or no-go decision time, well before the window closes
• Keep the source racks powered and cabled until the wave is signed off, not until it is moved
• Confirm backups are current and restorable before the window opens, not during it
• Agree in advance who has authority to call the rollback

Phase 5: the destination, early

The destination site needs work weeks before the first wave moves, and the long poles are almost always connectivity rather than space or power.

• Cross-connects and carrier circuits ordered early, because lead times run to weeks
• Rack elevations agreed and cabinets physically prepared
• Power circuits commissioned and A and B feeds confirmed
• Structured cabling and patching installed ahead of the equipment arriving
• Facility access lists updated for everyone who will be on site, including your movers

Phase 6: move day

1. 1Photograph every cabinet before anything is disconnected
2. 2Label both ends of every cable as it comes out, even ones you are replacing
3. 3Shut down cleanly and confirm shutdown before unracking, rather than assuming
4. 4Transport with appropriate packing and shock monitoring for the distance and equipment
5. 5Rack to the agreed elevation at the destination, cable and power up
6. 6Verify at three levels: device powers on, device is reachable, application actually works
7. 7Get explicit sign-off from the application owner before releasing the rollback

Phase 7: after the move

• Decommission the source racks and hand the space back, since you pay for it until you do
• Update the CMDB from the as-built record while it is still accurate
• Cancel the circuits and cross-connects at the source site, which is frequently forgotten and quietly expensive
• Hold a short review after the first wave and adjust the plan for the rest

That last point is worth protecting. The first wave is the most informative event in the whole project, and the plan for waves two onwards should look different afterwards. A plan that emerges from wave one unchanged usually means nobody was paying attention.