Planning a Cloud Migration That Does Not Stall
Most cloud migrations do not fail, they stall halfway while the business pays for two environments. A sequencing and governance approach that keeps the project moving.
Cloud migrations rarely fail outright. They stall. Six months in, forty percent of the estate has moved, the easy workloads are done, the hard ones have been deferred twice, and the organisation is paying for two environments while arguing about whether to continue. That half-finished state is more expensive and more fragile than either the old setup or the finished one.
Avoiding the stall is mostly a sequencing and governance problem rather than a technical one. The technology for moving a server has been solved for years. What has not been solved is deciding which thing moves next, who signs off, and what happens when the answer is inconvenient.
What stalling actually looks like
The pattern is consistent enough to recognise early. The first wave moves quickly because it was chosen for being easy. Momentum and executive attention peak. Then the team hits a workload with an unclear owner, a licence that does not transfer, a database nobody wants to touch, or a compliance question that needs a lawyer. Progress drops. Attention moves elsewhere. The remaining workloads become “phase two”, which never gets scheduled.
Three things cause this more often than anything else: an inventory that was never complete, dependencies discovered during a cutover rather than before it, and no agreed definition of done. All three are addressable before you move a single byte.
Build an inventory people cannot argue with
Every migration starts with a spreadsheet, and most of those spreadsheets are wrong. They are built from what the infrastructure team knows about, which is not the same as what exists. Shadow systems, a reporting server someone set up in 2019, a cron job on a developer’s machine, a third-party integration that writes to a file share: these are what turn a planned weekend into a three-day incident.
Build the inventory from evidence rather than memory. Pull from configuration management, DNS records, firewall rules, certificate inventories, billing line items and network flow logs. Flow logs in particular are hard to argue with. If something is talking to something else, it exists and it matters.
For each item, record the owner by name, the business function it serves, its data classification, its licence position, and its traffic profile. Mark anything with no identifiable owner. In most estates that list is longer than expected, and resolving it is cheaper now than during a cutover.
Classify by difficulty and by consequence
Once you have the inventory, score every workload on two independent axes. Difficulty covers technical coupling, data volume, licensing and how much the application assumes about its environment. Consequence covers what breaks in the business if it is down for an hour.
Those two axes give you four groups, and each gets a different treatment. Low difficulty and low consequence workloads are your proving ground. High consequence and low difficulty workloads move early, because the win is large and the risk is contained. High difficulty and low consequence workloads are where you spend engineering time re-architecting, since a mistake is survivable. High difficulty and high consequence workloads move last, with the most rehearsal, and sometimes they do not move at all.
That last admission is important. A migration plan that assumes everything moves is usually a plan that will stall on the one thing that should have been declared out of scope on day one.
The dependency map is the real schedule
Workloads do not move in isolation. They move in clusters bound together by latency sensitivity and shared data. Splitting a chatty pair across two environments introduces a network hop into a call path that was designed assuming a local one, and the symptom is a slow application rather than a broken one, which makes it harder to diagnose and easier to ignore.
Derive your clusters from observed traffic, not from architecture diagrams. Diagrams describe intent; flow data describes reality. Once you have clusters, your wave plan writes itself: each wave is one cluster plus whatever shared services it needs.
Pay specific attention to shared services, identity, DNS, logging, secrets, certificate issuance and artefact storage. These sit underneath everything, and they generally need to exist in the target environment before anything useful can run there. Set up identity and access structure first, following the principles in IAM fundamentals and least privilege, because retrofitting permission boundaries onto a live estate is far harder than establishing them in an empty one.
Data is the constraint, not compute
Compute is reproducible. You can build a new instance in minutes and throw it away if it is wrong. Data cannot be thrown away, has to arrive intact, and usually cannot stop changing while it moves.
Plan each dataset separately. Work out its size, its rate of change, how much downtime the owning business function will accept, and whether you can run replication to keep a target copy current while the source stays live. Replication plus a short cutover is almost always better than a long copy window, but it needs testing well in advance because the failure modes are subtle.
Watch the transfer cost too. Moving large volumes out of an existing provider is billed, and the bill is frequently larger than teams expect. Our guide to moving data between cloud providers safely covers transfer mechanics and verification, and cloud egress fees explained covers why the number lands where it does.
Prove the pattern on something small
Pick one workload from the low difficulty, low consequence group and take it all the way through: build, migrate, test, cut over, run in production for two weeks, decommission the old one. Not a proof of concept. A real production move, including the decommission, because the decommission is where organisations discover they are not actually confident.
What you are testing is the process rather than the technology. Does the change approval route work? Does monitoring pick the new environment up automatically? Does the on-call rota cover it? Does billing attribute it correctly? Does anyone notice when the old server is switched off? Each of those questions costs an hour to answer on a small workload and a weekend to answer on a large one.
Write down what you learned as a runbook template. Every subsequent wave should consume that template and improve it.
Cutover needs a rollback that someone has rehearsed
Every cutover plan should name the point of no return and state explicitly what happens on either side of it. Before that point, rollback means pointing traffic back at the source and losing nothing. After it, rollback means restoring from backup and losing whatever was written since. Teams get into trouble when they cross that line without noticing, usually because the source was decommissioned or writes were accepted in two places.
Rehearse the rollback at least once. An untested rollback is a hope. Set the decision deadline in advance too: if the migration is not verified by a stated clock time, you roll back, regardless of how close it feels. Without that rule, cutover windows extend until everyone is tired and judgement is poor.
Keep the source environment intact but read-only for an agreed period afterwards. A week is common. The cost of leaving it running is small compared with the cost of discovering on day four that a quarterly report job was never migrated.
Budget for the period when you pay twice
During a migration you run two environments. That overlap is unavoidable, and leaving it out of the budget is one of the most reliable ways to lose executive support halfway through. Model it honestly: source costs continue, target costs ramp, transfer costs spike during data moves, and staff time is consumed on both sides.
Attach a decommission date to every migrated workload and track it as a deliverable. Migrations stall financially when the old estate is never switched off. If you are not sure how to forecast the target side, building a cloud budget sets out a workable method, and right-sizing cloud instances explains why a direct capacity match from the old environment usually overspends.
Choosing the destination without over-deciding
Teams lose weeks to provider selection before they have enough information to choose well. A more useful order is to pick a plausible destination, move your proving-ground workload, and let real operating experience inform the bigger commitment. You will learn more from one production workload running for two weeks than from any amount of feature comparison.
What to weigh: where your data must legally sit, whether you need managed services or just reliable compute, how predictable you need the bill to be, and whether your team already knows a platform. A comparison like AWS and DigitalOcean is useful precisely because the two sit at opposite ends of that trade-off. On TrustCloudHub you can review verified listings including a Buy AWS Account, a Buy Hetzner Account or a Buy Digital Ocean Account to get a working environment quickly. We are an independent marketplace with no provider affiliation, so verify account state, limits and any credit balance inside the provider’s own console once an account is yours. Credits in particular are provider-controlled, carry expiry dates and service restrictions, and should never be assumed until you have seen them there.
Define done and say it out loud
A migration is finished when the source estate is switched off, the contracts are terminated, the runbooks describe the new environment, and the team has stopped maintaining two sets of knowledge. Write that definition at the start and report against it monthly. The projects that stall are almost always the ones that never agreed what finishing meant.