Lift and Shift vs Re-Architecting: Choosing a Migration Path
Lift and shift and re-architecting are the ends of a range, not a binary choice. How to pick an approach per workload instead of applying one rule to an estate.
The question gets framed as a choice between two philosophies. Move everything as it is and modernise later, or rebuild properly while you have the chance. Both camps have war stories, and both sets of war stories are true.
The framing is the problem. Lift and shift and re-architecting are the two ends of a range, and most of the useful options sit between them. More importantly, the decision belongs to each workload individually. A single answer applied across an estate is how migrations end up either expensive and slow or fast and disappointing.
What lift and shift really means
Lift and shift means running the same operating system, the same application binaries and the same configuration on infrastructure provided by someone else. A virtual machine moves to a virtual machine. The application does not know it has moved, which is precisely the point.
The benefits are real and often undersold. It is fast. It needs little application knowledge, which matters when the people who wrote the system left years ago. The risk is contained, because behaviour does not change and a rollback is comprehensible. It lets you exit a data centre lease or a hardware refresh on a fixed date, which is frequently the actual business driver.
It also buys information. Running a workload on cloud infrastructure for three months teaches you more about its real resource profile than any amount of capacity analysis, and that knowledge makes a later redesign much better targeted.
The ways it disappoints
Lift and shift fails to deliver when teams expect it to produce savings or resilience on its own. It does neither by default.
Cost usually goes up first. On-premises hardware was bought once and sized for peak, so running it is a sunk cost. Cloud compute is billed continuously, so a machine provisioned to match the old peak capacity bills at peak capacity around the clock. Unless you resize, you have converted a depreciated asset into a monthly charge. The correction is covered in right-sizing cloud instances, and it is not optional.
Availability does not improve either. An application that assumed a single reliable server still assumes that. Cloud instances fail more visibly and more often than enterprise hardware with redundant everything, so a naive move can reduce uptime. If the workload has no clustering, no health checks and no automated recovery, it inherits none from the platform.
There are also things that simply do not move. Hardware licence keys bound to physical identifiers, software with licensing terms that forbid virtualised or shared infrastructure, and appliances with no cloud equivalent. Find these during inventory, as described in planning a cloud migration that does not stall, not during the cutover.
What a rebuild actually costs
Re-architecting means changing the shape of the application to suit the platform: decomposing it, replacing self-managed components with managed ones, making it horizontally scalable, moving state out of the instance, rewriting the deployment model.
The gains are genuine. Elastic capacity that tracks demand. Less undifferentiated operational work. Faster deployments. Resilience built into the structure rather than bolted on. Over a long enough horizon the economics usually favour it.
The costs are consistently underestimated in three places. First, duration: rebuilds take two to four times the initial estimate, not because engineers are slow but because the existing system contains behaviour nobody documented and some of it turns out to be load-bearing. Second, opportunity cost: the team doing the rebuild is not shipping features, and that gap is paid by the business for the whole period. Third, risk: a rebuilt system is new software, and new software has new bugs, arriving at exactly the moment you are also changing infrastructure.
A rebuild during a migration means changing two variables at once. When something behaves oddly afterwards, you will not know whether it is the platform or the new code.
The options in between
Most workloads belong in the middle, and the middle has names worth knowing.
- Replatforming. Keep the application largely intact but swap its surroundings. Move the self-managed database onto a managed one, put the web tier behind a managed load balancer, replace a file share with object storage. Modest code changes, large operational payoff.
- Containerising without decomposing. Package the application as it is and run it on an orchestration platform. You get consistent deployment and recovery without touching the internal design.
- Repurchasing. Stop running it. Replace a self-hosted system with a vendor service. Often the correct answer for mail, wikis, ticketing and similar supporting systems, where running your own earns nothing.
- Retiring. Switch it off. Every estate contains systems that are migrated out of habit rather than need. Auditing usage first is usually worth the day it takes.
- Retaining. Leave it where it is for now. A legitimate decision if the workload is near end of life or the move has no payoff.
Replatforming is the quiet winner for a large share of estates. It captures most of the operational benefit of a rebuild at a fraction of the cost and risk, and it is reversible in a way a rebuild is not.
Matching the approach to the workload
| Workload characteristic | Usually the right approach |
|---|---|
| Stable, low change rate, predictable load | Lift and shift, then resize |
| No current owner, little documentation | Lift and shift, or retire if usage is low |
| Actively developed, team knows it well | Replatform, with selective redesign later |
| Highly variable or seasonal demand | Re-architect for elasticity, where the saving justifies it |
| Commodity function with a mature vendor option | Repurchase |
| Hard deadline driven by a lease or contract | Lift and shift, change nothing else |
| Heavy self-managed infrastructure components | Replatform onto managed equivalents |
| Scheduled for replacement within a year | Retain or retire |
The deadline row matters more than it looks. When a date is fixed externally, reducing scope is the only reliable lever, and the first thing to cut is architectural ambition.
Change the platform or the code, not both at once
If a workload needs both a move and a redesign, do them in sequence with a stable period between. Which order depends on where the risk sits.
Move first when the deadline is external, when the application is poorly understood, or when the current environment is unstable. You get the deadline met, you learn how the workload behaves on the new platform, and the redesign happens against a known baseline.
Rebuild first when the application genuinely cannot run on the target in its current form, or when it is small enough that a parallel build is realistic. Building the new version alongside the old and migrating users gradually is a clean pattern when scale allows it, and the techniques in blue green and canary deployments apply directly.
The anti-pattern is doing both simultaneously on a critical workload with a tight deadline. That is where migrations produce outages that take days to untangle.
Cost behaves differently, not just at a different level
Comparing the two approaches on monthly spend alone misses the real difference, which is in the shape of the cost curve.
A lifted workload has roughly flat cost. It runs the same instances whether traffic is high or low, so the bill is predictable and insensitive to demand. Predictability has value, and for a steady internal system it is often the better property.
A re-architected workload has cost that tracks usage. Quiet periods are cheap, busy periods are expensive. That is an advantage for genuinely variable demand and a disadvantage for anything steady, where you have added complexity to achieve elasticity nobody needs. The comparison in predictable pricing versus pay as you go works through when each shape is preferable.
Run the numbers on your actual demand curve rather than on the assumption that elastic is always cheaper. For many internal business systems, a fixed-price instance on a value-focused provider is both cheaper and simpler than an autoscaling design.
Making the call, workload by workload
For each item in your inventory, answer four questions. Is there an external deadline? Does anyone currently understand this system well enough to change it safely? Does the demand curve vary enough for elasticity to pay? Will this still exist in three years?
A hard deadline or no owner pushes you to lift and shift. Flat demand plus a short remaining life pushes you to lift and shift or retire. Good internal knowledge plus heavy self-managed components pushes you to replatform. Variable demand plus a long life and a team with capacity is the case for a rebuild.
Where you land also affects which provider suits. A lifted estate wants straightforward, predictably priced virtual machines, while a re-architected one wants depth of managed services. That contrast is the substance of our Microsoft Azure and Hetzner comparison. On TrustCloudHub you can review verified listings including a Buy Azure Account, a Buy Hetzner Account or a Buy Kamatera Account and have an environment ready for a pilot move quickly. We are an independent marketplace with no provider affiliation, so verify account limits, region access and any credit balance inside the provider’s own console after transfer. Credits are controlled by the provider, expire and carry service exclusions, so treat them as a bonus rather than a line in the migration budget.