The Ledger Nobody Opens: Why Infrastructure Migrations Triple Their Estimates Before the First Ticket Is Written
Photo by Photo by Tyler on Unsplash on Unsplash
Every engineering organization has a migration story that ends the same way: the six-week project that took eight months, the budget approved at $400K that landed closer to $1.2M, the cutover date that slipped three times before someone quietly stopped announcing dates at all.
The pattern is so consistent across the industry that it deserves to be treated as a law rather than a coincidence. Platform migrations — database moves, cloud replatforming, monolith decomposition — reliably cost two to four times their original estimate. Not because teams are careless, but because the estimating process itself excludes the phases that consume the most resources.
The Invisible Middle Phase
When leadership pictures a migration, they tend to imagine two moments: the beginning (assessment, tooling, planning) and the end (cutover, celebration, decommissioning the old system). The middle — the period when both systems run simultaneously — is where the money actually goes.
Dual-running is almost always framed as temporary. In practice, it behaves like a tenant who refuses to leave. Data synchronization pipelines must be built, monitored, and repaired. Schema drift between the old and new systems must be reconciled weekly. Incident response now spans two platforms, doubling the surface area of every 3 a.m. page. And because traffic can rarely be cut over all at once, teams maintain read-write paths against both systems for months, paying full operational tax on each.
The cruelest part: the longer the dual-run lasts, the more organizational pressure exists to keep it lasting. Rolling back becomes politically harder each week, which means the team absorbs escalating risk rather than admitting the timeline was fiction.
Where the Estimates Break
Four assumptions do most of the damage.
First, schema assumptions. Every legacy database contains columns nobody remembers creating, constraints enforced only in application code, and data that violates rules the documentation swears exist. Discovery of these anomalies is not a phase — it is a continuous tax that surfaces throughout the project.
Second, the integration census. The estimate usually counts the systems known to connect to the platform being migrated. It rarely counts the cron job someone wrote in 2019, the partner integration using an undocumented endpoint, or the analytics pipeline reading directly from a production replica.
Third, the human calendar. Migration work competes with feature delivery for the same engineers. A six-week migration staffed at 50% attention is a twelve-week migration, and the context-switching cost degrades everything else the team touches.
Fourth, the rollback nobody rehearsed. Every serious migration plan includes a rollback strategy. Almost none include a tested one. When cutover fails at 11 p.m. on a Sunday, the cost of an unrehearsed rollback is measured in days of engineering time and weeks of eroded executive confidence.
A Cost Model That Survives Contact With Reality
Before greenlighting a migration, engineering leaders should model four numbers explicitly:
- Sync cost per month. The fully loaded engineering and infrastructure cost of keeping two systems consistent. Multiply by a dual-run duration of at least six months, whatever the plan says.
- Discovery buffer. Add 30–40% to the schema and integration assessment estimate. Legacy systems hide their complexity by design.
- Opportunity cost. The delayed or descoped feature work attributable to the migration, expressed in revenue or roadmap terms. This is the number executives understand, and it is almost never on the slide.
- Exit cost. What it takes to reverse course at month three, month six, and month nine. If the exit cost at month six is prohibitive, leadership should know that before approving, not after.
A migration that only looks viable when dual-running lasts eight weeks and nothing goes wrong is not a viable migration. It is a wish with a Gantt chart.
Deciding With Open Eyes
None of this argues against migrating. Platforms outgrow their foundations, vendor pricing changes, and technical debt compounds. Some moves are genuinely necessary.
The argument is for honesty in the business case. A migration approved at three times its naive estimate and still worth doing is a sound investment. A migration approved at the naive estimate is a decision made with fabricated numbers — and the organization will pay the difference either way, just without having chosen to.
The best engineering leaders present both figures. The credibility earned by an accurate, uncomfortable estimate is worth more than the approval won by a flattering, fictional one.