practice

Cutover Planning

The plan for the moment of switching — including the decision points, the rollback, and the rehearsal that makes it survivable.

cutoverrollbackrehearsalgo-no-gomigration

Definition

A cutover plan is a timed sequence of steps to move from old to new, with defined decision points and a rollback that has been tested.

What the plan must contain

  • A step-by-step sequence with owners and durations, precise enough to follow at 3am.
  • Go/no-go decision points, with the criteria stated in advance — not judged in the moment under pressure.
  • The point of no return, named explicitly, so everyone knows when rollback stops being available.
  • A rollback procedure for every step before that point, tested rather than described.
  • Verification steps — what is checked after each stage, and by whom.
  • Communication — who is told what, when, including customers and support.
  • A named decision-maker, available, with authority to abort.

The rehearsal

Rehearse the whole thing on production-like data, including the rollback. A cutover rehearsed once takes half the time and produces a fraction of the surprises.

The rehearsal reliably finds: a step that takes four hours rather than one, a permission nobody has, a dependency that must be stopped first, and a verification that cannot actually be performed.

Time each step during the rehearsal, and add substantial contingency. Cutovers overrun.

The decisions to make in advance

When do you abort? "If the data verification has not passed by 04:00, we roll back." Decided in advance, this is a straightforward check. Decided at 04:00 with a tired team and a sunk cost, it is an argument that usually resolves toward pressing on — which is how cutovers become incidents.

What is acceptable degradation during the window, and who agreed it?

Reducing the need for a cutover at all

The best cutover is a small one. Gradual traffic shifting behind a flag, phased by segment, with the ability to revert instantly, converts a high-stakes event into a series of low-stakes ones.

Reserve the big-bang cutover for cases where phasing is genuinely impossible — and check that claim, because it is frequently a preference rather than a constraint.

Failure scenarios

  • Rollback never tested, so it fails when needed.
  • No abort criteria, so the decision is made under pressure.
  • Steps untimed, so the window overruns and the team continues into business hours.
  • The decision-maker unavailable, so an abort waits forty minutes for someone to be found.
  • Verification that cannot be performed in the window, so nobody knows whether it worked.

Interview question

"What is in a cutover plan besides the steps, and what would you rehearse?"