Decommissioning is a project

Nothing in an estate is harder to kill than a system that mostly works. Migrations get kickoffs; decommissions get a someday. Meanwhile the old thing hums along, billing for licenses, holding data, waiting for its certificate to expire at 2 a.m. Write this on the roadmap in pen: decommissioning is a project. Scope it like one.

The professionals treat shutdown as a milestone with a date. When 37signals finished pulling 18 petabytes out of S3 in 2025, the endpoint was not "mostly moved": it was deleting the AWS account, with the savings booked only when the meter actually stopped. That is the standard. Not migrated off; turned off. The dev team's own writeup reads like a logistics plan because it was one.

The project plan, in its usual shape:

1. Prove silence. Access logs, query logs, job history: sixty to ninety days of evidence that nothing real depends on the thing. "Real" matters; the uptime checker and the backup agent do not count as customers.
2. Scream test. Turn it off, on purpose, reversibly, during business hours, with a rollback that takes minutes. Announce it afterward, not before, or everyone suddenly remembers a dependency. Two silent weeks buys a demolition permit.
3. Data disposition, decided by policy, not by mood. What gets archived, in what format, readable by what, retained how long, deleted when. Legal signs this line. This step is half the project's actual work.
4. Contract and license teardown. Cancel the maintenance, the support tier, the DR replica's licenses. This is where the money is, and it does not cancel itself; procurement finds out from you or from the renewal notice.
5. The obituary. One page: what it was, what replaced it, where the archive lives, who to call. Filed where the inventory points at it. Future archaeologists send their thanks.

Fund it explicitly or it will not happen. The forces against are real: nobody's quarter improves, the risk feels asymmetric, and the people who remember why it is safe to turn off are leaving faster than the systems are. Our rule on every roadmap we write: each migration line carries its decommission line, with its own owner and date, and progress is counted in things turned off. Things turned on is a vanity metric. The estate shrinks or it does not.

The mainframe is not your problem. The seventeen things you already replaced but never unplugged, at four figures a month each: start there. It is the only work in modernization that pays for itself by Friday.