The Cost of Leaving

## Exit cost is a project, not a line on the bill **Exit cost** is the one-off money and calendar time spent moving a workload and its data off the platform it runs on today. It is a different number from the recurring bill, and the two get confused constantly. The recurring bill is what you stop paying once the move is finished; the exit cost is what you spend in order to start saving it. A team that says *we could leave in a quarter* has almost always priced only the part of the move that looks like engineering work. The example worth holding in mind is a media transcoding service with a long archive behind it, being moved off a platform after an acquisition decides to merge two estates onto one. The transcoding code is a few services. The archive is hundreds of terabytes that have been accumulating for years, most of it never read. The decision to merge was made on the recurring bills of the two estates; the number nobody produced is the one-off cost of getting out of one of them. ## The four lines | Line | Measured in | What sets it | What makes it grow | |---|---|---|---| | Data out | gigabytes moved at a per-gigabyte outbound rate, plus retrieval charges on anything in a cold tier | your storage inventory and the platform's published rates | total volume, and how much of it is archived rather than hot | | Rewrite | engineer-months | how much of the system leans on interfaces the target has no equivalent for | behaviour discovered during the port, not at planning time | | Overlap | months, multiplied by two full platform bills | first copy started to last consumer cut over | almost every slip elsewhere in the project | | Stranded commitment | months still to run, multiplied by the spend the term obliges | the contract signed for the discount | nothing — it was fixed on the day of signing | ## Why only one of the four ever gets estimated The rewrite is the only line that arrives as a backlog item. Somebody has to do it, so somebody sizes it. The other three are produced by a price list, a calendar and a contract respectively, and none of those is owned by the team doing the move. The data charge sits with whoever reads the bill. The overlap is invisible until two bills arrive in the same month. The commitment is in a document the engineering team has usually never seen. That is also why the estimate fails in a predictable direction: the missing lines are all costs, never savings, so an exit estimate is biased low by construction. ## The data line, and the direction that is priced Outbound bytes are normally charged per gigabyte; inbound bytes normally are not. That asymmetry is the whole reason the exit is the expensive direction — filling the platform you are joining looks free while emptying the one you are leaving does not. This is a priced asymmetry, not a technical one: the bytes move just as easily either way, and **data gravity** is a bill plus a calendar rather than a physical limit. Two details are routinely dropped. First, data held in a cold storage tier generally has to be retrieved into a readable state before it can be read out at all, and that retrieval is charged separately from the transfer. Second, the copy takes wall-clock time at the throughput the link actually sustains, and that elapsed time is what sets the minimum overlap window. ## The overlap is the residue Until the last consumer has moved, both platforms are billed in full. So the overlap line is not an independent estimate at all — it is the accumulator into which every other delay drains. A rewrite that slips a month adds that month's engineer cost **and** a month of two bills. A copy that runs slower than promised does the same. When an exit estimate turns out to be double, the doubling is usually here. ## Producing a number you can defend 1. **Inventory the data first**, split by storage tier and volume, because the tier decides whether a retrieval charge applies before anything can move. 2. **Measure sustained throughput on a real sample** rather than quoting the link's rating, and convert it to days before promising a cutover date. 3. **Size the rewrite by porting one representative component end to end**, then scale, rather than by counting call sites. 4. **Write the overlap as a formula**, months times both monthly bills, so that every slip discussed in a status meeting has a visible price. 5. **Read the term commitment** and state the months remaining as a separate line, since it is fixed and belongs in the total whatever the schedule turns out to be. State the assumptions next to the number. An exit estimate without its volume, throughput and month assumptions written beside it cannot be re-run when any of the three changes, and all three change.