When a job stops halfway, someone has to know where it stopped
Most multi-step work between two organizations is fine until something interrupts it in the middle. Then the usual fix, trying again, either does the first step twice or leaves the second one undone.
Escherbridge
Most handoff failures don't happen when everything is going well. They happen in the gap between two steps: one team has reserved the money, the next step is to pay the supplier, and the system restarts before anyone hears back. Nobody did anything wrong. The work simply stopped in the one place nobody was writing down.
When the system comes back, all it knows is that the job was interrupted. If it tries again, it may pay the supplier twice. If it assumes the payment failed, the reservation sits there forever. Either way, a person eventually finds out from the ledger, which is the most expensive place to find out anything.
Trying again is not a plan
The tempting fix is to retry harder. The people who write the web's own rulebook already said not to: a request that changes something, such as a payment, should not be repeated automatically just because the reply went missing. Stripe handles half of the problem by giving each payment request a reference number, so a repeat returns the first answer instead of charging again. That helps with one step. It doesn't help when step one happens in your organization and step three is declined three seconds later in someone else's.
Write down each step before you take it
Database researchers described the better habit in 1987: break a long piece of work into steps, record each one as it happens, and decide in advance what undoing each step looks like, or where the work simply halts. In business terms it comes down to three habits. Record what you intend to do before you do it. When you don't hear back, check what actually happened instead of guessing. And when an approval is unclear, stop rather than wave the work through. The Paid-Twice simulator shows the difference in about a minute.
What it costs
This takes more planning up front. Each step needs a defined undo, and work that gets stuck has to land in front of a person instead of retrying quietly in the background. AZOA keeps anything that moves real money switched off while it is still in alpha and under review, because finding out before launch is cheaper than unwinding duplicate payments afterward.
Sources
Related
- Run the Paid-Twice interactive simulatorSimulate dropped network replies, retry loops, and idempotent status checks.
- Field guide: Idempotent HTTP requestsWhat RFC 9110 §9.2.2 guarantees—and what application keys still have to store.
- Paid twice, or never, is the same bugWhy payment timeouts cannot be treated as failed requests.