A short defence of the boring retry policy
There is a popular shape of retry policy that involves machine learning and a lot of confidence. We have seen versions of it work and versions of it quietly underperform. The pattern that holds up over six months is almost always closer to a documented set of rules than to a black box.
The rules tend to be: short retry for soft declines, longer retry for issuer-side problems, different cadences by card brand, and a hard stop after a defined number of attempts. The total number of rules can be written on a single page. The discipline is in keeping it that simple while everyone in the room argues for an exception.
When the rules are documented and reviewed quarterly, the system tends to behave. When they are buried in the billing platform configuration and nobody can name them, the system drifts. Most retry projects we run are about moving the rules from the second state to the first.
Other recent briefs