Problem
Stripe processes billions of dollars a day, and a global success-rate metric hides trouble in small pockets of traffic. A card issuer can change its rules so that one payment type fails more often. Businesses that depend on that type see an acute drop, yet the global metric barely moves. Payment metrics also swing for valid reasons (new customers, fraud trends, a large free trial), so a plain comparison with history raises false alarms.
Example from the article: a UK card issuer starts declining recurring payments on prepaid cards at high rates, and only UK businesses with many prepaid-card payments notice.
Solution in brief
Split traffic into slices, then predict the success probability of every transaction with ML models. Run time-series anomaly detection on actual versus expected outcomes, so changes in the mix of transactions are not read as degradation. A finite state machine sums losses over time and raises an alert only for sustained loss, routed by urgency and likely root cause.
Questions this case answers
- How to detect degradations in many small segments of a high-dimensional population when aggregate metrics hide them and no stable baseline exists
- How to keep false positives manageable when thousands of segments are monitored in near real time
- How to detect a metric drop in a small traffic segment when the global metric does not move?
- How to monitor tens of thousands of segments without a stable baseline and without drowning the team in false alerts?
Tools and methods
- Slices of traffic — segments defined by combinations of issuer, currency, country, card product and payment features
- ML success-probability model — gives the expected outcome of each transaction, which acts as the baseline
- Time-series anomaly detection — compares actual with expected outcomes in near real time
- Finite state machine (FSM) — aggregates losses and decides when to alert
Google or a one-time email code. No password.