In this case the two points where we should have caught the issue would have been: * Integration testing with a specific data set that was different from actua

In this case the two points where we should have caught the issue would have been: * Integration testing with a specific data set that was different from actual production data. * Code review (they're not systematic yet but we're working on that) because the faulty code had a sketchy section and the conversation around it would have lead to a better version (that's what actually happened when fixing the bug)