Payment systems turn out to be particularly thorny because they hit an intersection of technical issues with conflicting goals:
* You want them to be as fast as possible (ideally: instantaneous)
* You want them to be validated (long-poll to a third party validation service)
* You need them to be sequentially auditable (this is a "... or you could go to jail" requirement)
* If a failure occurs, you don't want to be out your own money
While these are solvable, it's the reason that so many cloud services suddenly hit scaling issues at the payments layer: things are going great and then that layer gets involved and "oops, hold on, we're waiting on Visa's servers. Still waiting. Stillllll waiiiiiting..........." Or the team was certain they'd simultaneously solved speed and sequential auditability this time but, oops, there's yet another sequencing point that's actually a bottleneck.
Credit card based payment systems really shouldn't suffer these kinds of problems: The entire reason why credit cards exist in the first place (compared to debit cards) is to enable distributed, asynchronous, eventually-consistent payments. It should be perfectly normal for a business like steam to simply store the details necessary to process payment and then eventually reconcile with them, at a pace that the processors can handle. Instead we've built a world that operates under physics-violating assumptions. Gah.
* You want them to be as fast as possible (ideally: instantaneous)
* You want them to be validated (long-poll to a third party validation service)
* You need them to be sequentially auditable (this is a "... or you could go to jail" requirement)
* If a failure occurs, you don't want to be out your own money
While these are solvable, it's the reason that so many cloud services suddenly hit scaling issues at the payments layer: things are going great and then that layer gets involved and "oops, hold on, we're waiting on Visa's servers. Still waiting. Stillllll waiiiiiting..........." Or the team was certain they'd simultaneously solved speed and sequential auditability this time but, oops, there's yet another sequencing point that's actually a bottleneck.