Architecture walkthroughs tend to show you the polished version — the diagram that survived a whiteboard, cleaned up after the fact. What they skip is the sequence of decisions, tradeoffs, and dead ends that got there. That's what I want to share here, based on having built and operated Kafka-based payment pipelines at production scale in Brazilian fintech.
Start with Requirements, Not Technology
The first mistake teams make is starting with "we're going to use Kafka" before defining what correctness means for their payment events.
Key questions before drawing a single box:
What is the ordering requirement? Within a Kafka partition, ordering is guaranteed. Across partitions, it is not. If you need per-payment ordering (all events for a given payment ID in sequence), partition by payment ID and you get it for free.
