Start from the failure, not the diagram
Architecture diagrams describe the happy path. Systems are defined by what happens when one node is slow, not down — the case almost nobody designs for.
Consistency is per-operation
The useful question is never "should this system be consistent?" It is "for which read, against which write, across which regions?" Once consistency becomes a per-operation decision, the architecture gets dramatically simpler.
- Causal reads for feeds and timelines
- Linearizable reads for balances and quotas
- Local reads for anything a human will re-read in a second anyway
Latency is a distribution
Mean latency is a vanity metric. The p99.9 is where your users actually live, and the tail is almost always a queue you did not know you had.
Write it down
The most valuable artifact in a long-lived system is not the code, it is the record of why a trade-off was made. Six months later that document is the only thing standing between you and re-litigating the same decision.