Architecture
When to stop scaling horizontally and re-architect.
A practitioner's reading of the three load patterns that mean the cluster has run out of road and the workload needs a different shape, not more nodes.
Not every scaling problem is solved by more nodes. Three patterns tell you the workload has run out of horizontal road.
- Tail latency climbing while median stays flat — you have shard hot-spots, not capacity issues
- Cross-service fan-out multiplying with each new tenant — the topology, not the scale, is the ceiling
- Cache warm-up cost dominating the recovery budget — the working set no longer fits the shape you designed for