Goal: de-risk the head-row + S3-log design with the smallest set of experiments that could actually change it, and decide the schema-growth strategy from data rather than assumption.
- Two kinds of result, never mixed. Shape / correctness / relative results (HOT-update ratio, bloat trend, CAS correctness, MinIO-vs-SeaweedFS comparison) are valid in containers on modest hardware. Absolute throughput/latency needs a production-representative instance and same-region S3; a laptop number there is fiction. Every chart is labelled which it is.
- Earn the complexity. Start from the simplest viable schema (one unpartitioned table, delete-on-terminal, single primary). Partitioning, forward-migration, and sharding are each gated behind a specific metric that fails without them. If the simple thing holds at target scale, we ship the simple thing.
- Representative workload. Synthesize the real shape: per-run transition counts of 5 to 20