Local cache appendix
Hot payload-cache results
These measurements describe in-process RAM lookup and shared-buffer dispatch on the benchmark host. They are intentionally separated from remote-provider results.
Interpretation
A hot hit returns a shared Arc<Vec<u8>>. Millions of logical object reads per second do not mean that millions of full payloads were copied through memory. The figures measure lookup, synchronization, LRU bookkeeping, reference counting, and dispatch.
Provider labels are repeated host observations. The hot path does not call S3. Differences among labeled runs reflect ordinary host-process variance and cache state, not an object-store advantage.
Object dispatch rate
Payload sizes stay in separate panels. Each panel compares the writer head, reader head, and two independent replica processes under the same logical object size.
Lookup tail latency
Latency histograms retain the first 1,024 successes and then every 1,024th success, bounding instrumentation overhead during the highest-rate phases.
What belongs on the main report
Warm results stay on the main report because residency changes with cohort size and role. In particular, the 188 GiB 16 MiB cohort forces a real RAM-plus-NVMe working set; the 47 MiB 4 KiB cohort does not.