From the v0.8.7 recording (now merged, docs/cluster-persite-v087.md): the P1-persite-single bridge-point cell is internally inconsistent — 36% slower than the WRITE cell on the same node (true in no other lane), 11.3% rep spread, p99 6-9ms vs P2's 1.5ms. Taken at face value it would claim per-site clustering makes reads 58% faster, which has no credible mechanism.
The recording deliberately reports the read-path P1→P2 delta as NOT MEASURABLE. Re-record the P1 lane solo (cold host, no adjacent lanes) and update the reference table. The write-path numbers (0.2-0.3% spread) are unaffected.
(The other open item from that run — P3's 10x forwarded-write p99 tail — is tracked upstream as ephpm#439.)
From the v0.8.7 recording (now merged, docs/cluster-persite-v087.md): the P1-persite-single bridge-point cell is internally inconsistent — 36% slower than the WRITE cell on the same node (true in no other lane), 11.3% rep spread, p99 6-9ms vs P2's 1.5ms. Taken at face value it would claim per-site clustering makes reads 58% faster, which has no credible mechanism.
The recording deliberately reports the read-path P1→P2 delta as NOT MEASURABLE. Re-record the P1 lane solo (cold host, no adjacent lanes) and update the reference table. The write-path numbers (0.2-0.3% spread) are unaffected.
(The other open item from that run — P3's 10x forwarded-write p99 tail — is tracked upstream as ephpm#439.)