Reproduced committed divergence at LSN 9 (real-fs release soak): all 3 members term=3, minority ctrl-1 pcrc=0/PLEN=0 vs majority PLEN=3560 = PAYLOAD loss, term correct. Bisect: arrived via the reconcile seam under cross-term catch-up (stored term 3, leader term 4 / bug-1067 path), NOT the term-blind bypass. Real defect = cross-term catch-up payload loss in ha_stream_wal_to_standby (silent empty ship on short pread/malloc-fail). Independently verified against frozen WALs. Secondary: release-build log buffering breaks soak primary-detection. |
||
|---|---|---|
| .. | ||
| settings | ||
| specs | ||
| steering | ||