What cloudnative-pg/cloudnative-pg shipped
Generated by FoxPlug from public activity — not affiliated with Cloudnative-pg.
Week of September 14, 2026
What shipped
- Fixed the controller from stalling or reverting an in-progress failover by checking whether the failover could complete instead of requeuing blindly, and by not mistaking a pending marker for a missing pod. Pull request #11336
- Fixed a deadlock during multi-instance cluster restore from orphan PVCs where a replica's bootstrap would hang waiting for certificates that the restore gate prevented from being set. Pull request #11492
- Fixed failover to immediately shut down an unreachable old primary instead of waiting for checkpoint and fast-shutdown attempts that would never complete, so it can rejoin as a replica. Pull request #11347
- Fixed fencing to only take out a single instance instead of preventing the entire cluster from electing a new primary or recreating deleted pods. Pull request #11274
- Fixed a window during primary restart or reschedule where PostgreSQL starts before acquiring the primary lease, which could allow multiple primaries. Pull request #11356
- Fixed plugin status reporting to include the PostgresCapabilities field that was being dropped during updates. Pull request #11447
- Made the disk-space e2e test deterministic by stopping the postmaster once the disk is full, preventing timeouts from harmless temporary file writes. Pull request #11490
- Fixed pgbouncer SHOW POOLS parsing to bind columns by name instead of relying on column count, avoiding failures when future pgbouncer releases add new columns. Pull request #11464
- Fixed preflight CI to create image-pull secrets in the randomized namespace that preflight v1.20.0+ now uses. Pull request #11472
- Fixed Rook CRD race on AKS by waiting for CRD registration before applying custom resources. Pull request #11479
Why it matters
This week addresses critical failover and recovery scenarios that can leave clusters unavailable or stuck in deadlock. Fixes to fencing, primary lease acquisition, and restore operations ensure the cluster can reliably handle failures and maintenance events. These changes make the controller more resilient when instances are unreachable or the cluster needs to reorganize.
Changelog entry
- Fixed controller avoiding stalling or reverting an in-progress failover Pull request #11336
- Fixed certificate-wait deadlock in orphan-PVC restore Pull request #11492
- Fixed demotion of unreachable old primary immediately during failover Pull request #11347
- Fixed cluster recovery while an instance is fenced Pull request #11274
- Fixed PostgreSQL start to gate on holding the primary lease Pull request #11356
- Fixed plugin status reporting of PostgreSQL capabilities Pull request #11447
- Made disk-space e2e test deterministic Pull request #11490
- Fixed pgbouncer SHOW POOLS column binding by name Pull request #11464
- Fixed preflight image-pull secret creation in randomized namespace Pull request #11472
- Fixed Rook CRD race on AKS Pull request #11479
- Fixed incorrect field name in wal_archiving documentation Pull request #11477
This week: failover no longer stalls, fencing affects only the target instance, restore deadlocks are fixed, and primary startup is safer. Critical reliability improvements for PostgreSQL on Kubernetes.
This week we fixed several critical reliability issues in CloudNativePG: failover operations no longer stall or revert, fencing now correctly isolates only the target instance, restore from orphan PVCs no longer deadlocks, and primary startups now properly acquire the lease first. These changes ensure clusters can reliably recover from failures and maintenance events.