10 hours ago
- A company experienced data loss when a Redis master with disabled persistence was restarted, causing a split-brain and replication of an empty dataset to all slaves.
- Redis Sentinel does not protect against misconfiguration like disabling persistence; this is a byzantine failure.
- Current Sentinel slave selection uses replication offset, which resets on restart and does not verify the master's runid, leading to potential promotion of outdated slaves.
- Proposed improvement: tie replication offset to the master's runid to ensure slaves have continuity with the failed master.
- Better Sentinel documentation is needed, including a list of common misconfigurations to avoid, such as turning off persistence.