Data Redundancy Explained: Why Replicas Alone Cannot Prevent Data Loss

Data redundancy involves storing multiple copies of a dataset across different nodes so that a single node failure does not result in permanent data loss. The most common mechanism is replication, where a primary node streams write-ahead log records to replica nodes that apply changes locally. However, most real-world data loss incidents stem not from the absence of replicas, but from replicas falling out of sync before a primary failure or broken replication going undetected due to missing alerts. GitLab's January 2017 incident illustrates this clearly: an engineer accidentally deleted the primary data directory while all replicas had already stopped streaming, causing several hours of data to be permanently lost despite the system nominally having replicas. Replication strategies range from asynchronous, which prioritises speed but risks losing recently committed transactions, to synchronous and quorum-based approaches, which offer stronger durability guarantees at the cost of higher write latency.