Erasure coding not replication
10 data shards + 4 parity shards. Can lose any 4 shards, recover. More storage-efficient than 3x replication (1.4x vs 3x overhead).
Advertisement
Erasure coding not replication
10 data shards + 4 parity shards. Can lose any 4 shards, recover. More storage-efficient than 3x replication (1.4x vs 3x overhead).
Advertisement
Multi-AZ placement
Shards distributed across 3+ AZs. Whole AZ can fail — object still retrievable.
Index at scale
bucket/key → shard locations mapped in Index Service. Petabytes of index. Sharded + replicated. Fast lookup by prefix.
Continuous repair
Bit rot happens. Repair service scans shards, verifies checksums, replaces bad ones. Silent healing.
Storage classes as lifecycle
Standard → IA → Glacier → Deep Archive. Auto-transition via lifecycle rules. Trade retrieval latency for cost.