Delta: Highly available, strongly consistent storage using chain replication (2022)
a day ago
- Meta created Delta, a low-dependency object storage system for critical workloads and disaster recovery.
- Delta prioritizes simplicity and reliability over latency, efficiency, and storage density.
- It uses chain replication to ensure strong consistency, with reads from the tail and writes pipelined through the chain.
- Chain replication is compared with quorum replication: simpler but less efficient and higher write latency.
- Delta buckets consist of chains of servers spread across failure domains for durability.
- Failure detection uses heartbeats and voting; hosts are removed and later repaired via automated processes.
- Apportioned queries allow any node to serve reads by verifying committed versions with the tail, improving scalability.
- Automated repair via a control plane service maintains failure domain balance and chain distribution.
- Global replication supports synchronous replication to some regions and asynchronous to others for efficiency.
- Delta integrates with archival services for backup and restore, ensuring recoverability in degraded environments.
- Future work includes a centralized backup/restore service for core infrastructure at Meta.