Disaggregated storage

CockroachDB can offload its deepest LSM levels to shared cloud object storage. The on-node Pebble database keeps L0–L4 hot, while L5/L6 SSTables move to S3/GCS/Azure and are referenced by every replica that owns them.

Why

Inter-AZ replication traffic dominates infra cost on cloud platforms. Shared storage lets multiple replicas of the same range read the same SST without re-pushing it across the network at compaction time. Disaster recovery and bulk operations also benefit — a snapshot or restore can re-link existing files instead of re-shipping them.

Code

• pkg/storage/shared_storage.go — engine-level entry: register a remote storage and tell Pebble to use it for new SSTs above a configured level.
• pkg/storage/external_sst_reader.go — read SSTs that live outside the local LSM.
• pkg/cloud/ — adapters for cloud object stores (S3, GCS, Azure, NFS, …) used here and by backup, IMPORT, changefeeds.
• Pebble's "shared storage" support (upstream) is the engine-side counterpart.

Cluster settings

Selected knobs (current names may change; check pkg/storage/storage_config.go and pkg/cloud/):

• storage.shared_storage — URI and options.
• storage.shared_storage.uploads.batch_size — batching at compaction.
• Per-store --shared-storage flag on cockroach start.

Related pages

• Storage — the engine layer.
• features/backup-restore — biggest beneficiary.