Files
rustfs/docs/architecture/storage-control-data-plane.md
T
2026-06-21 11:25:44 +08:00

64 lines
2.2 KiB
Markdown

# Storage, Control Plane, And Background Controllers
This document defines migration boundaries for the storage hot path and adjacent
control-plane responsibilities.
## Storage API Contracts
Storage API contracts must not absorb implementation details from ECStore or the
reader pipeline.
Out of scope for the contract layer:
- KMS/SSE implementation.
- Range and compression behavior.
- Erasure coding and bitrot logic.
- Remote disk transport and recovery.
No-drift behavior:
- Object-to-set hash remains unchanged.
- Write quorum remains unchanged.
- Reader decryption, etag/checksum, version, and delete-marker behavior remain
unchanged.
- Public compatibility paths remain available through temporary re-exports or
wrappers during pure moves.
## Cluster Control Plane
ClusterControlPlane starts as a read-only facade inside `crates/ecstore/src/cluster`.
Do not create a standalone cluster crate until internal dependencies are stable.
Initial scope:
- Topology snapshot.
- Membership snapshot.
- Lock registry snapshot.
- Peer health snapshot.
- Pool state snapshot.
The first read-only implementation lives behind `rustfs_ecstore::api::cluster`.
It maps existing endpoint pools into the shared storage-api topology contract and
an ECStore-owned static membership snapshot. It must not expose local disk paths,
start health checks, mutate endpoint ownership, or change placement/readiness.
The same facade also owns static pool-state, local-node storage, and peer-health
status projections. Peer health remains explicitly unknown until a later slice
wires real health signals; this document does not authorize background probes or
RPC-based health checks.
Risk controls:
- Distributed lock quorum remains per set.
- RemoteDisk suspect/offline/recovery, timeout, and connection eviction semantics
must not be simplified.
- Health impact behavior must be feature-gated if it changes production behavior.
## Background Controllers
Scanner, heal, lifecycle, replication, config reload, metrics, and auto-tuning
controllers should move behind explicit controller boundaries after lifecycle
contracts are stable.
The first controller work should be read-only status and shutdown ordering, not
behavior changes.