Consistency Model

Write Consistency

A write is acknowledged only after:

WAL append is persisted (O_DIRECT + fsync)
Raft quorum commit (for replicated namespaces)

Single-node: linearizable writes for the shard leader. Replicated: linearizable writes within each Raft group.

Read Consistency Levels

-- Strong: leader-read after commit index >= required LSN (default for metadata/constraints)
SET read_consistency = 'strong';

-- Bounded staleness: follower read allowed if lag <= duration
SET read_consistency = 'bounded_staleness_5s';

-- Eventual: local read for CRDT edge sync (monotonic convergence)
SET read_consistency = 'eventual';

Level	Behavior	Use case
Strong	Leader read, waits for commit index	Metadata, constraints, OLTP
Bounded Staleness	Follower read if lag within configured bound	Read replicas, analytics
Eventual	Local read, monotonic convergence	CRDT edge sync, NodeDB-Lite

Snapshot Isolation

Cross-engine queries execute against a shared snapshot watermark. A query that combines vector search, graph traversal, and document field access sees a consistent point-in-time view across all engines.

Default isolation: Snapshot Isolation (SI) per query. Write skew detection is enforced for UNIQUE and FK constraints at commit.

Cross-Model Queries

RRF fusion and cross-engine joins include source watermarks in response metadata. This lets clients verify the freshness of results that span multiple engines.