Graph Engine

How the deterministic graph engine stores and queries data.

The graph engine is the core of Kremis. It stores nodes (entities), edges (relationships), and properties (attributes) in a fully deterministic structure.

Data Structures

All collections use BTreeMap for deterministic iteration order — no HashMap is used anywhere in the core.

pub struct Graph {
    nodes:        BTreeMap<NodeId, Node>,
    edges:        BTreeMap<NodeId, BTreeMap<NodeId, EdgeWeight>>,
    entity_index: BTreeMap<EntityId, NodeId>,
    properties:   BTreeMap<NodeId, BTreeMap<Attribute, Vec<Value>>>,
    next_node_id: u64,
}

Storage Backends

In-Memory

Graph struct in RAM. Fast, volatile. Used for testing and temporary sessions.

Persistent (redb)

ACID transactions, crash-safe. Copy-on-write B-trees with MVCC concurrent readers.

RedbGraph Tables

Table	Key	Value	Purpose
NODES	u64	&[u8] (postcard)	NodeId → serialized Node
EDGES	(u64, u64)	i64	(from, to) → weight
ENTITY_INDEX	u64	u64	EntityId → NodeId
METADATA	&str	u64	Counters (e.g. next_node_id)
PROPERTIES	(u64, u64)	&[u8] (postcard)	(node_id, attr_hash) → (Attribute, Vec<Value>)

Query Algorithms

Method	Algorithm	Details
compose	BFS	VecDeque queue, bounded by depth (max 100)
compose_filtered	BFS + weight filter	Skips edges below min_weight
strongest_path	Dijkstra	Cost = i64::MAX - weight (higher weight = preferred)
intersect	Set intersection	Neighbors of first node, intersect with remaining
related_context	BFS	Contextual alias for compose

All traversals return an Artifact containing the path and optional subgraph edges.

Export Formats

Canonical (bit-exact)

[header_len: u32 LE] [CanonicalHeader: postcard] [CanonicalGraph: postcard]

• Magic: b"KREX", version 2
• Checksum: XOR-based deterministic hash
• Import limits: 1M nodes, 10M edges (DoS protection)
• V1 backward compatibility (imports without properties)

JSON

SerializableGraph with serde — nodes, edges, next_node_id, properties.