Arrow and Parquet import

Build NodeTable and EdgeTable values, combine them with GraphData::try_new, then call Session::replace_graph_data. This replaces the selected open graph atomically; it is not an append operation. In an explicit transaction it stages the replacement until COMMIT.

Each table has one exact label set and property layout. Edge tables also have a directed flag. Multi-label elements occupy one table row, retaining one identity.

Column	Required type and meaning
__gf_id	Non-null UInt64 ID, unique per graph and node/edge kind
__gf_source	Non-null UInt64 source node ID on an edge table
__gf_destination	Non-null UInt64 destination node ID on an edge table
Property columns	Boolean, Int64, Float64, Utf8 or Binary

Import validates IDs across batches/tables, endpoints, labels, column names, schemas and property types. The __gf_ prefix is reserved. A property name must have one consistent scalar type within the same graph and element kind. Nodes and edges may use the same numeric ID because their identity kinds differ.

CLI manifest

{
  "nodes": [{"labels": ["Person"], "file": "people.parquet"}],
  "edges": [{"labels": ["Knows"], "file": "knows.parquet", "directed": true}]
}

Paths are relative to the manifest file. After creating an open graph:

cargo run -p graphfusion --locked -- import --database ./demo-db \
  --graph social --manifest ./data/graph.json

Persistent import reads and validates the external files, then writes database-owned immutable Parquet files. The database does not rely on the caller keeping the original files unchanged. Import materializes input; it is not a streaming ingest API. Persistent database paths must be UTF-8 without ASCII control characters.

The Rust social example builds actual Arrow tables. CSV/JSON graph-table import and typed graph import are not implemented.