GraphRecords

GraphRecords stores entities and their relationships as a graph. Nodes hold attributes. Edges connect nodes and can also hold attributes. Groups organize subsets of nodes and edges.

A GraphRecord is immutable. Every method that changes data returns a new GraphRecord and leaves the old one untouched. This makes records safe to share, cheap to snapshot, and free of locks.

When to Use GraphRecords

GraphRecords fits problems where:

• Data has natural relationships (users and products, documents and citations, components and dependencies)
• You need to query based on relationships ("find all users connected to products over $100")
• Different entity types have different attributes (users have age, products have price)

Installation

pip install graphrecords

Building a Graph

Since every call returns a new record, building is a chain:

import graphrecords as gr

# Nodes are tuples: (index, {attributes})
record = (
    gr.GraphRecord()
    .add_nodes_in_group(
        [
            ("alice", {"age": 30}),
            ("bob", {"age": 25}),
            ("carol", {"age": 35}),
        ],
        "users",
    )
    .add_nodes_in_group(
        [
            ("widget", {"price": 10.0}),
            ("gadget", {"price": 25.0}),
        ],
        "products",
    )
    # Edges are tuples: (source, target, {attributes})
    .add_edges_in_group(
        [
            ("alice", "widget", {"quantity": 1}),
            ("bob", "gadget", {"quantity": 2}),
            ("alice", "gadget", {"quantity": 1}),
        ],
        "purchases",
    )
)

add_nodes and add_edges do the same without a group.

You can also load Polars DataFrames, naming the index columns:

import polars as pl

users = pl.DataFrame({"id": ["alice", "bob"], "age": [30, 25]})
record = record.add_nodes((users, "id"))

purchases = pl.DataFrame({"user": ["alice"], "product": ["widget"], "qty": [1]})
record = record.add_edges((purchases, "user", "product"))

Anything that exports an Arrow stream works the same way.

Accessing Data

record.node_indices()  # ['alice', 'bob', 'carol', 'widget', 'gadget']

# node() returns a view of one node
alice = record.node("alice")
alice.attribute("age")  # 30
alice.attributes()  # {'age': 30}
alice.neighbors()  # ['widget', 'gadget']
alice.degree()  # 2

# group() returns a view of one group
record.group("users").nodes()  # ['alice', 'bob', 'carol']

# Edge indices are opaque values, not integers
edge_index = record.edge_indices()[0]
record.edge(edge_index).attributes()  # {'quantity': 1}

Querying

Queries are expressions. gr.nodes() and gr.edges() start free expressions that are not tied to any record; record.nodes() binds one to a record, which makes it a series. A series only runs when you call evaluate().

adult_users = record.nodes().filter(
    gr.nodes().in_group("users") & (gr.nodes().attribute("age") > 25)
)
list(adult_users.evaluate())  # ['alice', 'carol']

Expressions traverse the graph themselves — the chain reads as the question it asks:

bulk_buyers = (
    record.edges()
    .filter(gr.edges().in_group("purchases") & (gr.edges().attribute("quantity") >= 2))
    .source_node()
)
list(bulk_buyers.evaluate())  # ['bob']

And they aggregate:

record.nodes().filter(gr.nodes().in_group("users")).attribute("age").mean().evaluate()
# 30.0

Every operation is checked twice: the type checker rejects operations that do not fit the current shape of the expression while you write, and the engine optimizes the query before running it. explain() shows the optimized plan, explain_unoptimized() the raw one.

The result of evaluate() is consumed by iterating it once. Evaluate again for a fresh result, or collect into a list first.

Schema

Schemas define what attributes are allowed and their types.

Inferred mode (default): The schema learns from data as you add it. Any attribute is allowed.

Provided mode: The schema is fixed. Data that doesn't match is rejected.

from graphrecords.schema import AttributeDataType, AttributeType, GroupSchema, Schema
from graphrecords.datatype import Int, String

schema = Schema(
    groups={
        "users": GroupSchema(
            nodes={
                "age": AttributeDataType(Int(), AttributeType.Continuous),
                "name": AttributeDataType(String(), AttributeType.Unstructured),
            }
        )
    }
)

record = gr.GraphRecord.with_schema(schema)
record = record.freeze_schema()  # Switch to provided mode

# Now adding a user without 'age' or 'name' raises an error

Schemas are immutable like records: set_node_attribute, add_group, freeze, and the other schema methods all return a new schema.

Plugins

A plugin hooks into every change made to a record. A hook sees the pending change and can let it through, or return a modified one:

class Audit(Plugin):
    def pre_add_nodes(self, record, addition):
        print(f"adding {len(addition.batch)} nodes")


record = record.add_plugin("audit", Audit())

Hooks exist for every change (pre_add_nodes, post_remove_edges, ...). A plugin only pays for the hooks it defines. Returning None from a hook applies the change unchanged; returning a change replaces it, and returning a list of changes replaces it with several (an empty list drops it). The post_* hooks observe the records before and after together with the applied change, and can veto by raising.

Serialization

Save and load graphs using RON format:

record.to_ron("graph.ron")
loaded = gr.GraphRecord.from_ron("graph.ron")

Export to DataFrames, one pair of tables per group plus the ungrouped rest:

export = record.to_polars()
export["ungrouped"]["nodes"]  # a Polars DataFrame
export["groups"]["users"]["nodes"]

record.to_arrow()  # the same shape, as Arrow tables

Documentation

• User Guide
• API Reference

Background

GraphRecords started as MedRecord in the medmodels library. We realized it has applications beyond the medical domain and published it as a standalone library.

License

MIT. See LICENSE.