Preserving Non‑JSON‑Serializable Attributes in NetworkX JSON Serialization

Question

When migrating a small NetworkX application to a newer version without downtime, the JSON serialization helpers (`node_link_data`/`node_link_graph`) must retain all node and edge attributes.

However, the current API only preserves attributes that are JSON‑serializable; custom Python objects, lambdas, or mutable structures are omitted, and the format of the JSON output is not version‑locked, meaning upgrades may alter field names or nesting.

Uncertainty remains about how duplicate edges with differing attributes are handled in MultiGraphs and whether a stable schema exists for safe deserialization across releases.

How can non‑JSON‑serializable attributes be preserved during migration? Is there a documented version‑stable JSON schema for NetworkX graphs? What strategies exist for handling duplicate edge attributes in MultiGraph serialization?

Tasadduq Burney · Answer

Preserving Non‑JSON‑Serializable Attributes in NetworkX JSON Serialization
Non‑JSON‑serializable attributes are preserved by converting them into JSON‑compatible envelopes (type tags plus base64 or constructor data) before json.dumps, then running a restoration pass after node_link_graph.
Confirmed facts

node_link_data includes every attribute in the returned dict, but json.dumps does not.
No documented schema guarantees field stability across releases.
For MultiGraph and MultiDiGraph, one edge object per parallel edge is written, including a key field; parallel edges survive deserialization only when keys are unique and preserved.

Likely strategies for migration

Deep‑copy the graph, then walk nodes, edges, and graph attributes. Replace any non‑serializable value with an envelope such as {"_type": "custom", "_data": ""} or {"_type": "lambda", "_source": "inspect.getsource(lambda)"}.
Serialize with json.dumps using a default handler that recognizes the envelope format.
After node_link_graph, run a restoration pass that replaces envelopes back into the target attribute slots, using type information or constructor arguments.
Pin the NetworkX version in requirements.txt and store an application‑specific schema version. After round‑tripping, assert node/edge counts, parallel‑edge keys, and attribute key sets match the pre‑migration snapshot.

Parallel edges without deterministic keys are not safely round‑trippable. If your graph lacks explicit edge keys, assign them before serialization.
Missing diagnostic: Do any node or edge attributes contain lambdas, custom objects, or sets that must survive deserialization? This determines whether an envelope‑encoding restoration pass is required or if exclusion is acceptable.

Preserving Non‑JSON‑Serializable Attributes in NetworkX JSON Serialization

1 answer

Preserving Non‑JSON‑Serializable Attributes in NetworkX JSON Serialization

Your perspective belongs here.

1 question comment