Can networkx edge attributes survive gpickle loads across Python upgrades?

Answer

Edge attributes that are picklable (e.g., ints, floats, strings, lists, dicts, NumPy arrays, etc.) survive a gpickle round‑trip when moving from Python 3.8 to 3.12. Attributes that contain non‑picklable objects (open file handles, locks, threads, or custom class instances whose definition changes or cannot be imported) may be lost or cause deserialization errors.

Explanation

NetworkX’s write_gpickle and read_gpickle rely on Python’s pickle module. Newer Python versions can read older pickle protocols, so the basic structure of the graph and any picklable edge data are preserved. However, pickle cannot reconstruct objects that depend on runtime state or class definitions that differ between environments; those attributes are either omitted or raise an AttributeError during load.

NetworkX does not emit any warning or log message when expected edge attributes are missing after a read_gpickle; the returned graph simply lacks those keys, which can lead to silent failures.

Verification steps

1. Load the graph in the target Python version:

import networkx as nx
G = nx.read_gpickle('my_graph.gpickle')

2. Determine the set of expected attribute names (from a schema, documentation, or the original graph before pickling).
3. For each expected attribute attr, retrieve its values:

attrs = nx.get_edge_attributes(G, attr)
missing = [e for e in G.edges() if e not in attrs]
if missing:
    print(f'Attribute {attr} missing on edges: {missing}')
else:
    print(f'Attribute {attr} present on all {len(G.edges())} edges')

4. Optionally verify types with isinstance checks.

Missing diagnostic detail

If your edge attributes contain custom class instances, please confirm whether those class definitions are available and unchanged in the Python 3.12 environment. This detail changes the recommendation: you may need to replace such attributes with picklable equivalents or ensure the class can be imported.