Can networkx edge attributes survive gpickle loads across Python upgrades?

Deployment Failure Diagnosis

The goal is to identify why a recent deployment fails only after upgrading the runtime from Python 3.8 to 3.12. Logs point to missing edge attributes, yet the graph was serialized with nx.write_gpickle and deserialized with nx.read_gpickle in the new environment. The core uncertainty lies in whether the Gpickle format guarantees preservation of custom edge attributes across Python version changes, especially when the graph is rebuilt from edge data rather than the original object graph.

Constraints include the lack of explicit warnings from networkx when required attributes are absent, nondeterministic edge ordering in adjacency lists, and the absence of documented compatibility guarantees for pickle files between Python releases. This creates silent failures that only surface at runtime.

Which edge attributes are most at risk of being dropped during a gpickle round‑trip when upgrading from Python 3.8 to 3.12? Does networkx emit any diagnostic messages when required edge attributes are missing after loading a gpickle file? Can we programmatically verify that all expected edge attributes are present after deserialization?