Make Record pickle-safe (fix RecursionError via GCS IO manager)

[jirhiker]

Problem

Materializing a downstream asset (e.g. nm_waterlevel_trends) fails loading an upstream input:

RecursionError: maximum recursion depth exceeded
  backend/record.py, line 82, in __getattr__
    v = self._payload.get(attr)   [repeated 980+ times]

Record.__getattr__ reads self._payload. During unpickling (the GCS pickle IO manager passing data between assets), pickle probes getattr(obj, "__setstate__") before __dict__ is restored, so _payload doesn't exist yet → __getattr__ recurses on __getattr__('_payload') until it blows the stack. Triggers whenever a Record — or a payload dict nesting one — crosses between assets.

Fix

Guard __getattr__:

• dunder lookups raise AttributeError (so pickle falls back to its default state restore);
• _payload is read via self.__dict__.get("_payload"), which can't re-enter __getattr__.

Verification

• Bare and nested Records now round-trip through pickle.dumps/loads.
• Full backend suite passes (263 tests).

Note

This was committed to the #84 branch after that PR was squash-merged, so it didn't reach main. Standalone fix here — affects every IO-manager product (major-chemistry, waterlevel-trends, etc.), so worth merging promptly.

🤖 Generated with Claude Code

[github-actions]

Your pull request is automatically being deployed to Dagster Cloud.

Location	Status	Link	Updated
die-orchestration		View in Cloud	Jun 26, 2026 at 09:09 PM (UTC)