fix: GCS composite object has no MD5 hash

[yuechao-qin]

Summary

The orchestrator crashes with AttributeError: 'NoneType' object has no attribute 'encode' when collecting output artifacts from GCS that are composite objects. Composite objects (created by parallel uploads or compose operations) have no MD5 hash — blob.md5_hash is None.

The crash happens in the upstream SDK (cloud-pipelines==0.26.3.12) at GoogleCloudStorageProvider._get_info_from_uri, during output collection after a container completes. This is before any data reaches our code, so there's no way to fix it at the save/consume layer.

Fix

• New file: patched_google_cloud_storage.py — subclasses the SDK's GoogleCloudStorageProvider and overrides _get_info_from_uri to handle None MD5 gracefully
• When MD5 is missing, substitutes "no_md5_<UTC timestamp>" as the hash value — this forces a cache miss for any downstream task consuming that artifact (Solution A: safe, no false cache hits)
• Updated launchers (google_kubernetes_launchers.py, kubernetes_launchers.py) to use PatchedGoogleCloudStorageProvider instead of the upstream provider

Failing runs

• https://oasis.shopify.io/runs/019e8e17de7041a0dd8c
• https://oasis.shopify.io/runs/019e8f7be22321296b8a

Trade-offs

• Composite object outputs will always cause a cache miss for downstream tasks — this is intentional and safe
• Long-term fix: compute our own MD5 by downloading the file (cost scales with file size)

Tests

7 unit tests added covering:

• Normal blobs return correct MD5 hex digest
• Composite blobs return unique timestamp fallback
• Timestamps are unique across calls
• Warning is logged for composite objects
• Provider handles single files, composite files, and mixed directories

[yuechao-qin]

• fix: GCS composite object has no MD5 hash #271 👈 (View in Graphite)
• master

This stack of pull requests is managed by Graphite. Learn more about stacking.