Phase 2: Cloud provider abstraction and snapshot_utils refactor

[wdvr]

Summary

This PR implements Phase 2 of the cloud-agnostic migration plan:

• New providers/ module with abstract CloudProvider and AuthProvider interfaces
• AWSProvider implementation wrapping existing boto3 code
• GCPProvider and CustomProvider** stubs with extensive documentation
• Refactored snapshot_utils.py to use provider interface instead of direct boto3 calls

Provider Interface

from providers import get_cloud_provider

provider = get_cloud_provider()  # Returns AWS/GCP/Custom based on CLOUD_PROVIDER env

# Storage operations
volume = provider.create_volume(size_gb=100, availability_zone='us-east-2a')
snapshot = provider.create_snapshot(volume.volume_id, tags={'user': 'john'})
snapshots = provider.list_snapshots(filters={'gpu-dev-user': 'john'}, status=['completed'])

# Object storage
uri = provider.upload_to_object_storage('bucket', 'key', content, content_type='text/plain')

Changes to snapshot_utils.py

Before	After
import boto3; ec2_client = boto3.client("ec2")	from providers import get_cloud_provider; provider = _get_provider()
ec2_client.describe_snapshots(Filters=[...])	provider.list_snapshots(filters={...}, status=[...])
ec2_client.create_snapshot(VolumeId=...)	provider.create_snapshot(volume_id, tags={...})
s3_client.put_object(Bucket=..., Key=...)	provider.upload_to_object_storage(bucket, key, content)

Next Steps

• disk_reconciler.py needs refactoring (requires adding create_tags() / delete_tags() to interface)
• Phase 3: K8s-native VolumeSnapshot support
• Phase 4: GCP implementation (separate PR)
• Phase 5: OIDC integration (separate PR)

Test plan

• Unit tests for provider interface
• Verify snapshot operations work with AWS backend
• Test pagination for users with many snapshots

🤖 Generated with Claude Code