DB Schema

Key Design Decisions

IDs

All entities use MongoDB ObjectId (auto-generated, stored as String via @db.ObjectId).

Soft Deletes

All user-facing models carry deletedAt DateTime?. Active records have deletedAt IS NULL. This enables undo/recovery and audit trails without permanent data loss.

Asset Folder Integration

Video, Audio, LabelTrack, and TextTrack each have:

Field	Description
`folderId`	→ Node.id (null = project root level)
`filename`	Basename only (e.g. `"run01.mp4"`)
`name`	Full path (e.g. `"/camera/front/run01.mp4"`) — kept for backward compatibility

On asset creation, the server parses name into folder path + filename, auto-creates the folder hierarchy via ensureFolderHierarchy, and sets both folderId and filename.

Track Scoping

Tracks can be episode-scoped (raw recorded data) or space-scoped (generated/processed).

Track.episodeId = null → space-level track
Track.episodeId = "..." → episode-scoped track
S3 paths differ accordingly: space-root-relative vs. under episodes/{id}/

Track Chunk Manifest

Each track stores its chunks as Json on the record itself (no satellite chunk table):

[
  { "path": "00001", "src": "abc123def456" },
  { "path": "00002", "src": "789abc012def" }
]

Key	Description
`path`	Local sequential key (ordering/identity)
`src`	Content-addressed blob reference (S3 key)

Parameters

Flat key-value store with dot-separated keys. Nested dicts are flattened:

{ "model": { "lr": 0.001 } }  →  { "model.lr": 0.001 }

Each write is journaled as a ParameterEntry for audit trail. The data field on Parameters holds the current merged snapshot.

Storage Layer

All binary storage is delegated to BSS (Big Streaming Server). DreamLake Server is purely a metadata index — it does not manage S3 credentials, buckets, or storage backends. When multi-backend support is needed (S3, GCS, OSS), it will be implemented in BSS via a StorageAdapter interface.