DreamLake

Native Experiment run metadata

A native run has a human description and tags. Projects whose GET .../status lists the run-metadata capability accept them on create and on update. Both writes require a signed-in user with write access to the project, or a write session that user minted. A job session cannot set them and gets 403 NATIVE_FORBIDDEN.

PUT /native-experiments/ns/:namespace/projects/:project/runs/:runKey creates or returns the run, and PATCH .../runs/:runId updates it:

json
{"description": "lr sweep baseline", "tags": ["lr-sweep", "v2"]}
FieldAbsentnull / "" / []Value
descriptionunchangednull or "" clears ita string of at most 2,000 code points
tagsunchanged[] clears themreplaces the whole set

Tags are trimmed and de-duplicated in order; there can be at most 20 tags of at most 50 code points each. Writing the stored values again changes nothing, so a repeated PUT is safe. PATCH also accepts archived (a boolean). A change to description or tags restores an archived run unless the same body sets archived: true. A write-protected run rejects a change with 409 NATIVE_RUN_WRITE_PROTECTED, but can still be archived.

The same rules apply wherever the run is stored: in the main DreamLake database, or in the separate Dash database used for new runs when the server runs the native Dash store. A run keeps its storage, so a metadata write never moves it.

An invalid value returns 400 NATIVE_INVALID_RUN_METADATA. Any other field returns 400 NATIVE_INVALID_RUN on PUT and 400 NATIVE_INVALID_ARCHIVE on PATCH. Every run response and the run list return description (null when unset) and tags ([] when unset).

Additional metadata

The API also accepts a metadata object or null on run creation and updates. GraphQL Experiment.metadata and its experiment Node.metadata return that value from the run's original storage. me, namespace and project metadata use DreamLake's canonical records; workload and job credentials do not gain project or account metadata access.

Run, file and folder PATCH routes also accept priority, a signed 32-bit integer. GraphQL tree children put higher priorities first. Files initially use ml-dash's defaults: .dashrc 200, README.md 100 (case-insensitive), and other names 0. Renaming or moving a file retains its priority. File metadata can be any JSON value; run metadata remains an object or null.

Project folders

POST /native-experiments/ns/:namespace/projects/:project/folders and PATCH …/folders/:folderId also accept description, tags, metadata (any JSON) and priority (a signed 32-bit integer). Missing fields stay unchanged; null clears the description or metadata, and [] clears tags. These values stay with the folder's ID through renames and moves. GraphQL updateNode edits its description and tags. Writes require a person's WRITE access and refuse a protected, trashed or pending folder. Empty project folders are retained.

Inside a run, GET/POST …/runs/:runId/folders and GET/PATCH/DELETE …/runs/:runId/folders/:folderId manage file folders, including empty ones. Create with { "path": "/checkpoints/empty" }; PATCH accepts the same metadata fields and an optional new path. A move keeps folder and file IDs. Delete removes the subtree permanently. These operations stay in the run's original storage, honor its write protection, and restrict a run-bound writer or job to that run. The server advertises them as file-folders.

Run file prefixes accept Unicode, spaces, commas and percent signs. For example, /实验 记录/a,b/%2C keeps those names literally through uploads, folder moves, file listings and downloads. File names are single path segments; control characters and . or .. segments are rejected. Extended paths use the same lifecycle rollout gate as file folders.