DreamLake

Native Experiment aggregation

DreamLake computes cross-run metric statistics from native project Tracks. This endpoint reads DreamLake data only; historical Dash runs continue to use Dash. It requires a signed-in user with read access to the project.

POST /native-experiments/ns/:namespace/projects/:project/aggregate-series accepts:

json
{
  "experiments": [
    {"experiment_id": "<run-id>", "metrics": [{"metric_id": "<track-id>", "metric_name": "train"}]},
    {"experiment_id": "<run-id>", "metrics": [{"metric_id": "<track-id>", "metric_name": "train"}]}
  ],
  "x_key": "train.step",
  "y_keys": ["train.loss"],
  "alignment_method": "interpolate",
  "bins": 200,
  "agg_functions": ["MEAN", "STD", "PERCENTILE_25", "PERCENTILE_75"]
}

Get each run ID from GET .../runs and its Track ID from GET .../runs/:runId/metrics. metric_name prefixes the Track's logged column names in x_key and y_keys; use an empty name for bare columns. Runs and Tracks must all belong to the requested project. The response has series, each with an x array and the requested statistic arrays, plus metadata.

Alignment may be interpolate, common_steps, or none. Supported statistics are MEAN, MEDIAN, STD, VARIANCE, MIN, MAX, PERCENTILE_25, PERCENTILE_75, and COUNT; standard deviation and variance use the population formula. As in ml-dash, a request may include up to 100 runs (more returns 413 NATIVE_AGGREGATE_TOO_LARGE) and 1 to 10,000 interpolation bins, with no limit on Tracks, output columns or stored points. With interpolate, values are interpolated linearly on bins evenly spaced x values across all runs and are null outside a run's x range; when every x is an integer, the reported x values are then rounded half to even. Missing or unreadable Track chunks fail the request instead of returning partial statistics.