dataset

The remote dataset identity, or None for local observations.

Return a lazy time window, including start and excluding end.

Strings and datetimes are accepted; timezone-free bounds mean UTC. Explicit bounds replace those on this view. Remote selection happens on the server and requires its dataset selection capability.

Return a lazy view aggregated into fixed UTC-epoch bins.

Supported reducers are mean, first, last, min, max, sum, and count. Time filtering precedes aggregation; missing bins are omitted and no values are forward-filled. Bin labels are their left edges, which can precede a selection's start. Remote aggregation reduces transferred rows and requires a platform with dataset selection support.

Select indexed files and bounded observations without materialization.

Variables are filename stems reported by the source catalog. Path bounds are lexical integration paths; row bounds are per file. Sample bounds slice the first array dimension inside each observation, preserving all remaining dimensions. End bounds are exclusive. Steps subsample by stride; they do not average, filter or invent timestamps.

Iterate immutable source fingerprints without loading their contents.

Resolve source identity and schema without loading all observations.

Yield bounded, ordered batches pinned to one resolved source version.

Each remote page is checkpointed before yielding. A new iterator replays saved pages, then resumes at the saved cursor. Close the iterator when stopping early to release the download lock. progress=False silences tqdm in applications and automated jobs.

Clear this selection's saved pages without contacting the host.

Pages are shared by all prefix lengths. max_rows also selects an old-format materialized prefix entry to remove, when one exists.

Materialize a dataset or view, resuming saved pages automatically.

Every load resolves current metadata and permissions. Interrupted downloads resume at a pinned source version, even beyond cache_ttl. Without a content validator this is a resumable local snapshot, not a guarantee against upstream corrections. Completed downloads require a matching content token or an explicit TTL; cache_ttl=None opts into snapshot reuse until refresh or a source version change.

refresh=True restarts the download. max_rows caps output rows after selection/aggregation; whole downloaded pages are retained so a later larger request can reuse their tails. iter_batches uses the same checkpoints without materializing the full dataset in memory.

Create an Anomx TimeSeries, optionally resampling irregular observations.

The native forecasting container uses regular, timezone-free indices; aware times are normalized to UTC and their original timezone is stored in metadata. Resampling leaves empty bins as NaN unless fill is explicitly selected. Without a frequency, irregular data is rejected by TimeSeries; the original observations remain available as a frame.

Plot observations on a Matplotlib axes and return that axes.

Load and return the observations as a DataFrame.

Return observations as a NumPy array.

Create a local dataset, preserving a supplied temporal index.

Create a regularly indexed time series from a 1D or 2D array.

Resolve recorded channels through an API connection or host loader.

Return the first temporal index value; materializes a remote selection.

Return the last temporal index value; materializes a remote selection.

Convert temporal data for optional Darts models without a core dependency.