pvtend.tendency.open_months_ds

pvtend.tendency.open_months_ds(data_dir: Path, var_list: list[str], month_keys: list[tuple[int, int]], engine: str = 'netcdf4', chunks=None) → Dataset

Open multiple months of ERA5 data as a single dataset.

chunks=None (default) loads each month eagerly — fine for the compute path that touches most timesteps of the window. Pass a dask chunk spec (e.g. {"valid_time": 1}) to open lazily and only read the timesteps actually accessed — the PPVI path uses this so a worker materialises one (or a few) timesteps instead of the whole ~15 GB month, keeping per-worker RSS low enough to scale to many workers under the 2 TB cap.