FeatureTabulator#
- class rojak.atmosphere.feature_composite.extension.FeatureTabulator[source]#
Task side of a feature: tabulates its centres from the in-memory blocks of one time chunk.
It is sent to every chunk task, so it must hold parameters only, never a dask collection or a data-sized array. Implement it as a frozen dataclass: it then pickles cheaply and dask tokenises it deterministically from its fields.
It must also handle a chunk with no time steps, returning empty tables with the same columns and dtypes: the pipeline calls it once on such a chunk while building the graph, to find the columns and dtypes of the centre and pair tables without computing anything.
- __init__()#
Methods
__init__()