Yes, it caches each step of the pipeline and only recomputes the parts downstream from the parameter that changed. It’s different from existing HoloViews and Panel pipelines in that it evaluates eagerly. This means that it always caches the most up-to-date result based on the parameters/widgets that feed into the pipeline. This is required so that we can accurately mirror the API of the object that is being wrapped. For very expensive calculations it might therefore not be entirely suitable but in those scenarios you could switch to dask, which would ensure that the actual execution is still lazy.
PyData 2021 - Build polished, data-driven applications directly from your Pandas or XArray pipelines
2 Likes