Value filter support in NDArray has been proposed:
Is there a use case to motivate it?
This effectively requires that all read() return batches be sparse, as the value filter will remove values.
Where the requested batch_format is dense (i.e., the user wants a tensor back), this would require that we also provide coordinates and/or a mask in addition to the tensor (values). Or disallow that combination: if you specify a value filter, you can only ask for a sparse-capable batch_format.
Created issue as follow-up to a removed "TODO" in the abstract spec, per https://github.com/single-cell-data/TileDB-SOMA/issues/834#issuecomment-1453953693.
Value filter support in
NDArrayhas been proposed:read()return batches be sparse, as the value filter will remove values.batch_formatisdense(i.e., the user wants a tensor back), this would require that we also provide coordinates and/or a mask in addition to the tensor (values). Or disallow that combination: if you specify a value filter, you can only ask for a sparse-capablebatch_format.