Created issue as follow-up to a removed "TODO" in the abstract spec, per single-cell-data/TileDB-SOMA#834 (comment).
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.