Describe the enhancement requested
We are looking at implementing BFloat16 data type as an extension type based on a FSB[2] storage type. In Python we can write the repr method for the array, but ChunkedArray and Table will print the storage type, which looks like nonsense to users:
>>> tab = pa.table({ "x": bfloat16_array([1.1, None, 3.4]) })
>>> tab
pyarrow.Table
x: extension<lance.bfloat16<BFloat16Type>>
----
x: [[8D3F,null,5A40]]
>>> tab['x']
<pyarrow.lib.ChunkedArray object at 0x12a1f3a60>
[
[
8D3F,
null,
5A40
]
]
>>> tab['x'].chunk(0)
<lance.arrow.BFloat16Array object at 0x0000000129818a60>
[1.1015625, None, 3.40625]
lance-format/lance#1002
Component(s)
C++