All arrow2 arrays are defined roughly as the following:

pub struct Array {
    data_type: DataType,
    values: Buffer<T>,
    validity: Option<Bitmap>,
}

When you clone/slice/index an Array, you get another Array in roughly O(1) thanks to both the values and validity bitmaps being refcounted behind the scenes:

pub struct Buffer<T> {
    data: Arc<Bytes<T>>,
    offset: usize,
    length: usize,
}

pub struct Bitmap {
    bytes: Arc<Bytes<u8>>,
    offset: usize,
    length: usize,
    unset_bits: usize,
}

Well... not really, turns out the DataType is not refcounted, and it can get huge: it's a massive heap-recursive enum potentially filled with strings and such.

Say you have a ListArray that contains a bunch of StructArrays (i.e. a column of component data) and you want to extract references to the individual StructArrays in that list (i.e. the individual DataCells): each of these arrays is now going to carry a full copy of the StructArray's schema.

For tiny DataCells (which are very common in Rerun), the overhead is enormous.