arrow_data/

data.rs

// Licensed to the Apache Software Foundation (ASF) under one
// or more contributor license agreements.  See the NOTICE file
// distributed with this work for additional information
// regarding copyright ownership.  The ASF licenses this file
// to you under the Apache License, Version 2.0 (the
// "License"); you may not use this file except in compliance
// with the License.  You may obtain a copy of the License at
//
//   http://www.apache.org/licenses/LICENSE-2.0
//
// Unless required by applicable law or agreed to in writing,
// software distributed under the License is distributed on an
// "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
// KIND, either express or implied.  See the License for the
// specific language governing permissions and limitations
// under the License.

//! Contains [`ArrayData`], a generic representation of Arrow array data which encapsulates
//! common attributes and operations for Arrow array.

use crate::bit_iterator::BitSliceIterator;
use arrow_buffer::buffer::{BooleanBuffer, NullBuffer};
use arrow_buffer::{
    ArrowNativeType, Buffer, IntervalDayTime, IntervalMonthDayNano, MutableBuffer, bit_util, i256,
};
use arrow_schema::{ArrowError, DataType, UnionMode};
use std::mem;
use std::ops::Range;
use std::sync::Arc;

use crate::{equal, validate_binary_view, validate_string_view};

#[inline]
pub(crate) fn contains_nulls(
    null_bit_buffer: Option<&NullBuffer>,
    offset: usize,
    len: usize,
) -> bool {
    match null_bit_buffer {
        Some(buffer) => {
            match BitSliceIterator::new(buffer.validity(), buffer.offset() + offset, len).next() {
                Some((start, end)) => start != 0 || end != len,
                None => len != 0, // No non-null values
            }
        }
        None => false, // No null buffer
    }
}

#[inline]
pub(crate) fn count_nulls(
    null_bit_buffer: Option<&NullBuffer>,
    offset: usize,
    len: usize,
) -> usize {
    if let Some(buf) = null_bit_buffer {
        let buffer = buf.buffer();
        len - buffer.count_set_bits_offset(offset + buf.offset(), len)
    } else {
        0
    }
}

/// creates 2 [`MutableBuffer`]s with a given `capacity` (in slots).
#[inline]
pub(crate) fn new_buffers(data_type: &DataType, capacity: usize) -> [MutableBuffer; 2] {
    let empty_buffer = MutableBuffer::new(0);
    match data_type {
        DataType::Null => [empty_buffer, MutableBuffer::new(0)],
        DataType::Boolean => {
            let bytes = bit_util::ceil(capacity, 8);
            let buffer = MutableBuffer::new(bytes);
            [buffer, empty_buffer]
        }
        DataType::UInt8
        | DataType::UInt16
        | DataType::UInt32
        | DataType::UInt64
        | DataType::Int8
        | DataType::Int16
        | DataType::Int32
        | DataType::Int64
        | DataType::Float16
        | DataType::Float32
        | DataType::Float64
        | DataType::Decimal32(_, _)
        | DataType::Decimal64(_, _)
        | DataType::Decimal128(_, _)
        | DataType::Decimal256(_, _)
        | DataType::Date32
        | DataType::Time32(_)
        | DataType::Date64
        | DataType::Time64(_)
        | DataType::Duration(_)
        | DataType::Timestamp(_, _)
        | DataType::Interval(_) => [
            MutableBuffer::new(capacity * data_type.primitive_width().unwrap()),
            empty_buffer,
        ],
        DataType::Utf8 | DataType::Binary => {
            let mut buffer = MutableBuffer::new((1 + capacity) * mem::size_of::<i32>());
            // safety: `unsafe` code assumes that this buffer is initialized with one element
            buffer.push(0i32);
            [buffer, MutableBuffer::new(capacity * mem::size_of::<u8>())]
        }
        DataType::LargeUtf8 | DataType::LargeBinary => {
            let mut buffer = MutableBuffer::new((1 + capacity) * mem::size_of::<i64>());
            // safety: `unsafe` code assumes that this buffer is initialized with one element
            buffer.push(0i64);
            [buffer, MutableBuffer::new(capacity * mem::size_of::<u8>())]
        }
        DataType::BinaryView | DataType::Utf8View => [
            MutableBuffer::new(capacity * mem::size_of::<u128>()),
            empty_buffer,
        ],
        DataType::List(_) | DataType::Map(_, _) => {
            // offset buffer always starts with a zero
            let mut buffer = MutableBuffer::new((1 + capacity) * mem::size_of::<i32>());
            buffer.push(0i32);
            [buffer, empty_buffer]
        }
        DataType::ListView(_) => [
            MutableBuffer::new(capacity * mem::size_of::<i32>()),
            MutableBuffer::new(capacity * mem::size_of::<i32>()),
        ],
        DataType::LargeList(_) => {
            // offset buffer always starts with a zero
            let mut buffer = MutableBuffer::new((1 + capacity) * mem::size_of::<i64>());
            buffer.push(0i64);
            [buffer, empty_buffer]
        }
        DataType::LargeListView(_) => [
            MutableBuffer::new(capacity * mem::size_of::<i64>()),
            MutableBuffer::new(capacity * mem::size_of::<i64>()),
        ],
        DataType::FixedSizeBinary(size) => {
            if *size < 0 {
                panic!("cannot construct buffers from FixedSizeBinary({size})");
            }
            [MutableBuffer::new(capacity * *size as usize), empty_buffer]
        }
        DataType::Dictionary(k, _) => [
            MutableBuffer::new(capacity * k.primitive_width().unwrap()),
            empty_buffer,
        ],
        DataType::FixedSizeList(_, _) | DataType::Struct(_) | DataType::RunEndEncoded(_, _) => {
            [empty_buffer, MutableBuffer::new(0)]
        }
        DataType::Union(_, mode) => {
            let type_ids = MutableBuffer::new(capacity * mem::size_of::<i8>());
            match mode {
                UnionMode::Sparse => [type_ids, empty_buffer],
                UnionMode::Dense => {
                    let offsets = MutableBuffer::new(capacity * mem::size_of::<i32>());
                    [type_ids, offsets]
                }
            }
        }
    }
}

/// A generic representation of Arrow array data which encapsulates common attributes
/// and operations for Arrow array.
///
/// Specific operations for different arrays types (e.g., primitive, list, struct)
/// are implemented in `Array`.
///
/// # Memory Layout
///
/// `ArrayData` has references to one or more underlying data buffers
/// and optional child ArrayData, depending on type as illustrated
/// below. Bitmaps are not shown for simplicity but they are stored
/// similarly to the buffers.
///
/// ```text
///                        offset
///                       points to
/// ┌───────────────────┐ start of  ┌───────┐       Different
/// │                   │   data    │       │     ArrayData may
/// │ArrayData {        │           │....   │     also refers to
/// │  data_type: ...   │   ─ ─ ─ ─▶│1234   │  ┌ ─  the same
/// │  offset: ... ─ ─ ─│─ ┘        │4372   │      underlying
/// │  len: ...    ─ ─ ─│─ ┐        │4888   │  │     buffer with different offset/len
/// │  buffers: [       │           │5882   │◀─
/// │    ...            │  │        │4323   │
/// │  ]                │   ─ ─ ─ ─▶│4859   │
/// │  child_data: [    │           │....   │
/// │    ...            │           │       │
/// │  ]                │           └───────┘
/// │}                  │
/// │                   │            Shared Buffer uses
/// │               │   │            bytes::Bytes to hold
/// └───────────────────┘            actual data values
///           ┌ ─ ─ ┘
///
///           ▼
/// ┌───────────────────┐
/// │ArrayData {        │
/// │  ...              │
/// │}                  │
/// │                   │
/// └───────────────────┘
///
/// Child ArrayData may also have its own buffers and children
/// ```

#[derive(Debug, Clone)]
pub struct ArrayData {
    /// The data type
    data_type: DataType,

    /// The number of elements
    len: usize,

    /// The offset in number of items (not bytes).
    ///
    /// The offset applies to [`Self::child_data`] and [`Self::buffers`]. It
    /// does NOT apply to [`Self::nulls`].
    ///
    /// See [`Self::offset()`] for details and diagrams.
    offset: usize,

    /// The buffers that store the actual data for this array, as defined
    /// in the [Arrow Spec].
    ///
    /// Depending on the array types, [`Self::buffers`] can hold different
    /// kinds of buffers (e.g., value buffer, value offset buffer) at different
    /// positions.
    ///
    /// The buffer may be larger than needed.  Some items at the beginning may be skipped if
    /// there is an `offset`.  Some items at the end may be skipped if the buffer is longer than
    /// we need to satisfy `len`.
    ///
    /// [Arrow Spec](https://arrow.apache.org/docs/format/Columnar.html#physical-memory-layout)
    buffers: Vec<Buffer>,

    /// The child(ren) of this array.
    ///
    /// Only non-empty for nested types, such as `ListArray` and
    /// `StructArray`.
    ///
    /// The first logical element in each child element begins at `offset`.
    ///
    /// If the child element also has an offset then these offsets are
    /// cumulative.
    ///
    /// See [`Self::child_data()`] and [`Self::offset()`] for details.
    child_data: Vec<ArrayData>,

    /// The null bitmap.
    ///
    /// `None` indicates all values are non-null in this array.
    ///
    /// [`Self::offset()`] does not apply to the null bitmap. While the
    /// BooleanBuffer may be sliced (have its own offset) internally, this
    /// `NullBuffer` always represents exactly `len` elements.
    nulls: Option<NullBuffer>,
}

/// A thread-safe, shared reference to the Arrow array data.
pub type ArrayDataRef = Arc<ArrayData>;

fn checked_len_plus_offset(
    data_type: &DataType,
    len: usize,
    offset: usize,
) -> Result<usize, ArrowError> {
    len.checked_add(offset).ok_or_else(|| {
        ArrowError::InvalidArgumentError(format!(
            "Length {len} with offset {offset} overflows usize for {data_type}"
        ))
    })
}

/// Returns the first and last of `offsets`, and the bytes of `values` between
/// them as a `str`. Returns `None` if the offsets are not valid indexes into
/// `values` or those bytes are not valid UTF-8.
fn utf8_span<'a, T: ArrowNativeType>(
    offsets: &[T],
    values: &'a [u8],
) -> Option<(usize, usize, &'a str)> {
    let first = offsets.first()?.to_usize()?;
    let last = offsets.last()?.to_usize()?;
    let values_str = std::str::from_utf8(values.get(first..last)?).ok()?;
    Some((first, last, values_str))
}

impl ArrayData {
    /// Create a new ArrayData instance;
    ///
    /// If `null_count` is not specified, the number of nulls in
    /// null_bit_buffer is calculated.
    ///
    /// If the number of nulls is 0 then the null_bit_buffer
    /// is set to `None`.
    ///
    /// # Safety
    ///
    /// The input values *must* form a valid Arrow array for
    /// `data_type`, or undefined behavior can result.
    ///
    /// Note: This is a low level API and most users of the arrow
    /// crate should create arrays using the methods in the `array`
    /// module.
    pub unsafe fn new_unchecked(
        data_type: DataType,
        len: usize,
        null_count: Option<usize>,
        null_bit_buffer: Option<Buffer>,
        offset: usize,
        buffers: Vec<Buffer>,
        child_data: Vec<ArrayData>,
    ) -> Self {
        let builder = Self::inner_new_builder(
            data_type,
            len,
            null_count,
            null_bit_buffer,
            offset,
            buffers,
            child_data,
        );

        // SAFETY: caller responsible for ensuring data is valid
        unsafe { builder.build_unchecked() }
    }

    /// Create a new ArrayData, validating that the provided buffers form a valid
    /// Arrow array of the specified data type.
    ///
    /// If the number of nulls in `null_bit_buffer` is 0 then the null_bit_buffer
    /// is set to `None`.
    ///
    /// Internally this calls through to [`Self::validate_data`]
    ///
    /// Note: This is a low level API and most users of the arrow crate should create
    /// arrays using the builders found in [arrow_array](https://docs.rs/arrow-array)
    /// or [`ArrayDataBuilder`].
    ///
    /// See also [`Self::into_parts`] to recover the fields
    pub fn try_new(
        data_type: DataType,
        len: usize,
        null_bit_buffer: Option<Buffer>,
        offset: usize,
        buffers: Vec<Buffer>,
        child_data: Vec<ArrayData>,
    ) -> Result<Self, ArrowError> {
        let builder = Self::inner_new_builder(
            data_type,
            len,
            None,
            null_bit_buffer,
            offset,
            buffers,
            child_data,
        );

        assert!(!builder.skip_validation.get());

        // As the data is not trusted, do a full validation of its contents
        // We don't need to validate children as we can assume that the
        // [`ArrayData`] in `child_data` have already been validated through
        // a call to `ArrayData::try_new` or created using unsafe
        builder.build()
    }

    fn inner_new_builder(
        data_type: DataType,
        len: usize,
        null_count: Option<usize>,
        null_bit_buffer: Option<Buffer>,
        offset: usize,
        buffers: Vec<Buffer>,
        child_data: Vec<ArrayData>,
    ) -> ArrayDataBuilder {
        ArrayDataBuilder {
            data_type,
            len,
            null_count,
            null_bit_buffer,
            nulls: None,
            offset,
            buffers,
            child_data,
            align_buffers: false,
            skip_validation: UnsafeFlag::new(),
        }
    }

    /// Return the constituent parts of this ArrayData
    ///
    /// This is the inverse of [`ArrayData::try_new`].
    ///
    /// Returns `(data_type, len, nulls, offset, buffers, child_data)`
    pub fn into_parts(
        self,
    ) -> (
        DataType,
        usize,
        Option<NullBuffer>,
        usize,
        Vec<Buffer>,
        Vec<ArrayData>,
    ) {
        let Self {
            data_type,
            len,
            nulls,
            offset,
            buffers,
            child_data,
        } = self;

        (data_type, len, nulls, offset, buffers, child_data)
    }

    /// Returns a builder to construct a [`ArrayData`] instance of the same [`DataType`]
    #[inline]
    pub const fn builder(data_type: DataType) -> ArrayDataBuilder {
        ArrayDataBuilder::new(data_type)
    }

    /// Returns a reference to the [`DataType`] of this [`ArrayData`]
    #[inline]
    pub const fn data_type(&self) -> &DataType {
        &self.data_type
    }

    /// Returns the [`Buffer`] storing data for this [`ArrayData`]
    pub fn buffers(&self) -> &[Buffer] {
        &self.buffers
    }

    /// Returns a slice of children [`ArrayData`]. This will be non
    /// empty for type such as lists and structs.
    ///
    /// Note: For nested types where the parent element `i` corresponds directly
    /// to child element `i` (such as structs), both the parent's offset and
    /// each child's own offset apply when locating child values — see
    /// [`Self::offset`] for details.
    pub fn child_data(&self) -> &[ArrayData] {
        &self.child_data[..]
    }

    /// Returns whether the element at index `i` is null
    #[inline]
    pub fn is_null(&self, i: usize) -> bool {
        match &self.nulls {
            Some(v) => v.is_null(i),
            None => false,
        }
    }

    /// Returns a reference to the null buffer of this [`ArrayData`] if any
    ///
    /// Note: [`ArrayData::offset`] does NOT apply to the returned [`NullBuffer`]
    #[inline]
    pub fn nulls(&self) -> Option<&NullBuffer> {
        self.nulls.as_ref()
    }

    /// Returns whether the element at index `i` is not null
    #[inline]
    pub fn is_valid(&self, i: usize) -> bool {
        !self.is_null(i)
    }

    /// Returns the length (i.e., number of elements) of this [`ArrayData`].
    #[inline]
    pub const fn len(&self) -> usize {
        self.len
    }

    /// Returns whether this [`ArrayData`] is empty
    #[inline]
    pub const fn is_empty(&self) -> bool {
        self.len == 0
    }

    /// Returns the offset in elements of this [`ArrayData`]
    ///
    /// The offset applies to [`Self::buffers`] and [`Self::child_data`],
    /// but does NOT apply to [`Self::nulls`], which always represents exactly
    /// [`Self::len`] elements.
    ///
    /// # Offsets for Non-nested types
    ///
    /// For non-nested types, the offset skips leading elements in the buffers.
    /// Logical element `i` is stored at physical position `offset + i`.
    ///
    /// For example, with `offset = 2` and `len = 3` the following array
    /// represents elements `[C, D, E]`:
    ///
    /// ```text
    ///                     offset: 2        len: 3
    ///                   ◀───────────▶◀────────────────▶
    ///                   ┌─────┬─────┬─────┬─────┬─────┬─────┐
    ///    values buffer  │  A  │  B  │  C  │  D  │  E  │  F  │
    ///                   └─────┴─────┴─────┴─────┴─────┴─────┘
    ///    physical index    0     1     2     3     4     5
    ///    logical index                 0     1     2
    /// ```
    ///
    /// # Offsets for Struct types
    ///
    /// For [struct]s, logical element `i` of the parent corresponds directly to
    /// element `i` of each child, with no indirection in between. Since a
    /// struct has no buffers of its own, its offset applies to each child,
    /// composing cumulatively with any child offset. Logical element `i` of the
    /// struct corresponds to element `offset + i` of each child.
    ///
    /// For example, a struct with `offset = 2` and `len = 3` whose children
    /// `c1` and `c2` themselves each have an offset of `1` represents the
    /// elements `{c1: D, c2: d}`, `{c1: E, c2: e}`, `{c1: F, c2: f}`:
    ///
    /// ```text
    ///                        struct offset: 2    len: 3
    ///                          ◀───────────▶◀────────────────▶
    ///    child offset: 1 ◀─────▶
    ///                    ┌─────┬─────┬─────┬─────┬─────┬─────┬─────┐
    ///    child c1        │  A  │  B  │  C  │  D  │  E  │  F  │  G  │
    ///                    ├─────┼─────┼─────┼─────┼─────┼─────┼─────┤
    ///    child c2        │  a  │  b  │  c  │  d  │  e  │  f  │  g  │
    ///                    └─────┴─────┴─────┴─────┴─────┴─────┴─────┘
    ///    physical index     0     1     2     3     4     5     6
    ///    child index              0     1     2     3     4     5
    ///    struct index                         0     1     2
    /// ```
    ///
    /// [struct]: https://arrow.apache.org/docs/format/Columnar.html#struct-layout
    #[inline]
    pub const fn offset(&self) -> usize {
        self.offset
    }

    /// Returns the total number of nulls in this array
    #[inline]
    pub fn null_count(&self) -> usize {
        self.nulls
            .as_ref()
            .map(|x| x.null_count())
            .unwrap_or_default()
    }

    /// Returns the total number of bytes of memory occupied by the
    /// buffers owned by this [`ArrayData`] and all of its
    /// children. (See also diagram on [`ArrayData`]).
    ///
    /// Note that this [`ArrayData`] may only refer to a subset of the
    /// data in the underlying [`Buffer`]s (due to `offset` and
    /// `length`), but the size returned includes the entire size of
    /// the buffers.
    ///
    /// If multiple [`ArrayData`]s refer to the same underlying
    /// [`Buffer`]s they will both report the same size.
    pub fn get_buffer_memory_size(&self) -> usize {
        let mut size = 0;
        for buffer in &self.buffers {
            size += buffer.capacity();
        }
        if let Some(bitmap) = &self.nulls {
            size += bitmap.buffer().capacity()
        }
        for child in &self.child_data {
            size += child.get_buffer_memory_size();
        }
        size
    }

    /// Returns the total number of the bytes of memory occupied by
    /// the buffers by this slice of [`ArrayData`] (See also diagram on [`ArrayData`]).
    ///
    /// This is approximately the number of bytes if a new
    /// [`ArrayData`] was formed by creating new [`Buffer`]s with
    /// exactly the data needed. For variadic layouts, this includes the full
    /// capacity of every variadic buffer retained by a zero-copy slice, without
    /// inspecting which buffers or ranges are referenced by the slice.
    ///
    /// For example, a [`DataType::Int64`] with `100` elements,
    /// [`Self::get_slice_memory_size`] would return `100 * 8 = 800`. If
    /// the [`ArrayData`] was then [`Self::slice`]ed to refer to its
    /// first `20` elements, then [`Self::get_slice_memory_size`] on the
    /// sliced [`ArrayData`] would return `20 * 8 = 160`.
    pub fn get_slice_memory_size(&self) -> Result<usize, ArrowError> {
        let mut result: usize = 0;
        let layout = layout(&self.data_type);

        for spec in &layout.buffers {
            match spec {
                BufferSpec::FixedWidth { byte_width, .. } => {
                    // Offset buffers contain len+1 elements: one boundary per element
                    // plus a final boundary marking the end of the last element.
                    let len = match self.data_type {
                        DataType::Utf8
                        | DataType::LargeUtf8
                        | DataType::Binary
                        | DataType::LargeBinary
                        | DataType::List(_)
                        | DataType::LargeList(_)
                        | DataType::Map(_, _) => self.len + 1,
                        _ => self.len,
                    };
                    let buffer_size = len.checked_mul(*byte_width).ok_or_else(|| {
                        ArrowError::ComputeError(
                            "Integer overflow computing buffer size".to_string(),
                        )
                    })?;
                    result += buffer_size;
                }
                BufferSpec::VariableWidth => {
                    let buffer_len = match self.data_type {
                        DataType::Utf8 | DataType::Binary => {
                            let offsets = self.typed_offsets::<i32>()?;
                            (offsets[self.len] - offsets[0]) as usize
                        }
                        DataType::LargeUtf8 | DataType::LargeBinary => {
                            let offsets = self.typed_offsets::<i64>()?;
                            (offsets[self.len] - offsets[0]) as usize
                        }
                        _ => {
                            return Err(ArrowError::NotYetImplemented(format!(
                                "Invalid data type for VariableWidth buffer. Expected Utf8, LargeUtf8, Binary or LargeBinary. Got {}",
                                self.data_type
                            )));
                        }
                    };
                    result += buffer_len;
                }
                BufferSpec::BitMap => {
                    let buffer_size = bit_util::ceil(self.len, 8);
                    result += buffer_size;
                }
                BufferSpec::AlwaysNull => {
                    // Nothing to do
                }
            }
        }

        if layout.variadic {
            // Slicing view arrays retains all variadic data buffers unchanged.
            for buffer in self.buffers.iter().skip(layout.buffers.len()) {
                result += buffer.capacity();
            }
        }

        if self.nulls().is_some() {
            result += bit_util::ceil(self.len, 8);
        }

        for child in &self.child_data {
            result += child.get_slice_memory_size()?;
        }
        Ok(result)
    }

    /// Returns the total number of bytes of memory occupied
    /// physically by this [`ArrayData`] and all its [`Buffer`]s and
    /// children. (See also diagram on [`ArrayData`]).
    ///
    /// Equivalent to:
    ///  `size_of_val(self)` +
    ///  [`Self::get_buffer_memory_size`] +
    ///  `size_of_val(child)` for all children
    pub fn get_array_memory_size(&self) -> usize {
        let mut size = mem::size_of_val(self);

        // Calculate rest of the fields top down which contain actual data
        for buffer in &self.buffers {
            size += mem::size_of::<Buffer>();
            size += buffer.capacity();
        }
        if let Some(nulls) = &self.nulls {
            size += nulls.buffer().capacity();
        }
        for child in &self.child_data {
            size += child.get_array_memory_size();
        }

        size
    }

    /// Creates a zero-copy slice of itself. This creates a new
    /// [`ArrayData`] pointing at the same underlying [`Buffer`]s with a
    /// different offset and len
    ///
    /// # Panics
    ///
    /// Panics if `offset + length` overflows or is greater than `self.len()`.
    pub fn slice(&self, offset: usize, length: usize) -> ArrayData {
        self.clone().sliced(offset, length)
    }

    /// Creates a zero-copy slice of this array, reusing its buffer and child allocations.
    ///
    /// This is equivalent to [`Self::slice`], but consumes `self`.
    ///
    /// # Panics
    ///
    /// Panics if `offset + length` overflows or is greater than `self.len()`.
    pub fn sliced(mut self, offset: usize, length: usize) -> Self {
        self.slice_in_place(offset, length);
        self
    }

    fn slice_in_place(&mut self, offset: usize, length: usize) {
        let end = offset
            .checked_add(length)
            .expect("offset + length overflow");
        assert!(end <= self.len());

        if let DataType::Struct(_) = self.data_type() {
            // A struct has no buffers of its own, and reading child element `i`
            // combines this array's offset with the child's own offset. Applying
            // the slice to both would count it twice, so the cumulative offset
            // goes to the children and this array's offset is reset to 0.
            let child_offset = self.offset + offset;
            for child in &mut self.child_data {
                child.slice_in_place(child_offset, length);
            }
            self.offset = 0;
        } else {
            self.offset += offset;
        }
        self.len = length;
        // `nulls` belongs to this array rather than to the children, so
        // it is sliced by `offset` alone.
        self.nulls = self.nulls.as_ref().map(|x| x.slice(offset, length));
    }

    /// Returns the `buffer` as a slice of type `T` starting at self.offset
    ///
    /// # Panics
    /// This function panics if:
    /// * the buffer is not byte-aligned with type T, or
    /// * the datatype is `Boolean` (it corresponds to a bit-packed buffer where the offset is not applicable)
    pub fn buffer<T: ArrowNativeType>(&self, buffer: usize) -> &[T] {
        &self.buffers()[buffer].typed_data()[self.offset..]
    }

    /// Returns a new [`ArrayData`] valid for `data_type` containing `len` null values
    ///
    /// # Panics
    /// This function panics if:
    /// * the datatype `data_type` has incorrect layout
    pub fn new_null(data_type: &DataType, len: usize) -> Self {
        let bit_len = bit_util::ceil(len, 8);
        let zeroed = |len: usize| Buffer::from(MutableBuffer::from_len_zeroed(len));

        let (buffers, child_data, has_nulls) = match data_type.primitive_width() {
            Some(width) => (vec![zeroed(width * len)], vec![], true),
            None => match data_type {
                DataType::Null => (vec![], vec![], false),
                DataType::Boolean => (vec![zeroed(bit_len)], vec![], true),
                DataType::Binary | DataType::Utf8 => {
                    (vec![zeroed((len + 1) * 4), zeroed(0)], vec![], true)
                }
                DataType::BinaryView | DataType::Utf8View => (vec![zeroed(len * 16)], vec![], true),
                DataType::LargeBinary | DataType::LargeUtf8 => {
                    (vec![zeroed((len + 1) * 8), zeroed(0)], vec![], true)
                }
                DataType::FixedSizeBinary(i) => {
                    if *i < 0 {
                        panic!("cannot construct null data from FixedSizeBinary({i})");
                    }
                    (vec![zeroed(*i as usize * len)], vec![], true)
                }
                DataType::List(f) | DataType::Map(f, _) => (
                    vec![zeroed((len + 1) * 4)],
                    vec![ArrayData::new_empty(f.data_type())],
                    true,
                ),
                DataType::LargeList(f) => (
                    vec![zeroed((len + 1) * 8)],
                    vec![ArrayData::new_empty(f.data_type())],
                    true,
                ),
                DataType::ListView(f) => (
                    vec![zeroed(len * 4), zeroed(len * 4)],
                    vec![ArrayData::new_empty(f.data_type())],
                    true,
                ),
                DataType::LargeListView(f) => (
                    vec![zeroed(len * 8), zeroed(len * 8)],
                    vec![ArrayData::new_empty(f.data_type())],
                    true,
                ),
                DataType::FixedSizeList(f, list_len) => (
                    vec![],
                    vec![ArrayData::new_null(f.data_type(), *list_len as usize * len)],
                    true,
                ),
                DataType::Struct(fields) => (
                    vec![],
                    fields
                        .iter()
                        .map(|f| Self::new_null(f.data_type(), len))
                        .collect(),
                    true,
                ),
                DataType::Dictionary(k, v) => (
                    vec![zeroed(k.primitive_width().unwrap() * len)],
                    vec![ArrayData::new_empty(v.as_ref())],
                    true,
                ),
                DataType::Union(f, mode) => match f.iter().next() {
                    // Every slot carries a type id naming one of the children,
                    // so an empty union has nothing to put in a slot and can
                    // only be the empty array.
                    None => {
                        assert_eq!(
                            len, 0,
                            "cannot construct null data from an empty union of length {len}, a slot has no type id to carry"
                        );
                        let buffers = match mode {
                            UnionMode::Sparse => vec![zeroed(0)],
                            UnionMode::Dense => vec![zeroed(0), zeroed(0)],
                        };
                        (buffers, vec![], false)
                    }
                    Some((id, _)) => {
                        let ids = Buffer::from_iter(std::iter::repeat_n(id, len));
                        let buffers = match mode {
                            UnionMode::Sparse => vec![ids],
                            UnionMode::Dense => {
                                let end_offset = i32::from_usize(len).unwrap();
                                vec![ids, Buffer::from_iter(0_i32..end_offset)]
                            }
                        };

                        let children = f
                            .iter()
                            .enumerate()
                            .map(|(idx, (_, f))| {
                                if idx == 0 || *mode == UnionMode::Sparse {
                                    Self::new_null(f.data_type(), len)
                                } else {
                                    Self::new_empty(f.data_type())
                                }
                            })
                            .collect();

                        (buffers, children, false)
                    }
                },
                DataType::RunEndEncoded(r, v) => {
                    if len == 0 {
                        // For empty arrays, create zero-length child arrays.
                        let runs = ArrayData::new_empty(r.data_type());
                        let values = ArrayData::new_empty(v.data_type());
                        (vec![], vec![runs, values], false)
                    } else {
                        let runs = match r.data_type() {
                            DataType::Int16 => {
                                let i = i16::from_usize(len).expect("run overflow");
                                Buffer::from_slice_ref([i])
                            }
                            DataType::Int32 => {
                                let i = i32::from_usize(len).expect("run overflow");
                                Buffer::from_slice_ref([i])
                            }
                            DataType::Int64 => {
                                let i = i64::from_usize(len).expect("run overflow");
                                Buffer::from_slice_ref([i])
                            }
                            dt => unreachable!("Invalid run ends data type {dt}"),
                        };

                        let builder = ArrayData::builder(r.data_type().clone())
                            .len(1)
                            .buffers(vec![runs]);

                        // SAFETY:
                        // Valid by construction
                        let runs = unsafe { builder.build_unchecked() };
                        (
                            vec![],
                            vec![runs, ArrayData::new_null(v.data_type(), 1)],
                            false,
                        )
                    }
                }
                // Handled by Some(width) branch above
                DataType::Int8
                | DataType::Int16
                | DataType::Int32
                | DataType::Int64
                | DataType::UInt8
                | DataType::UInt16
                | DataType::UInt32
                | DataType::UInt64
                | DataType::Float16
                | DataType::Float32
                | DataType::Float64
                | DataType::Timestamp(_, _)
                | DataType::Date32
                | DataType::Date64
                | DataType::Time32(_)
                | DataType::Time64(_)
                | DataType::Duration(_)
                | DataType::Interval(_)
                | DataType::Decimal32(_, _)
                | DataType::Decimal64(_, _)
                | DataType::Decimal128(_, _)
                | DataType::Decimal256(_, _) => unreachable!("{data_type}"),
            },
        };

        let mut builder = ArrayDataBuilder::new(data_type.clone())
            .len(len)
            .buffers(buffers)
            .child_data(child_data);

        if has_nulls {
            builder = builder.nulls(Some(NullBuffer::new_null(len)))
        }

        // SAFETY:
        // Data valid by construction
        unsafe { builder.build_unchecked() }
    }

    /// Returns a new empty [ArrayData] valid for `data_type`.
    pub fn new_empty(data_type: &DataType) -> Self {
        Self::new_null(data_type, 0)
    }

    /// Verifies that the buffers meet the minimum alignment requirements for the data type
    ///
    /// Buffers that are not adequately aligned will be copied to a new aligned allocation
    ///
    /// This can be useful for when interacting with data sent over IPC or FFI, that may
    /// not meet the minimum alignment requirements
    ///
    /// This also aligns buffers of children data
    pub fn align_buffers(&mut self) {
        let layout = layout(&self.data_type);
        for (buffer, spec) in self.buffers.iter_mut().zip(&layout.buffers) {
            if let BufferSpec::FixedWidth { alignment, .. } = spec
                && buffer.as_ptr().align_offset(*alignment) != 0
            {
                *buffer = Buffer::from_slice_ref(buffer.as_ref());
            }
        }
        // align children data recursively
        for data in &mut self.child_data {
            data.align_buffers()
        }
    }

    /// "cheap" validation of an `ArrayData`. Ensures buffers are
    /// sufficiently sized to store `len` + `offset` total elements of
    /// `data_type` and performs other inexpensive consistency checks.
    ///
    /// This check is "cheap" in the sense that it does not validate the
    /// contents of the buffers (e.g. that all offsets for UTF8 arrays
    /// are within the bounds of the values buffer).
    ///
    /// See [ArrayData::validate_data] to validate fully the offset content
    /// and the validity of utf8 data
    pub fn validate(&self) -> Result<(), ArrowError> {
        // Need at least this much space in each buffer
        let len_plus_offset = checked_len_plus_offset(&self.data_type, self.len, self.offset)?;

        // Check that the data layout conforms to the spec
        let layout = layout(&self.data_type);

        if !layout.can_contain_null_mask && self.nulls.is_some() {
            return Err(ArrowError::InvalidArgumentError(format!(
                "Arrays of type {:?} cannot contain a null bitmask",
                self.data_type,
            )));
        }

        // Check data buffers length for view types and other types
        if self.buffers.len() < layout.buffers.len()
            || (!layout.variadic && self.buffers.len() != layout.buffers.len())
        {
            return Err(ArrowError::InvalidArgumentError(format!(
                "Expected {} buffers in array of type {:?}, got {}",
                layout.buffers.len(),
                self.data_type,
                self.buffers.len(),
            )));
        }

        for (i, (buffer, spec)) in self.buffers.iter().zip(layout.buffers.iter()).enumerate() {
            match spec {
                BufferSpec::FixedWidth {
                    byte_width,
                    alignment,
                } => {
                    let min_buffer_size = len_plus_offset.saturating_mul(*byte_width);

                    if buffer.len() < min_buffer_size {
                        return Err(ArrowError::InvalidArgumentError(format!(
                            "Need at least {} bytes in buffers[{}] in array of type {:?}, but got {}",
                            min_buffer_size,
                            i,
                            self.data_type,
                            buffer.len()
                        )));
                    }

                    let align_offset = buffer.as_ptr().align_offset(*alignment);
                    if align_offset != 0 {
                        return Err(ArrowError::InvalidArgumentError(format!(
                            "Misaligned buffers[{i}] in array of type {:?}, offset from expected alignment of {alignment} by {}",
                            self.data_type,
                            align_offset.min(alignment - align_offset)
                        )));
                    }
                }
                BufferSpec::VariableWidth => {
                    // not cheap to validate (need to look at the
                    // data). Partially checked in validate_offsets
                    // called below. Can check with `validate_full`
                }
                BufferSpec::BitMap => {
                    let min_buffer_size = bit_util::ceil(len_plus_offset, 8);
                    if buffer.len() < min_buffer_size {
                        return Err(ArrowError::InvalidArgumentError(format!(
                            "Need at least {} bytes for bitmap in buffers[{}] in array of type {:?}, but got {}",
                            min_buffer_size,
                            i,
                            self.data_type,
                            buffer.len()
                        )));
                    }
                }
                BufferSpec::AlwaysNull => {
                    // Nothing to validate
                }
            }
        }

        // check null bit buffer size
        if let Some(nulls) = self.nulls() {
            if nulls.null_count() > self.len {
                return Err(ArrowError::InvalidArgumentError(format!(
                    "null_count {} for an array exceeds length of {} elements",
                    nulls.null_count(),
                    self.len
                )));
            }

            if nulls.len() != self.len {
                return Err(ArrowError::InvalidArgumentError(format!(
                    "null buffer incorrect size. got {} expected {}",
                    nulls.len(),
                    self.len
                )));
            }
        }

        self.validate_child_data()?;

        // Additional Type specific checks
        match &self.data_type {
            DataType::Utf8 | DataType::Binary => {
                self.validate_offsets::<i32>(self.buffers[1].len())?;
            }
            DataType::LargeUtf8 | DataType::LargeBinary => {
                self.validate_offsets::<i64>(self.buffers[1].len())?;
            }
            DataType::Dictionary(key_type, _value_type) => {
                // At the moment, constructing a DictionaryArray will also check this
                if !DataType::is_dictionary_key_type(key_type) {
                    return Err(ArrowError::InvalidArgumentError(format!(
                        "Dictionary key type must be integer, but was {key_type}"
                    )));
                }
            }
            DataType::RunEndEncoded(run_ends_type, _) => {
                if run_ends_type.is_nullable() {
                    return Err(ArrowError::InvalidArgumentError(
                        "The nullable should be set to false for the field defining run_ends array.".to_string()
                    ));
                }
                if !DataType::is_run_ends_type(run_ends_type.data_type()) {
                    return Err(ArrowError::InvalidArgumentError(format!(
                        "RunArray run_ends types must be Int16, Int32 or Int64, but was {}",
                        run_ends_type.data_type()
                    )));
                }
            }
            DataType::Map(f, _) if f.is_nullable() => {
                return Err(ArrowError::InvalidArgumentError(
                    "The nullable should be set to false for the map entries field.".to_string(),
                ));
            }
            _ => {}
        }

        Ok(())
    }

    /// Returns a reference to the data in `buffer` as a typed slice
    /// (typically `&[i32]` or `&[i64]`) after validating. The
    /// returned slice is guaranteed to have at least `self.len + 1`
    /// entries.
    ///
    /// For an empty array, the `buffer` can also be empty.
    fn typed_offsets<T: ArrowNativeType + num_traits::Num>(&self) -> Result<&[T], ArrowError> {
        // An empty list-like array can have 0 offsets
        if self.len == 0 && self.buffer_at(0)?.is_empty() {
            return Ok(&[]);
        }

        let len = checked_len_plus_offset(&self.data_type, self.len, 1)?;

        self.typed_buffer(0, len)
    }

    /// Returns a reference to the data in `buffers[idx]` as a typed slice after validating
    fn typed_buffer<T: ArrowNativeType + num_traits::Num>(
        &self,
        idx: usize,
        len: usize,
    ) -> Result<&[T], ArrowError> {
        let buffer = self.buffer_at(idx)?;

        let required_elements = checked_len_plus_offset(&self.data_type, len, self.offset)?;
        let byte_width = mem::size_of::<T>();
        let required_len = required_elements.checked_mul(byte_width).ok_or_else(|| {
            ArrowError::InvalidArgumentError(format!(
                "Buffer {idx} of {} byte length overflow: {} elements of {} bytes exceeds usize",
                self.data_type, required_elements, byte_width
            ))
        })?;

        if buffer.len() < required_len {
            return Err(ArrowError::InvalidArgumentError(format!(
                "Buffer {} of {} isn't large enough. Expected {} bytes got {}",
                idx,
                self.data_type,
                required_len,
                buffer.len()
            )));
        }

        Ok(&buffer.typed_data::<T>()[self.offset..required_elements])
    }

    /// Does a cheap sanity check that the `self.len` values in `buffer` are valid
    /// offsets (of type T) into some other buffer of `values_length` bytes long
    fn validate_offsets<T: ArrowNativeType + num_traits::Num + std::fmt::Display>(
        &self,
        values_length: usize,
    ) -> Result<(), ArrowError> {
        // Justification: buffer size was validated above
        let offsets = self.typed_offsets::<T>()?;
        if offsets.is_empty() {
            return Ok(());
        }

        let first_offset = offsets[0].to_usize().ok_or_else(|| {
            ArrowError::InvalidArgumentError(format!(
                "Error converting offset[0] ({}) to usize for {}",
                offsets[0], self.data_type
            ))
        })?;

        let last_offset = offsets[self.len].to_usize().ok_or_else(|| {
            ArrowError::InvalidArgumentError(format!(
                "Error converting offset[{}] ({}) to usize for {}",
                self.len, offsets[self.len], self.data_type
            ))
        })?;

        if first_offset > values_length {
            return Err(ArrowError::InvalidArgumentError(format!(
                "First offset {} of {} is larger than values length {}",
                first_offset, self.data_type, values_length,
            )));
        }

        if last_offset > values_length {
            return Err(ArrowError::InvalidArgumentError(format!(
                "Last offset {} of {} is larger than values length {}",
                last_offset, self.data_type, values_length,
            )));
        }

        if first_offset > last_offset {
            return Err(ArrowError::InvalidArgumentError(format!(
                "First offset {} in {} is smaller than last offset {}",
                first_offset, self.data_type, last_offset,
            )));
        }

        Ok(())
    }

    /// Does a cheap sanity check that the `self.len` values in `buffer` are valid
    /// offsets and sizes (of type T) into some other buffer of `values_length` bytes long
    fn validate_offsets_and_sizes<T: ArrowNativeType + num_traits::Num + std::fmt::Display>(
        &self,
        values_length: usize,
    ) -> Result<(), ArrowError> {
        let offsets: &[T] = self.typed_buffer(0, self.len)?;
        let sizes: &[T] = self.typed_buffer(1, self.len)?;
        if offsets.len() != sizes.len() {
            return Err(ArrowError::ComputeError(format!(
                "ListView offsets len {} does not match sizes len {}",
                offsets.len(),
                sizes.len()
            )));
        }

        for i in 0..sizes.len() {
            let size = sizes[i].to_usize().ok_or_else(|| {
                ArrowError::InvalidArgumentError(format!(
                    "Error converting size[{}] ({}) to usize for {}",
                    i, sizes[i], self.data_type
                ))
            })?;
            let offset = offsets[i].to_usize().ok_or_else(|| {
                ArrowError::InvalidArgumentError(format!(
                    "Error converting offset[{}] ({}) to usize for {}",
                    i, offsets[i], self.data_type
                ))
            })?;
            if size
                .checked_add(offset)
                .expect("Offset and size have exceeded the usize boundary")
                > values_length
            {
                return Err(ArrowError::InvalidArgumentError(format!(
                    "Size {} at index {} is larger than the remaining values for {}",
                    size, i, self.data_type
                )));
            }
        }
        Ok(())
    }

    /// Validates the layout of `child_data` ArrayData structures
    fn validate_child_data(&self) -> Result<(), ArrowError> {
        match &self.data_type {
            DataType::List(field) => {
                let values_data = self.get_single_valid_child_data(field.data_type())?;
                self.validate_offsets::<i32>(values_data.len)?;
            }
            DataType::LargeList(field) => {
                let values_data = self.get_single_valid_child_data(field.data_type())?;
                self.validate_offsets::<i64>(values_data.len)?;
            }
            DataType::Map(field, _) => {
                let DataType::Struct(entries_fields) = field.data_type() else {
                    return Err(ArrowError::InvalidArgumentError(format!(
                        "Map field should be a entries struct data type, got {:?} instead",
                        field.data_type()
                    )));
                };
                if entries_fields.len() != 2 {
                    return Err(ArrowError::InvalidArgumentError(format!(
                        "Map entries data type should be a struct containing 2 fields, got {} fields",
                        entries_fields.len()
                    )));
                }

                // Key field
                if entries_fields[0].is_nullable() {
                    return Err(ArrowError::InvalidArgumentError(
                        "Map key field must not be nullable".to_string(),
                    ));
                }
                let values_data = self.get_single_valid_child_data(field.data_type())?;
                self.validate_offsets::<i32>(values_data.len)?;
            }
            DataType::ListView(field) => {
                let values_data = self.get_single_valid_child_data(field.data_type())?;
                self.validate_offsets_and_sizes::<i32>(values_data.len)?;
            }
            DataType::LargeListView(field) => {
                let values_data = self.get_single_valid_child_data(field.data_type())?;
                self.validate_offsets_and_sizes::<i64>(values_data.len)?;
            }
            DataType::FixedSizeList(field, list_size) => {
                let values_data = self.get_single_valid_child_data(field.data_type())?;

                let list_size: usize = (*list_size).try_into().map_err(|_| {
                    ArrowError::InvalidArgumentError(format!(
                        "{} has a negative list_size {}",
                        self.data_type, list_size
                    ))
                })?;

                let len_plus_offset =
                    checked_len_plus_offset(&self.data_type, self.len, self.offset)?;
                let expected_values_len = len_plus_offset
                    .checked_mul(list_size)
                    .expect("integer overflow computing expected number of expected values in FixedListSize");

                if values_data.len < expected_values_len {
                    return Err(ArrowError::InvalidArgumentError(format!(
                        "Values length {} is less than the length + offset ({}) multiplied by the value size ({}) for {}",
                        values_data.len, len_plus_offset, list_size, self.data_type
                    )));
                }
            }
            DataType::Struct(fields) => {
                self.validate_num_child_data(fields.len())?;
                let len_plus_offset =
                    checked_len_plus_offset(&self.data_type, self.len, self.offset)?;
                for (i, field) in fields.iter().enumerate() {
                    let field_data = self.get_valid_child_data(i, field.data_type())?;

                    // Ensure child field has sufficient size
                    if field_data.len < len_plus_offset {
                        return Err(ArrowError::InvalidArgumentError(format!(
                            "{} child array #{} for field {} has length smaller than expected for struct array ({} < {})",
                            self.data_type,
                            i,
                            field.name(),
                            field_data.len,
                            len_plus_offset
                        )));
                    }
                }
            }
            DataType::RunEndEncoded(run_ends_field, values_field) => {
                self.validate_num_child_data(2)?;
                let run_ends_data = self.get_valid_child_data(0, run_ends_field.data_type())?;
                let values_data = self.get_valid_child_data(1, values_field.data_type())?;
                if run_ends_data.len != values_data.len {
                    return Err(ArrowError::InvalidArgumentError(format!(
                        "The run_ends array length should be the same as values array length. Run_ends array length is {}, values array length is {}",
                        run_ends_data.len, values_data.len
                    )));
                }
                if run_ends_data.nulls.is_some() {
                    return Err(ArrowError::InvalidArgumentError(
                        "Found null values in run_ends array. The run_ends array should not have null values.".to_string(),
                    ));
                }
            }
            DataType::Union(fields, mode) => {
                self.validate_num_child_data(fields.len())?;

                for (i, (_, field)) in fields.iter().enumerate() {
                    let field_data = self.get_valid_child_data(i, field.data_type())?;

                    if mode == &UnionMode::Sparse {
                        let len_plus_offset =
                            checked_len_plus_offset(&self.data_type, self.len, self.offset)?;
                        if field_data.len < len_plus_offset {
                            return Err(ArrowError::InvalidArgumentError(format!(
                                "Sparse union child array #{} has length smaller than expected for union array ({} < {})",
                                i, field_data.len, len_plus_offset
                            )));
                        }
                    }
                }
            }
            DataType::Dictionary(_key_type, value_type) => {
                self.get_single_valid_child_data(value_type)?;
            }
            _ => {
                // other types do not have child data
                if !self.child_data.is_empty() {
                    return Err(ArrowError::InvalidArgumentError(format!(
                        "Expected no child arrays for type {} but got {}",
                        self.data_type,
                        self.child_data.len()
                    )));
                }
            }
        }
        Ok(())
    }

    /// Ensures that this array data has a single child_data with the
    /// expected type, and calls `validate()` on it. Returns a
    /// reference to that child_data
    fn get_single_valid_child_data(
        &self,
        expected_type: &DataType,
    ) -> Result<&ArrayData, ArrowError> {
        self.validate_num_child_data(1)?;
        self.get_valid_child_data(0, expected_type)
    }

    /// Returns `buffers[idx]`, or an error if there is no such buffer.
    ///
    /// [`Self::validate_values`] can be called on its own, without the buffer counts
    /// having been checked by [`Self::validate`] first, so the index may be missing.
    fn buffer_at(&self, idx: usize) -> Result<&Buffer, ArrowError> {
        self.buffers.get(idx).ok_or_else(|| {
            ArrowError::InvalidArgumentError(format!(
                "{} should contain at least {} buffer(s), had {}",
                self.data_type,
                idx + 1,
                self.buffers.len()
            ))
        })
    }

    /// Returns `child_data[idx]`, or an error if there is no such child.
    ///
    /// [`Self::validate_values`] can be called on its own, without the child counts
    /// having been checked by [`Self::validate`] first, so the index may be missing.
    fn child_at(&self, idx: usize) -> Result<&ArrayData, ArrowError> {
        self.child_data.get(idx).ok_or_else(|| {
            ArrowError::InvalidArgumentError(format!(
                "{} should contain at least {} child data array(s), had {}",
                self.data_type,
                idx + 1,
                self.child_data.len()
            ))
        })
    }

    /// Returns `Err` if self.child_data does not have exactly `expected_len` elements
    fn validate_num_child_data(&self, expected_len: usize) -> Result<(), ArrowError> {
        if self.child_data.len() != expected_len {
            Err(ArrowError::InvalidArgumentError(format!(
                "Value data for {} should contain {} child data array(s), had {}",
                self.data_type,
                expected_len,
                self.child_data.len()
            )))
        } else {
            Ok(())
        }
    }

    /// Ensures that `child_data[i]` has the expected type, calls
    /// `validate()` on it, and returns a reference to that child_data
    fn get_valid_child_data(
        &self,
        i: usize,
        expected_type: &DataType,
    ) -> Result<&ArrayData, ArrowError> {
        let values_data = self.child_data.get(i).ok_or_else(|| {
            ArrowError::InvalidArgumentError(format!(
                "{} did not have enough child arrays. Expected at least {} but had only {}",
                self.data_type,
                i + 1,
                self.child_data.len()
            ))
        })?;

        if expected_type != &values_data.data_type {
            return Err(ArrowError::InvalidArgumentError(format!(
                "Child type mismatch for {}. Expected {} but child data had {}",
                self.data_type, expected_type, values_data.data_type
            )));
        }

        values_data.validate()?;
        Ok(values_data)
    }

    /// Validate that the data contained within this [`ArrayData`] is valid
    ///
    /// 1. Null count is correct
    /// 2. All offsets are valid
    /// 3. All String data is valid UTF-8
    /// 4. All dictionary offsets are valid
    ///
    /// Internally this calls:
    ///
    /// * [`Self::validate`]
    /// * [`Self::validate_nulls`]
    /// * [`Self::validate_values`]
    ///
    /// Note: this does not recurse into children, for a recursive variant
    /// see [`Self::validate_full`]
    pub fn validate_data(&self) -> Result<(), ArrowError> {
        self.validate()?;

        self.validate_nulls()?;
        self.validate_values()?;
        Ok(())
    }

    /// Performs a full recursive validation of this [`ArrayData`] and all its children
    ///
    /// This is equivalent to calling [`Self::validate_data`] on this [`ArrayData`]
    /// and all its children recursively
    pub fn validate_full(&self) -> Result<(), ArrowError> {
        self.validate_data()?;
        // validate all children recursively
        self.child_data
            .iter()
            .enumerate()
            .try_for_each(|(i, child_data)| {
                child_data.validate_full().map_err(|e| {
                    ArrowError::InvalidArgumentError(format!(
                        "{} child #{} invalid: {}",
                        self.data_type, i, e
                    ))
                })
            })?;
        Ok(())
    }

    /// Validates the values stored within this [`ArrayData`] are valid
    /// without recursing into child [`ArrayData`]
    ///
    /// Does not (yet) check
    /// 1. Union type_ids are valid see [#85](https://github.com/apache/arrow-rs/issues/85)
    /// 2. the null count is correct and that any
    /// 3. nullability requirements of its children are correct
    ///
    /// [#85]: https://github.com/apache/arrow-rs/issues/85
    pub fn validate_nulls(&self) -> Result<(), ArrowError> {
        if let Some(nulls) = &self.nulls {
            let actual = nulls.len() - nulls.inner().count_set_bits();
            if actual != nulls.null_count() {
                return Err(ArrowError::InvalidArgumentError(format!(
                    "null_count value ({}) doesn't match actual number of nulls in array ({})",
                    nulls.null_count(),
                    actual
                )));
            }
        }

        // In general non-nullable children should not contain nulls, however, for certain
        // types, such as StructArray and FixedSizeList, nulls in the parent take up
        // space in the child. As such we permit nulls in the children in the corresponding
        // positions for such types
        match &self.data_type {
            DataType::List(f) | DataType::LargeList(f) | DataType::Map(f, _) => {
                if !f.is_nullable() {
                    let child = &self.child_data[0];
                    self.validate_non_nullable(None, child, child.nulls())?
                }
            }
            DataType::FixedSizeList(field, len) => {
                let child = &self.child_data[0];
                if !field.is_nullable() {
                    let element_len = *len as usize;
                    let child_nulls = child.nulls().map(|nulls| {
                        nulls.slice(self.offset * element_len, self.len * element_len)
                    });
                    let expanded = self.nulls.as_ref().map(|nulls| nulls.expand(element_len));
                    self.validate_non_nullable(expanded.as_ref(), child, child_nulls.as_ref())?;
                }
            }
            DataType::Struct(fields) => {
                for (field, child) in fields.iter().zip(&self.child_data) {
                    if !field.is_nullable() {
                        let child_nulls = child
                            .nulls()
                            .map(|nulls| nulls.slice(self.offset, self.len));
                        self.validate_non_nullable(self.nulls(), child, child_nulls.as_ref())?
                    }
                }
            }
            _ => {}
        }

        Ok(())
    }

    /// Verifies that `child` contains no nulls not present in `mask`
    fn validate_non_nullable(
        &self,
        mask: Option<&NullBuffer>,
        child: &ArrayData,
        child_nulls: Option<&NullBuffer>,
    ) -> Result<(), ArrowError> {
        let Some(mask) = mask else {
            return match child_nulls.map(NullBuffer::null_count).unwrap_or_default() {
                0 => Ok(()),
                _ => Err(ArrowError::InvalidArgumentError(format!(
                    "non-nullable child of type {} contains nulls not present in parent {}",
                    child.data_type, self.data_type
                ))),
            };
        };

        match child_nulls {
            Some(nulls) if !mask.contains(nulls) => Err(ArrowError::InvalidArgumentError(format!(
                "non-nullable child of type {} contains nulls not present in parent",
                child.data_type
            ))),
            _ => Ok(()),
        }
    }

    /// Validates the values stored within this [`ArrayData`] are valid
    /// without recursing into child [`ArrayData`]
    ///
    /// Does not (yet) check
    /// 1. Union type_ids are valid see [#85](https://github.com/apache/arrow-rs/issues/85)
    pub fn validate_values(&self) -> Result<(), ArrowError> {
        match &self.data_type {
            DataType::Utf8 => self.validate_utf8::<i32>(),
            DataType::LargeUtf8 => self.validate_utf8::<i64>(),
            DataType::Binary => self.validate_offsets_full::<i32>(self.buffer_at(1)?.len()),
            DataType::LargeBinary => self.validate_offsets_full::<i64>(self.buffer_at(1)?.len()),
            DataType::BinaryView => {
                let views = self.typed_buffer::<u128>(0, self.len)?;
                validate_binary_view(views, &self.buffers[1..])
            }
            DataType::Utf8View => {
                let views = self.typed_buffer::<u128>(0, self.len)?;
                validate_string_view(views, &self.buffers[1..])
            }
            DataType::List(_) | DataType::Map(_, _) => {
                let child = self.child_at(0)?;
                self.validate_offsets_full::<i32>(child.len)
            }
            DataType::LargeList(_) => {
                let child = self.child_at(0)?;
                self.validate_offsets_full::<i64>(child.len)
            }
            DataType::Union(fields, mode) => {
                let mut child_lens: Vec<Option<usize>> = vec![None; 128];
                for (index, (type_id, _)) in fields.iter().enumerate() {
                    let child = self.child_at(index)?;
                    child_lens[type_id as usize] = Some(child.len());
                }

                // Type id values must match one of the fields.
                let type_ids = self.typed_buffer::<i8>(0, self.len)?;
                for (index, &type_id) in type_ids.iter().enumerate() {
                    if type_id < 0 {
                        return Err(ArrowError::InvalidArgumentError(format!(
                            "Type id values cannot be negative, got {type_id} at index {index}"
                        )));
                    }
                    let type_id = type_id as usize;
                    if child_lens[type_id].is_none() {
                        return Err(ArrowError::InvalidArgumentError(format!(
                            "Type id values must match one of the field type ids, got {type_id} at index {index}"
                        )));
                    }
                }

                // Check the value offsets are in bounds.
                if mode == &UnionMode::Dense {
                    let offsets = self.typed_buffer::<i32>(1, self.len)?;

                    if type_ids
                        .iter()
                        .zip(offsets.iter())
                        .any(|(&type_id, &offset)| {
                            offset < 0 || offset as usize >= child_lens[type_id as usize].unwrap()
                        })
                    {
                        return Err(ArrowError::InvalidArgumentError(
                            "Offsets must be non-negative and within the length of the child array"
                                .to_string(),
                        ));
                    }
                }

                Ok(())
            }
            DataType::Dictionary(key_type, _value_type) => {
                let dictionary_length = self.child_at(0)?.len;
                let dictionary_length = i64::try_from(dictionary_length).map_err(|_| {
                    ArrowError::InvalidArgumentError(format!(
                        "Dictionary of {dictionary_length} values is too long for an i64"
                    ))
                })?;
                let max_value = dictionary_length - 1;
                match key_type.as_ref() {
                    DataType::UInt8 => self.check_bounds::<u8>(max_value),
                    DataType::UInt16 => self.check_bounds::<u16>(max_value),
                    DataType::UInt32 => self.check_bounds::<u32>(max_value),
                    DataType::UInt64 => self.check_bounds::<u64>(max_value),
                    DataType::Int8 => self.check_bounds::<i8>(max_value),
                    DataType::Int16 => self.check_bounds::<i16>(max_value),
                    DataType::Int32 => self.check_bounds::<i32>(max_value),
                    DataType::Int64 => self.check_bounds::<i64>(max_value),
                    _ => Err(ArrowError::InvalidArgumentError(format!(
                        "Dictionary key type must be an integer, got {key_type}"
                    ))),
                }
            }
            DataType::RunEndEncoded(run_ends, _values) => {
                let len_plus_offset =
                    checked_len_plus_offset(&self.data_type, self.len, self.offset)?;
                let run_ends_data = self.child_at(0)?;
                match run_ends.data_type() {
                    DataType::Int16 => {
                        run_ends_data.check_run_ends::<i16>(self.len, len_plus_offset)
                    }
                    DataType::Int32 => {
                        run_ends_data.check_run_ends::<i32>(self.len, len_plus_offset)
                    }
                    DataType::Int64 => {
                        run_ends_data.check_run_ends::<i64>(self.len, len_plus_offset)
                    }
                    data_type => Err(ArrowError::InvalidArgumentError(format!(
                        "Run end type must be Int16, Int32 or Int64, got {data_type}"
                    ))),
                }
            }
            _ => {
                // No extra validation check required for other types
                Ok(())
            }
        }
    }

    /// Calls the `validate(item_index, range)` function for each of
    /// the ranges specified in the arrow offsets buffer of type
    /// `T`. Also validates that each offset is smaller than
    /// `offset_limit`
    ///
    /// For an empty array, the offsets buffer can either be empty
    /// or contain a single `0`.
    ///
    /// For example, the offsets buffer contained `[1, 2, 4]`, this
    /// function would call `validate([1,2])`, and `validate([2,4])`
    fn validate_each_offset<T, V>(&self, offset_limit: usize, validate: V) -> Result<(), ArrowError>
    where
        T: ArrowNativeType + TryInto<usize> + num_traits::Num + std::fmt::Display,
        V: Fn(usize, Range<usize>) -> Result<(), ArrowError>,
    {
        self.typed_offsets::<T>()?
            .iter()
            .enumerate()
            .map(|(i, x)| {
                // check if the offset can be converted to usize
                let r = x.to_usize().ok_or_else(|| {
                    ArrowError::InvalidArgumentError(format!(
                        "Offset invariant failure: Could not convert offset {x} to usize at position {i}"))}
                    );
                // check if the offset exceeds the limit
                match r {
                    Ok(n) if n <= offset_limit => Ok((i, n)),
                    Ok(_) => Err(ArrowError::InvalidArgumentError(format!(
                        "Offset invariant failure: offset at position {i} out of bounds: {x} > {offset_limit}"))
                    ),
                    Err(e) => Err(e),
                }
            })
            .scan(0_usize, |start, end| {
                // check offsets are monotonically increasing
                match end {
                    Ok((i, end)) if *start <= end => {
                        let range = Some(Ok((i, *start..end)));
                        *start = end;
                        range
                    }
                    Ok((i, end)) => Some(Err(ArrowError::InvalidArgumentError(format!(
                        "Offset invariant failure: non-monotonic offset at slot {}: {} > {}",
                        i - 1, start, end))
                    )),
                    Err(err) => Some(Err(err)),
                }
            })
            .skip(1) // the first element is meaningless
            .try_for_each(|res: Result<(usize, Range<usize>), ArrowError>| {
                let (item_index, range) = res?;
                validate(item_index-1, range)
            })
    }

    /// Ensures that all strings formed by the offsets in `buffers[0]`
    /// into `buffers[1]` are valid utf8 sequences. Bytes of `buffers[1]`
    /// outside the range of the offsets are not checked.
    fn validate_utf8<T>(&self) -> Result<(), ArrowError>
    where
        T: ArrowNativeType + TryInto<usize> + num_traits::Num + std::fmt::Display,
    {
        let values_buffer = &self.buffer_at(1)?.as_slice();
        let offsets = self.typed_offsets::<T>()?;
        if let Some((first, last, values_str)) = utf8_span(offsets, values_buffer) {
            self.validate_each_offset::<T, _>(values_buffer.len(), |string_index, range| {
                // `values_str` ends at the last offset. Checking each pair of offsets doesn't
                // show that `range.end <= last`: a later offset can still be smaller, which
                // `validate_each_offset` reports when it gets there.
                if range.end > last {
                    return Ok(());
                }
                let (start, end) = (range.start - first, range.end - first);
                if !values_str.is_char_boundary(start) || !values_str.is_char_boundary(end) {
                    return Err(ArrowError::InvalidArgumentError(format!(
                        "incomplete utf-8 byte sequence from index {string_index}"
                    )));
                }
                Ok(())
            })
        } else {
            // Find specific offset that failed utf8 validation
            self.validate_each_offset::<T, _>(values_buffer.len(), |string_index, range| {
                std::str::from_utf8(&values_buffer[range.clone()]).map_err(|e| {
                    ArrowError::InvalidArgumentError(format!(
                        "Invalid UTF8 sequence at string index {string_index} ({range:?}): {e}"
                    ))
                })?;
                Ok(())
            })
        }
    }

    /// Ensures that all offsets in `buffers[0]` into `buffers[1]` are
    /// between `0` and `offset_limit`
    fn validate_offsets_full<T>(&self, offset_limit: usize) -> Result<(), ArrowError>
    where
        T: ArrowNativeType + TryInto<usize> + num_traits::Num + std::fmt::Display,
    {
        self.validate_each_offset::<T, _>(offset_limit, |_string_index, _range| {
            // No validation applied to each value, but the iteration
            // itself applies bounds checking to each range
            Ok(())
        })
    }

    /// Validates that each value in self.buffers (typed as T)
    /// is within the range [0, max_value], inclusive
    fn check_bounds<T>(&self, max_value: i64) -> Result<(), ArrowError>
    where
        T: ArrowNativeType + TryInto<i64> + num_traits::Num + std::fmt::Display,
    {
        // `validate()` checks the buffer size too, but `validate_values()` can be called
        // on its own, so do not assume it has run.
        let indexes: &[T] = self.typed_buffer::<T>(0, self.len)?;

        indexes.iter().enumerate().try_for_each(|(i, &dict_index)| {
            // Do not check the value is null (value can be arbitrary)
            if self.is_null(i) {
                return Ok(());
            }
            let dict_index: i64 = dict_index.try_into().map_err(|_| {
                ArrowError::InvalidArgumentError(format!(
                    "Value at position {i} out of bounds: {dict_index} (can not convert to i64)"
                ))
            })?;

            if dict_index < 0 || dict_index > max_value {
                return Err(ArrowError::InvalidArgumentError(format!(
                    "Value at position {i} out of bounds: {dict_index} (should be in [0, {max_value}])"
                )));
            }
            Ok(())
        })
    }

    /// Validates that each value in run_ends array is positive and strictly increasing.
    fn check_run_ends<T>(
        &self,
        parent_len: usize,
        parent_len_plus_offset: usize,
    ) -> Result<(), ArrowError>
    where
        T: ArrowNativeType + TryInto<i64> + num_traits::Num + std::fmt::Display,
    {
        let values = self.typed_buffer::<T>(0, self.len)?;
        let mut prev_value = 0_i64;
        values.iter().enumerate().try_for_each(|(ix, &inp_value)| {
            let value: i64 = inp_value.try_into().map_err(|_| {
                ArrowError::InvalidArgumentError(format!(
                    "Value at position {ix} out of bounds: {inp_value} (can not convert to i64)"
                ))
            })?;
            if value <= 0_i64 {
                return Err(ArrowError::InvalidArgumentError(format!(
                    "The values in run_ends array should be strictly positive. Found value {value} at index {ix} that does not match the criteria."
                )));
            }
            if ix > 0 && value <= prev_value {
                return Err(ArrowError::InvalidArgumentError(format!(
                    "The values in run_ends array should be strictly increasing. Found value {value} at index {ix} with previous value {prev_value} that does not match the criteria."
                )));
            }

            prev_value = value;
            Ok(())
        })?;

        if parent_len > 0 && prev_value.as_usize() < parent_len_plus_offset {
            return Err(ArrowError::InvalidArgumentError(format!(
                "The offset + length of array should be less or equal to last value in the run_ends array. The last value of run_ends array is {prev_value} and offset + length of array is {parent_len_plus_offset}."
            )));
        }
        Ok(())
    }

    /// Returns true if this `ArrayData` is equal to `other`, using pointer comparisons
    /// to determine buffer equality. This is cheaper than `PartialEq::eq` but may
    /// return false when the arrays are logically equal
    pub fn ptr_eq(&self, other: &Self) -> bool {
        if self.offset != other.offset
            || self.len != other.len
            || self.data_type != other.data_type
            || self.buffers.len() != other.buffers.len()
            || self.child_data.len() != other.child_data.len()
        {
            return false;
        }

        match (&self.nulls, &other.nulls) {
            (Some(a), Some(b)) if !a.inner().ptr_eq(b.inner()) => return false,
            (Some(_), None) | (None, Some(_)) => return false,
            _ => {}
        }

        if !self
            .buffers
            .iter()
            .zip(other.buffers.iter())
            .all(|(a, b)| a.as_ptr() == b.as_ptr())
        {
            return false;
        }

        self.child_data
            .iter()
            .zip(other.child_data.iter())
            .all(|(a, b)| a.ptr_eq(b))
    }

    /// Converts this [`ArrayData`] into an [`ArrayDataBuilder`]
    pub fn into_builder(self) -> ArrayDataBuilder {
        self.into()
    }

    /// Claim memory used by this ArrayData in the provided memory pool.
    ///
    /// This claims memory for:
    /// - All buffers in self.buffers
    /// - All child ArrayData recursively
    /// - The null buffer if present
    #[cfg(feature = "pool")]
    pub fn claim(&self, pool: &dyn arrow_buffer::MemoryPool) {
        // Claim all data buffers
        for buffer in &self.buffers {
            buffer.claim(pool);
        }

        // Claim null buffer if present
        if let Some(nulls) = &self.nulls {
            nulls.claim(pool);
        }

        // Recursively claim child data
        for child in &self.child_data {
            child.claim(pool);
        }
    }
}

/// Return the expected [`DataTypeLayout`] Arrays of this data
/// type are expected to have
pub fn layout(data_type: &DataType) -> DataTypeLayout {
    // based on C/C++ implementation in
    // https://github.com/apache/arrow/blob/661c7d749150905a63dd3b52e0a04dac39030d95/cpp/src/arrow/type.h (and .cc)
    use arrow_schema::IntervalUnit::*;

    match data_type {
        DataType::Null => DataTypeLayout {
            buffers: vec![],
            can_contain_null_mask: false,
            variadic: false,
        },
        DataType::Boolean => DataTypeLayout {
            buffers: vec![BufferSpec::BitMap],
            can_contain_null_mask: true,
            variadic: false,
        },
        DataType::Int8 => DataTypeLayout::new_fixed_width::<i8>(),
        DataType::Int16 => DataTypeLayout::new_fixed_width::<i16>(),
        DataType::Int32 => DataTypeLayout::new_fixed_width::<i32>(),
        DataType::Int64 => DataTypeLayout::new_fixed_width::<i64>(),
        DataType::UInt8 => DataTypeLayout::new_fixed_width::<u8>(),
        DataType::UInt16 => DataTypeLayout::new_fixed_width::<u16>(),
        DataType::UInt32 => DataTypeLayout::new_fixed_width::<u32>(),
        DataType::UInt64 => DataTypeLayout::new_fixed_width::<u64>(),
        DataType::Float16 => DataTypeLayout::new_fixed_width::<half::f16>(),
        DataType::Float32 => DataTypeLayout::new_fixed_width::<f32>(),
        DataType::Float64 => DataTypeLayout::new_fixed_width::<f64>(),
        DataType::Timestamp(_, _) => DataTypeLayout::new_fixed_width::<i64>(),
        DataType::Date32 => DataTypeLayout::new_fixed_width::<i32>(),
        DataType::Date64 => DataTypeLayout::new_fixed_width::<i64>(),
        DataType::Time32(_) => DataTypeLayout::new_fixed_width::<i32>(),
        DataType::Time64(_) => DataTypeLayout::new_fixed_width::<i64>(),
        DataType::Interval(YearMonth) => DataTypeLayout::new_fixed_width::<i32>(),
        DataType::Interval(DayTime) => DataTypeLayout::new_fixed_width::<IntervalDayTime>(),
        DataType::Interval(MonthDayNano) => {
            DataTypeLayout::new_fixed_width::<IntervalMonthDayNano>()
        }
        DataType::Duration(_) => DataTypeLayout::new_fixed_width::<i64>(),
        DataType::Decimal32(_, _) => DataTypeLayout::new_fixed_width::<i32>(),
        DataType::Decimal64(_, _) => DataTypeLayout::new_fixed_width::<i64>(),
        DataType::Decimal128(_, _) => DataTypeLayout::new_fixed_width::<i128>(),
        DataType::Decimal256(_, _) => DataTypeLayout::new_fixed_width::<i256>(),
        DataType::FixedSizeBinary(size) => {
            let spec = BufferSpec::FixedWidth {
                byte_width: (*size).try_into().unwrap(),
                alignment: mem::align_of::<u8>(),
            };
            DataTypeLayout {
                buffers: vec![spec],
                can_contain_null_mask: true,
                variadic: false,
            }
        }
        DataType::Binary => DataTypeLayout::new_binary::<i32>(),
        DataType::LargeBinary => DataTypeLayout::new_binary::<i64>(),
        DataType::Utf8 => DataTypeLayout::new_binary::<i32>(),
        DataType::LargeUtf8 => DataTypeLayout::new_binary::<i64>(),
        DataType::BinaryView | DataType::Utf8View => DataTypeLayout::new_view(),
        DataType::FixedSizeList(_, _) => DataTypeLayout::new_nullable_empty(), // all in child data
        DataType::List(_) => DataTypeLayout::new_fixed_width::<i32>(),
        DataType::ListView(_) => DataTypeLayout::new_list_view::<i32>(),
        DataType::LargeListView(_) => DataTypeLayout::new_list_view::<i64>(),
        DataType::LargeList(_) => DataTypeLayout::new_fixed_width::<i64>(),
        DataType::Map(_, _) => DataTypeLayout::new_fixed_width::<i32>(),
        DataType::Struct(_) => DataTypeLayout::new_nullable_empty(), // all in child data,
        DataType::RunEndEncoded(_, _) => DataTypeLayout::new_empty(), // all in child data,
        DataType::Union(_, mode) => {
            let type_ids = BufferSpec::FixedWidth {
                byte_width: mem::size_of::<i8>(),
                alignment: mem::align_of::<i8>(),
            };

            DataTypeLayout {
                buffers: match mode {
                    UnionMode::Sparse => {
                        vec![type_ids]
                    }
                    UnionMode::Dense => {
                        vec![
                            type_ids,
                            BufferSpec::FixedWidth {
                                byte_width: mem::size_of::<i32>(),
                                alignment: mem::align_of::<i32>(),
                            },
                        ]
                    }
                },
                can_contain_null_mask: false,
                variadic: false,
            }
        }
        DataType::Dictionary(key_type, _value_type) => layout(key_type),
    }
}

/// Layout specification for a data type
#[derive(Debug, PartialEq, Eq)]
// Note: Follows structure from C++: https://github.com/apache/arrow/blob/master/cpp/src/arrow/type.h#L91
pub struct DataTypeLayout {
    /// A vector of buffer layout specifications, one for each expected buffer
    pub buffers: Vec<BufferSpec>,

    /// Can contain a null bitmask
    pub can_contain_null_mask: bool,

    /// This field only applies to the view type [`DataType::BinaryView`] and [`DataType::Utf8View`]
    /// If `variadic` is true, the number of buffers expected is only lower-bounded by
    /// buffers.len(). Buffers that exceed the lower bound are legal.
    pub variadic: bool,
}

impl DataTypeLayout {
    /// Describes a basic numeric array where each element has type `T`
    pub fn new_fixed_width<T>() -> Self {
        Self {
            buffers: vec![BufferSpec::FixedWidth {
                byte_width: mem::size_of::<T>(),
                alignment: mem::align_of::<T>(),
            }],
            can_contain_null_mask: true,
            variadic: false,
        }
    }

    /// Describes arrays which have no data of their own
    /// but may still have a Null Bitmap (e.g. FixedSizeList)
    pub fn new_nullable_empty() -> Self {
        Self {
            buffers: vec![],
            can_contain_null_mask: true,
            variadic: false,
        }
    }

    /// Describes arrays which have no data of their own
    /// (e.g. RunEndEncoded).
    pub fn new_empty() -> Self {
        Self {
            buffers: vec![],
            can_contain_null_mask: false,
            variadic: false,
        }
    }

    /// Describes a basic numeric array where each element has a fixed
    /// with offset buffer of type `T`, followed by a
    /// variable width data buffer
    pub fn new_binary<T>() -> Self {
        Self {
            buffers: vec![
                // offsets
                BufferSpec::FixedWidth {
                    byte_width: mem::size_of::<T>(),
                    alignment: mem::align_of::<T>(),
                },
                // values
                BufferSpec::VariableWidth,
            ],
            can_contain_null_mask: true,
            variadic: false,
        }
    }

    /// Describes a view type
    pub fn new_view() -> Self {
        Self {
            buffers: vec![BufferSpec::FixedWidth {
                byte_width: mem::size_of::<u128>(),
                alignment: mem::align_of::<u128>(),
            }],
            can_contain_null_mask: true,
            variadic: true,
        }
    }

    /// Describes a list view type
    pub fn new_list_view<T>() -> Self {
        Self {
            buffers: vec![
                BufferSpec::FixedWidth {
                    byte_width: mem::size_of::<T>(),
                    alignment: mem::align_of::<T>(),
                },
                BufferSpec::FixedWidth {
                    byte_width: mem::size_of::<T>(),
                    alignment: mem::align_of::<T>(),
                },
            ],
            can_contain_null_mask: true,
            variadic: false,
        }
    }
}

/// Layout specification for a single data type buffer
#[derive(Debug, PartialEq, Eq)]
pub enum BufferSpec {
    /// Each element is a fixed width primitive, with the given `byte_width` and `alignment`
    ///
    /// `alignment` is the alignment required by Rust for an array of the corresponding primitive,
    /// see [`Layout::array`](std::alloc::Layout::array) and [`std::mem::align_of`].
    ///
    /// Arrow-rs requires that all buffers have at least this alignment, to allow for
    /// [slice](std::slice) based APIs. Alignment in excess of this is not required to allow
    /// for array slicing and interoperability with `Vec`, which cannot be over-aligned.
    ///
    /// Note that these alignment requirements will vary between architectures
    FixedWidth {
        /// The width of each element in bytes
        byte_width: usize,
        /// The alignment required by Rust for an array of the corresponding primitive
        alignment: usize,
    },
    /// Variable width, such as string data for utf8 data
    VariableWidth,
    /// Buffer holds a bitmap.
    ///
    /// Note: Unlike the C++ implementation, the null/validity buffer
    /// is handled specially rather than as another of the buffers in
    /// the spec, so this variant is only used for the Boolean type.
    BitMap,
    /// Buffer is always null. Unused currently in Rust implementation,
    /// (used in C++ for Union type)
    AlwaysNull,
}

impl PartialEq for ArrayData {
    fn eq(&self, other: &Self) -> bool {
        equal::equal(self, other)
    }
}

/// A boolean flag that cannot be mutated outside of unsafe code.
///
/// Defaults to a value of false.
///
/// This structure is used to enforce safety in the [`ArrayDataBuilder`]
///
/// [`ArrayDataBuilder`]: super::ArrayDataBuilder
///
/// # Example
/// ```rust
/// use arrow_data::UnsafeFlag;
/// assert!(!UnsafeFlag::default().get()); // default is false
/// let mut flag = UnsafeFlag::new();
/// assert!(!flag.get()); // defaults to false
/// // can only set it to true in unsafe code
/// unsafe { flag.set(true) };
/// assert!(flag.get()); // now true
/// ```
#[derive(Debug, Clone)]
#[doc(hidden)]
pub struct UnsafeFlag(bool);

impl UnsafeFlag {
    /// Creates a new `UnsafeFlag` with the value set to `false`.
    ///
    /// See examples on [`Self::new`]
    #[inline]
    pub const fn new() -> Self {
        Self(false)
    }

    /// Sets the value of the flag to the given value
    ///
    /// Note this can purposely only be done in `unsafe` code
    ///
    /// # Safety
    ///
    /// If set, the flag will be set to the given value. There is nothing
    /// immediately unsafe about doing so, however, the flag can be used to
    /// subsequently bypass safety checks in the [`ArrayDataBuilder`].
    #[inline]
    pub unsafe fn set(&mut self, val: bool) {
        self.0 = val;
    }

    /// Returns the value of the flag
    #[inline]
    pub fn get(&self) -> bool {
        self.0
    }
}

// Manual impl to make it clear you can not construct unsafe with true
impl Default for UnsafeFlag {
    fn default() -> Self {
        Self::new()
    }
}

/// Builder for [`ArrayData`] type
#[derive(Debug)]
pub struct ArrayDataBuilder {
    data_type: DataType,
    len: usize,
    null_count: Option<usize>,
    null_bit_buffer: Option<Buffer>,
    nulls: Option<NullBuffer>,
    offset: usize,
    buffers: Vec<Buffer>,
    child_data: Vec<ArrayData>,
    /// Should buffers be realigned (copying if necessary)?
    ///
    /// Defaults to false.
    align_buffers: bool,
    /// Should data validation be skipped for this [`ArrayData`]?
    ///
    /// Defaults to false.
    ///
    /// # Safety
    ///
    /// This flag can only be set to true using `unsafe` APIs. However, once true
    /// subsequent calls to `build()` may result in undefined behavior if the data
    /// is not valid.
    skip_validation: UnsafeFlag,
}

impl ArrayDataBuilder {
    #[inline]
    /// Creates a new array data builder
    pub const fn new(data_type: DataType) -> Self {
        Self {
            data_type,
            len: 0,
            null_count: None,
            null_bit_buffer: None,
            nulls: None,
            offset: 0,
            buffers: vec![],
            child_data: vec![],
            align_buffers: false,
            skip_validation: UnsafeFlag::new(),
        }
    }

    /// Creates a new array data builder from an existing one, changing the data type
    pub fn data_type(self, data_type: DataType) -> Self {
        Self { data_type, ..self }
    }

    #[inline]
    /// Sets the length of the [ArrayData]
    pub const fn len(mut self, n: usize) -> Self {
        self.len = n;
        self
    }

    /// Sets the null buffer of the [ArrayData]
    pub fn nulls(mut self, nulls: Option<NullBuffer>) -> Self {
        self.nulls = nulls;
        self.null_count = None;
        self.null_bit_buffer = None;
        self
    }

    /// Sets the null count of the [ArrayData]
    pub fn null_count(mut self, null_count: usize) -> Self {
        self.null_count = Some(null_count);
        self
    }

    /// Sets the `null_bit_buffer` of the [ArrayData]
    pub fn null_bit_buffer(mut self, buf: Option<Buffer>) -> Self {
        self.nulls = None;
        self.null_bit_buffer = buf;
        self
    }

    /// Sets the offset of the [ArrayData]
    #[inline]
    pub const fn offset(mut self, n: usize) -> Self {
        self.offset = n;
        self
    }

    /// Sets the buffers of the [ArrayData]
    pub fn buffers(mut self, v: Vec<Buffer>) -> Self {
        self.buffers = v;
        self
    }

    /// Adds a single buffer to the [ArrayData]'s buffers
    pub fn add_buffer(mut self, b: Buffer) -> Self {
        self.buffers.push(b);
        self
    }

    /// Adds multiple buffers to the [ArrayData]'s buffers
    pub fn add_buffers<I: IntoIterator<Item = Buffer>>(mut self, bs: I) -> Self {
        self.buffers.extend(bs);
        self
    }

    /// Sets the child data of the [ArrayData]
    pub fn child_data(mut self, v: Vec<ArrayData>) -> Self {
        self.child_data = v;
        self
    }

    /// Adds a single child data to the [ArrayData]'s child data
    pub fn add_child_data(mut self, r: ArrayData) -> Self {
        self.child_data.push(r);
        self
    }

    /// Creates an array data, without any validation
    ///
    /// Note: This is shorthand for
    /// ```rust
    /// # #[expect(unsafe_op_in_unsafe_fn)]
    /// # let mut builder = arrow_data::ArrayDataBuilder::new(arrow_schema::DataType::Null);
    /// # let _ = unsafe {
    /// builder.skip_validation(true).build().unwrap()
    /// # };
    /// ```
    ///
    /// # Safety
    ///
    /// The same caveats as [`ArrayData::new_unchecked`]
    /// apply.
    pub unsafe fn build_unchecked(self) -> ArrayData {
        unsafe { self.skip_validation(true) }.build().unwrap()
    }

    /// Creates an `ArrayData`, consuming `self`
    ///
    /// # Undefined behavior
    ///
    /// By default the underlying buffers are checked to ensure they are valid
    /// Arrow data. However, if the [`Self::skip_validation`] flag has been set
    /// to true (by the `unsafe` API) this validation is skipped. If the data is
    /// not valid, undefined behavior will result.
    pub fn build(self) -> Result<ArrayData, ArrowError> {
        let Self {
            data_type,
            len,
            null_count,
            null_bit_buffer,
            nulls,
            offset,
            buffers,
            child_data,
            align_buffers,
            skip_validation,
        } = self;

        // SAFETY: `skip_validation` is only set to true using `unsafe` APIs.
        let validate = !skip_validation.get() || cfg!(feature = "force_validate");
        if validate && let Some(buffer) = null_bit_buffer.as_ref() {
            // Check before constructing the BooleanBuffer, which would otherwise panic.
            let len_plus_offset = checked_len_plus_offset(&data_type, len, offset)?;
            let needed_len = bit_util::ceil(len_plus_offset, 8);
            if buffer.len() < needed_len {
                return Err(ArrowError::InvalidArgumentError(format!(
                    "null_bit_buffer size too small. got {} needed {}",
                    buffer.len(),
                    needed_len
                )));
            }
        }

        let nulls = nulls
            .or_else(|| {
                let buffer = null_bit_buffer?;
                let buffer = BooleanBuffer::new(buffer, offset, len);
                Some(match null_count {
                    Some(n) => {
                        // SAFETY: call to `data.validate_data()` below validates the null buffer is valid
                        unsafe { NullBuffer::new_unchecked(buffer, n) }
                    }
                    None => NullBuffer::new(buffer),
                })
            })
            .filter(|b| b.null_count() != 0);

        let mut data = ArrayData {
            data_type,
            len,
            offset,
            buffers,
            child_data,
            nulls,
        };

        if align_buffers {
            data.align_buffers();
        }

        if validate {
            data.validate_data()?;
        }
        Ok(data)
    }

    /// Ensure that all buffers are aligned, copying data if necessary
    ///
    /// Rust requires that arrays are aligned to their corresponding primitive,
    /// see [`Layout::array`](std::alloc::Layout::array) and [`std::mem::align_of`].
    ///
    /// [`ArrayData`] therefore requires that all buffers have at least this alignment,
    /// to allow for [slice](std::slice) based APIs. See [`BufferSpec::FixedWidth`].
    ///
    /// As this alignment is architecture specific, and not guaranteed by all arrow implementations,
    /// this flag is provided to automatically copy buffers to a new correctly aligned allocation
    /// when necessary, making it useful when interacting with buffers produced by other systems,
    /// e.g. IPC or FFI.
    ///
    /// If this flag is not enabled, `[Self::build`] return an error on encountering
    /// insufficiently aligned buffers.
    pub fn align_buffers(mut self, align_buffers: bool) -> Self {
        self.align_buffers = align_buffers;
        self
    }

    /// Skips validation of the data.
    ///
    /// If this flag is enabled, `[Self::build`] will skip validation of the
    /// data
    ///
    /// If this flag is not enabled, `[Self::build`] will validate that all
    /// buffers are valid and will return an error if any data is invalid.
    /// Validation can be expensive.
    ///
    /// # Safety
    ///
    /// If validation is skipped, the buffers must form a valid Arrow array,
    /// otherwise undefined behavior will result
    pub unsafe fn skip_validation(mut self, skip_validation: bool) -> Self {
        unsafe {
            self.skip_validation.set(skip_validation);
        }
        self
    }
}

impl From<ArrayData> for ArrayDataBuilder {
    fn from(d: ArrayData) -> Self {
        Self {
            data_type: d.data_type,
            len: d.len,
            offset: d.offset,
            buffers: d.buffers,
            child_data: d.child_data,
            nulls: d.nulls,
            null_bit_buffer: None,
            null_count: None,
            align_buffers: false,
            skip_validation: UnsafeFlag::new(),
        }
    }
}

/// Get byte width of FixedSizeBinary size
/// # Panics:
/// - Panics if the `data_type` is not FixedSizeBinary
/// - Panics if byte width is negative
pub(crate) fn get_fixed_size_binary_width(data_type: &DataType) -> usize {
    match data_type {
        DataType::FixedSizeBinary(i) => {
            if *i < 0 {
                panic!("cannot compare FixedSizeBinary({})", *i);
            }
            *i as usize
        }
        _ => unreachable!(),
    }
}

#[cfg(test)]
mod tests {
    use arrow_schema::UnionFields;

    use super::*;
    use crate::ByteView;
    use crate::transform::MutableArrayData;
    use arrow_buffer::{OffsetBuffer, ScalarBuffer};
    use arrow_schema::{Field, Fields};

    // See arrow/tests/array_data_validation.rs for test of array validation

    /// returns a buffer initialized with some constant value for tests
    fn make_i32_buffer(n: usize) -> Buffer {
        Buffer::from_slice_ref(vec![42i32; n])
    }

    /// returns a buffer initialized with some constant value for tests
    fn make_f32_buffer(n: usize) -> Buffer {
        Buffer::from_slice_ref(vec![42f32; n])
    }

    #[test]
    fn test_builder() {
        // Buffer needs to be at least 25 long
        let v = (0..25).collect::<Vec<i32>>();
        let b1 = Buffer::from_slice_ref(&v);
        let arr_data = ArrayData::builder(DataType::Int32)
            .len(20)
            .offset(5)
            .add_buffer(b1)
            .null_bit_buffer(Some(Buffer::from([
                0b01011111, 0b10110101, 0b01100011, 0b00011110,
            ])))
            .build()
            .unwrap();

        assert_eq!(20, arr_data.len());
        assert_eq!(10, arr_data.null_count());
        assert_eq!(5, arr_data.offset());
        assert_eq!(1, arr_data.buffers().len());
        assert_eq!(
            Buffer::from_slice_ref(&v).as_slice(),
            arr_data.buffers()[0].as_slice()
        );
    }

    #[test]
    fn test_builder_with_child_data() {
        let child_arr_data = ArrayData::try_new(
            DataType::Int32,
            5,
            None,
            0,
            vec![Buffer::from_slice_ref([1i32, 2, 3, 4, 5])],
            vec![],
        )
        .unwrap();

        let field = Arc::new(Field::new("x", DataType::Int32, true));
        let data_type = DataType::Struct(vec![field].into());

        let arr_data = ArrayData::builder(data_type)
            .len(5)
            .offset(0)
            .add_child_data(child_arr_data.clone())
            .build()
            .unwrap();

        assert_eq!(5, arr_data.len());
        assert_eq!(1, arr_data.child_data().len());
        assert_eq!(child_arr_data, arr_data.child_data()[0]);
    }

    #[test]
    fn test_struct_validation_accounts_for_parent_offset() {
        let data_type =
            DataType::Struct(Fields::from(vec![Field::new("x", DataType::Int32, false)]));
        let child = ArrayData::builder(DataType::Int32)
            .len(5)
            .add_buffer(Buffer::from_slice_ref([0, 1, 2, 3, 4]))
            .build()
            .unwrap();

        // The parent needs child elements 1..6, but the child only has five.
        let err = ArrayData::builder(data_type)
            .len(5)
            .offset(1)
            .add_child_data(child)
            .build()
            .unwrap_err()
            .to_string();

        assert!(err.contains(
            "child array #0 for field x has length smaller than expected for struct array (5 < 6)"
        ));
    }

    #[test]
    fn test_struct_non_nullable_child_nulls_account_for_parent_offset() {
        let build = |parent_nulls| {
            let child = ArrayData::builder(DataType::Int32)
                .len(5)
                .add_buffer(Buffer::from_slice_ref([0, 1, 2, 3, 4]))
                .nulls(Some(NullBuffer::new(BooleanBuffer::from(vec![
                    true, true, false, true, true,
                ]))))
                .build()
                .unwrap();

            ArrayData::builder(DataType::Struct(Fields::from(vec![Field::new(
                "x",
                DataType::Int32,
                false,
            )])))
            .len(4)
            .offset(1)
            .nulls(Some(NullBuffer::new(BooleanBuffer::from(parent_nulls))))
            .add_child_data(child)
            .build()
        };

        assert!(build(vec![true, false, true, true]).is_ok());
        assert!(build(vec![true, true, false, true]).is_err());
    }

    #[test]
    fn test_fixed_size_list_non_nullable_child_nulls_account_for_parent_offset() {
        for child_offset in [0, 1] {
            let child = ArrayData::builder(DataType::Int32)
                .len(6)
                .offset(child_offset)
                .add_buffer(Buffer::from_slice_ref([0i32; 7]))
                .nulls(Some(NullBuffer::from(vec![
                    true, true, false, false, true, true,
                ])))
                .build()
                .unwrap();
            let build = |parent_nulls| {
                ArrayData::builder(DataType::FixedSizeList(
                    Arc::new(Field::new("item", DataType::Int32, false)),
                    2,
                ))
                .offset(1)
                .len(2)
                .nulls(Some(NullBuffer::from(parent_nulls)))
                .add_child_data(child.clone())
                .build()
            };

            // The first visible list contains both child nulls.
            assert!(build(vec![false, true]).is_ok());
            assert!(build(vec![true, false]).is_err());
        }
    }

    #[test]
    fn test_fixed_size_list_non_nullable_child_nulls_without_parent_nulls() {
        let child = ArrayData::builder(DataType::Int32)
            .len(6)
            .add_buffer(Buffer::from_slice_ref([0i32; 6]))
            .nulls(Some(NullBuffer::from(vec![
                false, false, true, true, false, false,
            ])))
            .build()
            .unwrap();
        let build = |offset, len, size| {
            ArrayData::builder(DataType::FixedSizeList(
                Arc::new(Field::new("item", DataType::Int32, false)),
                size,
            ))
            .offset(offset)
            .len(len)
            .add_child_data(child.clone())
            .build()
        };

        // Nulls before and after the visible child range are irrelevant.
        assert!(build(1, 1, 2).is_ok());
        assert!(build(0, 1, 2).is_err());
        assert!(build(2, 1, 2).is_err());
        assert!(build(3, 0, 2).is_ok());
        assert!(build(1, 2, 0).is_ok());
    }

    #[test]
    fn test_fixed_size_list_validation_accounts_for_parent_offset() {
        for nulls in [None, Some(NullBuffer::new_null(4))] {
            let child = ArrayData::builder(DataType::Int32)
                .len(4)
                .add_buffer(Buffer::from_slice_ref([0i32; 4]))
                .nulls(nulls)
                .build()
                .unwrap();
            let err = ArrayData::builder(DataType::FixedSizeList(
                Arc::new(Field::new("item", DataType::Int32, false)),
                2,
            ))
            .offset(1)
            .len(2)
            .add_child_data(child)
            .build()
            .unwrap_err()
            .to_string();
            assert!(err.contains("Values length 4 is less than"));
        }
    }

    #[test]
    fn test_struct_equal_accounts_for_parent_offset() {
        let data_type =
            DataType::Struct(Fields::from(vec![Field::new("x", DataType::Int32, false)]));

        let child1 = ArrayData::builder(DataType::Int32)
            .len(5)
            .add_buffer(Buffer::from_slice_ref([0, 1, 2, 3, 4]))
            .build()
            .unwrap();
        let child2 = child1.slice(1, 4);

        // data1 has offset at parent level; data2 has offset at child level.
        let data1 = ArrayData::builder(data_type.clone())
            .len(4)
            .offset(1)
            .add_child_data(child1)
            .build()
            .unwrap();
        let data2 = ArrayData::builder(data_type)
            .len(4)
            .add_child_data(child2)
            .build()
            .unwrap();

        assert_eq!(data1, data2);
    }

    #[test]
    fn test_extend_struct_accounts_for_parent_offset() {
        let data_type =
            DataType::Struct(Fields::from(vec![Field::new("x", DataType::Int32, false)]));
        let child = ArrayData::builder(DataType::Int32)
            .len(5)
            .add_buffer(Buffer::from_slice_ref([0, 1, 2, 3, 4]))
            .build()
            .unwrap();

        let data = ArrayData::builder(data_type)
            .len(4)
            .offset(1)
            .add_child_data(child)
            .build()
            .unwrap();

        let mut mutable = MutableArrayData::new(vec![&data], false, data.len());
        mutable.try_extend(0, 0, data.len()).unwrap();
        let output = mutable.freeze();

        assert_eq!(output.child_data()[0].buffer::<i32>(0), &[1, 2, 3, 4]);
    }

    #[test]
    fn test_extend_fixed_size_list_accounts_for_parent_offset() {
        let field = Arc::new(Field::new("item", DataType::Int32, false));
        let child = ArrayData::builder(DataType::Int32)
            .len(6)
            .add_buffer(Buffer::from_slice_ref([0, 1, 2, 3, 4, 5]))
            .build()
            .unwrap();

        let data = ArrayData::builder(DataType::FixedSizeList(field, 2))
            .len(2)
            .offset(1)
            .add_child_data(child)
            .build()
            .unwrap();

        let mut mutable = MutableArrayData::new(vec![&data], false, data.len());
        mutable.try_extend(0, 0, data.len()).unwrap();
        let output = mutable.freeze();

        assert_eq!(output.child_data()[0].buffer::<i32>(0), &[2, 3, 4, 5]);
    }

    #[test]
    fn test_extend_sparse_union_accounts_for_parent_offset() {
        let fields = UnionFields::try_new(
            [0, 1],
            [
                Field::new("a", DataType::Int32, false),
                Field::new("b", DataType::Int32, false),
            ],
        )
        .unwrap();
        let child_a = ArrayData::builder(DataType::Int32)
            .len(3)
            .add_buffer(Buffer::from_slice_ref([10, 20, 30]))
            .build()
            .unwrap();
        let child_b = ArrayData::builder(DataType::Int32)
            .len(3)
            .add_buffer(Buffer::from_slice_ref([100, 200, 300]))
            .build()
            .unwrap();

        // The children are left whole, as `ArrayData::slice` and C Data
        // Interface producers leave them for a sparse union
        let data = ArrayData::builder(DataType::Union(fields, UnionMode::Sparse))
            .len(2)
            .offset(1)
            .add_buffer(Buffer::from_slice_ref([0i8, 1, 0]))
            .add_child_data(child_a)
            .add_child_data(child_b)
            .build()
            .unwrap();

        let mut mutable = MutableArrayData::new(vec![&data], false, data.len());
        mutable.try_extend(0, 0, data.len()).unwrap();
        let output = mutable.freeze();

        assert_eq!(output.buffer::<i8>(0), &[1, 0]);
        assert_eq!(output.child_data()[0].buffer::<i32>(0), &[20, 30]);
        assert_eq!(output.child_data()[1].buffer::<i32>(0), &[200, 300]);
    }

    #[test]
    fn test_null_count() {
        let mut bit_v: [u8; 2] = [0; 2];
        bit_util::set_bit(&mut bit_v, 0);
        bit_util::set_bit(&mut bit_v, 3);
        bit_util::set_bit(&mut bit_v, 10);
        let arr_data = ArrayData::builder(DataType::Int32)
            .len(16)
            .add_buffer(make_i32_buffer(16))
            .null_bit_buffer(Some(Buffer::from(bit_v)))
            .build()
            .unwrap();
        assert_eq!(13, arr_data.null_count());

        // Test with offset
        let mut bit_v: [u8; 2] = [0; 2];
        bit_util::set_bit(&mut bit_v, 0);
        bit_util::set_bit(&mut bit_v, 3);
        bit_util::set_bit(&mut bit_v, 10);
        let arr_data = ArrayData::builder(DataType::Int32)
            .len(12)
            .offset(2)
            .add_buffer(make_i32_buffer(14)) // requires at least 14 bytes of space,
            .null_bit_buffer(Some(Buffer::from(bit_v)))
            .build()
            .unwrap();
        assert_eq!(10, arr_data.null_count());
    }

    #[test]
    fn test_null_buffer_ref() {
        let mut bit_v: [u8; 2] = [0; 2];
        bit_util::set_bit(&mut bit_v, 0);
        bit_util::set_bit(&mut bit_v, 3);
        bit_util::set_bit(&mut bit_v, 10);
        let arr_data = ArrayData::builder(DataType::Int32)
            .len(16)
            .add_buffer(make_i32_buffer(16))
            .null_bit_buffer(Some(Buffer::from(bit_v)))
            .build()
            .unwrap();
        assert!(arr_data.nulls().is_some());
        assert_eq!(&bit_v, arr_data.nulls().unwrap().validity());
    }

    #[test]
    fn test_slice() {
        let mut bit_v: [u8; 2] = [0; 2];
        bit_util::set_bit(&mut bit_v, 0);
        bit_util::set_bit(&mut bit_v, 3);
        bit_util::set_bit(&mut bit_v, 10);
        let data = ArrayData::builder(DataType::Int32)
            .len(16)
            .add_buffer(make_i32_buffer(16))
            .null_bit_buffer(Some(Buffer::from(bit_v)))
            .build()
            .unwrap();
        let new_data = data.slice(1, 15);
        assert_eq!(data.len() - 1, new_data.len());
        assert_eq!(1, new_data.offset());
        assert_eq!(data.null_count(), new_data.null_count());

        // slice of a slice (removes one null)
        let new_data = new_data.slice(1, 14);
        assert_eq!(data.len() - 2, new_data.len());
        assert_eq!(2, new_data.offset());
        assert_eq!(data.null_count() - 1, new_data.null_count());
    }

    #[test]
    #[should_panic(expected = "offset + length overflow")]
    fn test_slice_panics_on_offset_length_overflow() {
        let data = ArrayData::builder(DataType::Int32)
            .len(4)
            .add_buffer(make_i32_buffer(4))
            .build()
            .unwrap();
        let sliced = data.slice(1, 3);

        sliced.slice(1, usize::MAX);
    }

    #[test]
    fn test_sliced() {
        for nulls in [
            None,
            Some(NullBuffer::from(vec![true, false, true, false, true])),
        ] {
            let data = ArrayData::builder(DataType::Int32)
                .len(5)
                .offset(1)
                .add_buffer(make_i32_buffer(6))
                .nulls(nulls)
                .build()
                .unwrap();

            for offset in 0..=data.len() {
                for length in 0..=data.len() - offset {
                    let owned = data.clone();
                    let buffers = owned.buffers().as_ptr();
                    let sliced = owned.sliced(offset, length);
                    assert_eq!(sliced, data.slice(offset, length));
                    assert_eq!(sliced.len(), length);
                    assert_eq!(sliced.offset(), 1 + offset);
                    assert_eq!(sliced.buffers().as_ptr(), buffers);
                    assert_eq!(sliced.buffers()[0].as_ptr(), data.buffers()[0].as_ptr());
                    if let Some(nulls) = data.nulls() {
                        assert_eq!(
                            sliced.nulls().unwrap().iter().collect::<Vec<_>>(),
                            nulls.iter().skip(offset).take(length).collect::<Vec<_>>()
                        );
                    }
                    sliced.validate_full().unwrap();
                }
            }
        }

        let empty = ArrayData::new_empty(&DataType::Int32);
        assert_eq!(empty.clone().sliced(0, 0), empty);
    }

    #[test]
    fn test_sliced_nested_struct() {
        let child = ArrayData::builder(DataType::Int32)
            .len(8)
            .offset(1)
            .add_buffer(make_i32_buffer(9))
            .nulls(Some(NullBuffer::from(vec![
                true, false, true, false, true, true, false, true,
            ])))
            .build()
            .unwrap();
        let inner = ArrayData::builder(DataType::Struct(Fields::from(vec![Field::new(
            "x",
            DataType::Int32,
            true,
        )])))
        .len(6)
        .offset(1)
        .nulls(Some(NullBuffer::from(vec![
            true, false, true, true, false, true,
        ])))
        .add_child_data(child)
        .build()
        .unwrap();
        let data = ArrayData::builder(DataType::Struct(Fields::from(vec![Field::new(
            "inner",
            inner.data_type().clone(),
            true,
        )])))
        .len(4)
        .offset(1)
        .nulls(Some(NullBuffer::from(vec![false, true, false, true])))
        .add_child_data(inner)
        .build()
        .unwrap();

        for (offset, length) in [(0, 4), (1, 2), (4, 0), (0, 0)] {
            let owned = data.clone();
            let children = owned.child_data().as_ptr();
            let grandchildren = owned.child_data()[0].child_data().as_ptr();
            let buffers = owned.child_data()[0].child_data()[0].buffers().as_ptr();
            let sliced = owned.sliced(offset, length);
            let inner = &sliced.child_data()[0];
            let child = &inner.child_data()[0];
            assert_eq!(sliced, data.slice(offset, length));
            assert_eq!(sliced.offset(), 0);
            assert_eq!(inner.offset(), 0);
            assert_eq!(child.offset(), 3 + offset);
            assert_eq!(sliced.child_data().as_ptr(), children);
            assert_eq!(inner.child_data().as_ptr(), grandchildren);
            assert_eq!(child.buffers().as_ptr(), buffers);
            sliced.validate_full().unwrap();
        }

        let sliced = data.sliced(1, 2).sliced(1, 1);
        assert_eq!(sliced.null_count(), 1);
        assert_eq!(sliced.child_data()[0].null_count(), 0);
        assert_eq!(sliced.child_data()[0].child_data()[0].offset(), 5);
        assert_eq!(sliced.child_data()[0].child_data()[0].null_count(), 0);
    }

    #[test]
    fn test_sliced_list_preserves_children() {
        let child = ArrayData::builder(DataType::Int32)
            .len(6)
            .offset(1)
            .add_buffer(make_i32_buffer(7))
            .nulls(Some(NullBuffer::from(vec![
                true, false, true, false, true, true,
            ])))
            .build()
            .unwrap();
        let data = ArrayData::builder(DataType::List(Arc::new(Field::new(
            "item",
            DataType::Int32,
            true,
        ))))
        .len(4)
        .offset(1)
        .add_buffer(Buffer::from_slice_ref([0i32, 1, 2, 3, 4, 5]))
        .nulls(Some(NullBuffer::from(vec![false, true, false, true])))
        .add_child_data(child)
        .build()
        .unwrap();
        let buffers = data.buffers().as_ptr();
        let children = data.child_data().as_ptr();
        let child_buffers = data.child_data()[0].buffers().as_ptr();
        let expected = data.slice(1, 2);
        let sliced = data.sliced(1, 2);
        assert_eq!(sliced, expected);
        assert_eq!(sliced.offset(), 2);
        assert_eq!(sliced.child_data()[0].len(), 6);
        assert_eq!(sliced.child_data()[0].offset(), 1);
        assert_eq!(sliced.buffers().as_ptr(), buffers);
        assert_eq!(sliced.child_data().as_ptr(), children);
        assert_eq!(sliced.child_data()[0].buffers().as_ptr(), child_buffers);
        sliced.validate_full().unwrap();
    }

    #[test]
    #[should_panic(expected = "offset + length overflow")]
    fn test_sliced_panics_on_offset_length_overflow() {
        ArrayData::new_null(&DataType::Int32, 4).sliced(1, usize::MAX);
    }

    #[test]
    #[should_panic(expected = "end <= self.len()")]
    fn test_sliced_panics_on_offset_out_of_bounds() {
        ArrayData::new_null(&DataType::Int32, 4).sliced(5, 0);
    }

    #[test]
    #[should_panic(expected = "end <= self.len()")]
    fn test_sliced_panics_on_length_out_of_bounds() {
        ArrayData::new_null(&DataType::Int32, 4).sliced(3, 2);
    }

    #[test]
    fn test_typed_offsets_length_overflow() {
        let data = ArrayData {
            data_type: DataType::Binary,
            len: usize::MAX,
            offset: 0,
            buffers: vec![Buffer::from_slice_ref([0_i32])],
            child_data: vec![],
            nulls: None,
        };
        let err = data.typed_offsets::<i32>().unwrap_err();

        assert_eq!(
            err.to_string(),
            format!(
                "Invalid argument error: Length {} with offset 1 overflows usize for Binary",
                usize::MAX
            )
        );
    }

    #[test]
    fn test_validate_typed_buffer_length_overflow() {
        let data = ArrayData {
            data_type: DataType::Binary,
            len: 0,
            offset: 2,
            buffers: vec![Buffer::from_slice_ref([0_i32])],
            child_data: vec![],
            nulls: None,
        };
        let err = data.typed_buffer::<i32>(0, usize::MAX).unwrap_err();

        assert_eq!(
            err.to_string(),
            format!(
                "Invalid argument error: Length {} with offset 2 overflows usize for Binary",
                usize::MAX
            )
        );
    }

    // Exercises ArrayData::try_new with len + offset overflowing
    fn try_new_binary_length_offset_overflow() -> Result<ArrayData, ArrowError> {
        ArrayData::try_new(
            DataType::Binary,
            usize::MAX,
            None,
            1,
            vec![
                Buffer::from_slice_ref([0_i32]),
                Buffer::from_iter(std::iter::empty::<u8>()),
            ],
            vec![],
        )
    }

    #[cfg(not(feature = "force_validate"))]
    #[test]
    fn test_try_new_length_offset_overflow() {
        let err = try_new_binary_length_offset_overflow().unwrap_err();

        assert_eq!(
            err.to_string(),
            format!(
                "Invalid argument error: Length {} with offset 1 overflows usize for Binary",
                usize::MAX
            )
        );
    }

    #[cfg(feature = "force_validate")]
    #[test]
    #[should_panic(
        expected = "Length 18446744073709551615 with offset 1 overflows usize for Binary"
    )]
    fn test_try_new_length_offset_overflow_force_validate() {
        try_new_binary_length_offset_overflow().unwrap();
    }

    #[test]
    fn test_equality() {
        let int_data = ArrayData::builder(DataType::Int32)
            .len(1)
            .add_buffer(make_i32_buffer(1))
            .build()
            .unwrap();

        let float_data = ArrayData::builder(DataType::Float32)
            .len(1)
            .add_buffer(make_f32_buffer(1))
            .build()
            .unwrap();
        assert_ne!(int_data, float_data);
        assert!(!int_data.ptr_eq(&float_data));
        assert!(int_data.ptr_eq(&int_data));

        let int_data_clone = int_data.clone();
        assert_eq!(int_data, int_data_clone);
        assert!(int_data.ptr_eq(&int_data_clone));
        assert!(int_data_clone.ptr_eq(&int_data));

        let int_data_slice = int_data_clone.slice(1, 0);
        assert!(int_data_slice.ptr_eq(&int_data_slice));
        assert!(!int_data.ptr_eq(&int_data_slice));
        assert!(!int_data_slice.ptr_eq(&int_data));

        let data_buffer = Buffer::from_slice_ref(b"abcdef");
        let offsets_buffer = Buffer::from_slice_ref([0_i32, 2_i32, 2_i32, 5_i32]);
        let string_data = ArrayData::try_new(
            DataType::Utf8,
            3,
            Some(Buffer::from_iter(vec![true, false, true])),
            0,
            vec![offsets_buffer, data_buffer],
            vec![],
        )
        .unwrap();

        assert_ne!(float_data, string_data);
        assert!(!float_data.ptr_eq(&string_data));

        assert!(string_data.ptr_eq(&string_data));

        let string_data_cloned = string_data.clone();
        assert!(string_data_cloned.ptr_eq(&string_data));
        assert!(string_data.ptr_eq(&string_data_cloned));

        let string_data_slice = string_data.slice(1, 2);
        assert!(string_data_slice.ptr_eq(&string_data_slice));
        assert!(!string_data_slice.ptr_eq(&string_data))
    }

    #[test]
    fn test_slice_memory_size_view_payload_buffers() {
        for data_type in [DataType::Utf8View, DataType::BinaryView] {
            let inline_only = ArrayData::builder(data_type.clone())
                .len(2)
                .add_buffer(Buffer::from_vec(vec![0_u128; 2]))
                .build()
                .unwrap();
            assert_eq!(
                inline_only.get_slice_memory_size().unwrap(),
                2 * mem::size_of::<u128>()
            );

            let mut first_payload = Vec::with_capacity(32);
            first_payload.extend_from_slice(b"first payload");
            let first_view =
                ByteView::new(first_payload.len().try_into().unwrap(), &first_payload[..4])
                    .as_u128();
            let first_payload = Buffer::from_vec(first_payload);
            assert!(first_payload.capacity() > first_payload.len());
            let first_payload_capacity = first_payload.capacity();

            let mut second_payload = Vec::with_capacity(64);
            second_payload.extend_from_slice(b"second payload");
            let second_view = ByteView::new(
                second_payload.len().try_into().unwrap(),
                &second_payload[..4],
            )
            .with_buffer_index(1)
            .as_u128();
            let second_payload = Buffer::from_vec(second_payload);
            assert!(second_payload.capacity() > second_payload.len());
            let second_payload_capacity = second_payload.capacity();

            let data = ArrayData::builder(data_type)
                .len(3)
                .add_buffer(Buffer::from_vec(vec![first_view, 0_u128, second_view]))
                .add_buffer(first_payload)
                .add_buffer(second_payload)
                .build()
                .unwrap();
            let sliced = data.slice(1, 1);

            assert_eq!(
                sliced.get_slice_memory_size().unwrap(),
                mem::size_of::<u128>() + first_payload_capacity + second_payload_capacity
            );
        }
    }

    #[test]
    fn test_slice_memory_size_utf8_offset_buffer_len_plus_one() {
        // 2-element array ["hello", "world"]: array len = 2, 10 bytes
        let data_buffer = Buffer::from_slice_ref(b"helloworld");
        // offsets need array_len+1 entries to mark the end of every string:
        //   [0, 5, 10] -> 3 i32s = 12 bytes
        let offsets_buffer = Buffer::from_slice_ref([0_i32, 5_i32, 10_i32]);
        let array = ArrayData::try_new(
            DataType::Utf8,
            2,
            None,
            0,
            vec![offsets_buffer, data_buffer],
            vec![],
        )
        .unwrap();
        assert_eq!(array.get_slice_memory_size().unwrap(), 22); // 12 + 10
    }

    #[test]
    fn test_slice_memory_size_binary_offset_buffer_len_plus_one() {
        // 2-element array: array len = 2, not 3
        // values: 5 bytes
        let data_buffer = Buffer::from_slice_ref([0u8, 1, 2, 3, 4]);
        // offsets need array_len+1 entries to mark the end of every element:
        let offsets_buffer = Buffer::from_slice_ref([0_i32, 2_i32, 5_i32]);
        let array = ArrayData::try_new(
            DataType::Binary,
            2,
            None,
            0,
            vec![offsets_buffer, data_buffer],
            vec![],
        )
        .unwrap();
        assert_eq!(array.get_slice_memory_size().unwrap(), 17); // 12 + 5
    }

    #[test]
    fn test_slice_memory_size() {
        let mut bit_v: [u8; 2] = [0; 2];
        bit_util::set_bit(&mut bit_v, 0);
        bit_util::set_bit(&mut bit_v, 3);
        bit_util::set_bit(&mut bit_v, 10);
        let data = ArrayData::builder(DataType::Int32)
            .len(16)
            .add_buffer(make_i32_buffer(16))
            .null_bit_buffer(Some(Buffer::from(bit_v)))
            .build()
            .unwrap();
        let new_data = data.slice(1, 14);
        assert_eq!(
            data.get_slice_memory_size().unwrap() - 8,
            new_data.get_slice_memory_size().unwrap()
        );
        let data_buffer = Buffer::from_slice_ref(b"abcdef");
        let offsets_buffer = Buffer::from_slice_ref([0_i32, 2_i32, 2_i32, 5_i32]);
        let string_data = ArrayData::try_new(
            DataType::Utf8,
            3,
            Some(Buffer::from_iter(vec![true, false, true])),
            0,
            vec![offsets_buffer, data_buffer],
            vec![],
        )
        .unwrap();
        let string_data_slice = string_data.slice(1, 2);
        //4 bytes of offset and 2 bytes of data reduced by slicing.
        assert_eq!(
            string_data.get_slice_memory_size().unwrap() - 6,
            string_data_slice.get_slice_memory_size().unwrap()
        );
    }

    #[test]
    fn test_builder_rejects_short_null_bit_buffer() {
        for (len, offset) in [(8000, 0), (8, 1)] {
            let err = ArrayData::builder(DataType::Int32)
                .len(len)
                .offset(offset)
                .add_buffer(make_i32_buffer(len + offset))
                .null_bit_buffer(Some(Buffer::from([0_u8])))
                .build()
                .unwrap_err();
            assert_eq!(
                err.to_string(),
                format!(
                    "Invalid argument error: null_bit_buffer size too small. got 1 needed {}",
                    bit_util::ceil(len + offset, 8)
                )
            );
        }
    }

    #[test]
    fn test_builder_null_bit_buffer_length_overflow() {
        let err = ArrayData::builder(DataType::Int32)
            .len(usize::MAX)
            .offset(1)
            .null_bit_buffer(Some(Buffer::default()))
            .build()
            .unwrap_err();
        assert_eq!(
            err.to_string(),
            format!(
                "Invalid argument error: Length {} with offset 1 overflows usize for Int32",
                usize::MAX
            )
        );
    }

    #[test]
    fn test_builder_accepts_valid_null_bit_buffer() {
        for (len, offset) in [(8, 0), (7, 1)] {
            let data = ArrayData::builder(DataType::Int32)
                .len(len)
                .offset(offset)
                .add_buffer(make_i32_buffer(len + offset))
                .null_bit_buffer(Some(Buffer::from([0_u8])))
                .build()
                .unwrap();
            assert_eq!(data.len(), len);
            assert_eq!(data.offset(), offset);
            assert_eq!(data.null_count(), len);
        }
    }

    #[test]
    fn test_count_nulls() {
        let buffer = Buffer::from([0b00010110, 0b10011111]);
        let buffer = NullBuffer::new(BooleanBuffer::new(buffer, 0, 16));
        let count = count_nulls(Some(&buffer), 0, 16);
        assert_eq!(count, 7);

        let count = count_nulls(Some(&buffer), 4, 8);
        assert_eq!(count, 3);
    }

    #[test]
    fn test_contains_nulls() {
        let buffer: Buffer =
            MutableBuffer::from_iter([false, false, false, true, true, false]).into();
        let buffer = NullBuffer::new(BooleanBuffer::new(buffer, 0, 6));
        assert!(contains_nulls(Some(&buffer), 0, 6));
        assert!(contains_nulls(Some(&buffer), 0, 3));
        assert!(!contains_nulls(Some(&buffer), 3, 2));
        assert!(!contains_nulls(Some(&buffer), 0, 0));
    }

    #[test]
    fn test_alignment() {
        let buffer = Buffer::from_vec(vec![1_i32, 2_i32, 3_i32]);
        let sliced = buffer.slice(1);

        let mut data = ArrayData {
            data_type: DataType::Int32,
            len: 0,
            offset: 0,
            buffers: vec![buffer],
            child_data: vec![],
            nulls: None,
        };
        data.validate_full().unwrap();

        // break alignment in data
        data.buffers[0] = sliced;
        let err = data.validate().unwrap_err();

        assert_eq!(
            err.to_string(),
            "Invalid argument error: Misaligned buffers[0] in array of type Int32, offset from expected alignment of 4 by 1"
        );

        data.align_buffers();
        data.validate_full().unwrap();
    }

    #[test]
    fn test_alignment_struct() {
        let buffer = Buffer::from_vec(vec![1_i32, 2_i32, 3_i32]);
        let sliced = buffer.slice(1);

        let child_data = ArrayData {
            data_type: DataType::Int32,
            len: 0,
            offset: 0,
            buffers: vec![buffer],
            child_data: vec![],
            nulls: None,
        };

        let schema = DataType::Struct(Fields::from(vec![Field::new("a", DataType::Int32, false)]));
        let mut data = ArrayData {
            data_type: schema,
            len: 0,
            offset: 0,
            buffers: vec![],
            child_data: vec![child_data],
            nulls: None,
        };
        data.validate_full().unwrap();

        // break alignment in child data
        data.child_data[0].buffers[0] = sliced;
        let err = data.validate().unwrap_err();

        assert_eq!(
            err.to_string(),
            "Invalid argument error: Misaligned buffers[0] in array of type Int32, offset from expected alignment of 4 by 1"
        );

        data.align_buffers();
        data.validate_full().unwrap();
    }

    #[test]
    fn test_null_view_types() {
        let array_len = 32;
        let array = ArrayData::new_null(&DataType::BinaryView, array_len);
        assert_eq!(array.len(), array_len);
        for i in 0..array.len() {
            assert!(array.is_null(i));
        }

        let array = ArrayData::new_null(&DataType::Utf8View, array_len);
        assert_eq!(array.len(), array_len);
        for i in 0..array.len() {
            assert!(array.is_null(i));
        }

        let array = ArrayData::new_null(
            &DataType::ListView(Arc::new(Field::new_list_field(DataType::Int32, true))),
            array_len,
        );
        assert_eq!(array.len(), array_len);
        for i in 0..array.len() {
            assert!(array.is_null(i));
        }

        let array = ArrayData::new_null(
            &DataType::LargeListView(Arc::new(Field::new_list_field(DataType::Int32, true))),
            array_len,
        );
        assert_eq!(array.len(), array_len);
        for i in 0..array.len() {
            assert!(array.is_null(i));
        }
    }

    // Even when `force_validate` feature is on
    #[test]
    fn test_dont_panic_on_bad_input_when_using_try_new() {
        let empty_bytes = Buffer::default();

        let array_data = ArrayData::try_new(
            DataType::Utf8,
            1, // len
            None,
            0,
            // the offsets says that we have 2 bytes but the buffer is empty
            vec![Buffer::from_vec(vec![0i32, 2i32]), empty_bytes],
            vec![],
        );

        let res = array_data.expect_err("should get error");

        assert_eq!(
            res.to_string(),
            "Invalid argument error: Last offset 2 of Utf8 is larger than values length 0"
        );
    }

    /// Without `force_validate`, `build_unchecked` skips validation, so these can
    /// reach `validate_values` with a data type that has an invalid child type.
    #[test]
    #[cfg(not(feature = "force_validate"))]
    fn test_validate_values_rejects_a_non_integer_dictionary_key() {
        let values = valid_non_nullable_int32_array_data(2);
        let data_type = DataType::Dictionary(Box::new(DataType::Utf8), Box::new(DataType::Int32));
        let dictionary = unsafe {
            ArrayData::builder(data_type)
                .len(1)
                .add_child_data(values)
                .build_unchecked()
        };

        let err = dictionary.validate_values().expect_err("should get error");
        assert_eq!(
            err.to_string(),
            "Invalid argument error: Dictionary key type must be an integer, got Utf8"
        );
    }

    #[test]
    #[cfg(not(feature = "force_validate"))]
    fn test_validate_values_rejects_a_non_integer_run_end() {
        let data_type = DataType::RunEndEncoded(
            Arc::new(Field::new(
                Field::REE_RUN_ENDS_FIELD_DEFAULT_NAME,
                DataType::Utf8,
                false,
            )),
            Arc::new(Field::new(
                Field::REE_VALUES_FIELD_DEFAULT_NAME,
                DataType::Int32,
                true,
            )),
        );
        let run_end_encoded = unsafe {
            ArrayData::builder(data_type)
                .len(1)
                .add_child_data(valid_non_nullable_int32_array_data(1))
                .add_child_data(valid_non_nullable_int32_array_data(1))
                .build_unchecked()
        };

        let err = run_end_encoded
            .validate_values()
            .expect_err("should get error");
        assert_eq!(
            err.to_string(),
            "Invalid argument error: Run end type must be Int16, Int32 or Int64, got Utf8"
        );
    }

    /// Returns a `RunEndEncoded<Int32, Float32>` data type and run ends / values
    /// children whose runs end at `[1, 3]`, i.e. a logical length of 3.
    fn run_array_test_children() -> (DataType, Vec<ArrayData>) {
        let dt = DataType::RunEndEncoded(
            Arc::new(Field::new(
                Field::REE_RUN_ENDS_FIELD_DEFAULT_NAME,
                DataType::Int32,
                false,
            )),
            Arc::new(Field::new(
                Field::REE_VALUES_FIELD_DEFAULT_NAME,
                DataType::Float32,
                true,
            )),
        );

        let runs = ArrayData::builder(DataType::Int32)
            .len(2)
            .add_buffer(vec![1_i32, 3].into())
            .build()
            .unwrap();
        let values = ArrayData::builder(DataType::Float32)
            .len(2)
            .add_buffer(vec![1.5_f32, 2.5].into())
            .build()
            .unwrap();

        (dt, vec![runs, values])
    }

    #[test]
    fn test_run_array_validates_last_run() {
        let (dt, children) = run_array_test_children();

        let err = ArrayDataBuilder::new(dt)
            .len(4)
            .offset(0)
            .child_data(children)
            .build()
            .unwrap_err();
        assert_eq!(
            err.to_string(),
            "Invalid argument error: The offset + length of array should be less or equal to last value in the run_ends array. The last value of run_ends array is 3 and offset + length of array is 4."
        );
    }

    #[test]
    fn test_run_array_empty_validates() {
        let (dt, children) = run_array_test_children();

        let data = ArrayDataBuilder::new(dt)
            .len(0)
            .offset(4)
            .child_data(children)
            .build()
            .unwrap();
        assert_eq!(data.len(), 0);
    }

    /// The parent's offset must be included when checking the last run end:
    /// offset 2 + length 2 = 4 exceeds the last run end of 3.
    #[test]
    fn test_run_array_validates_last_run_with_offset() {
        let (dt, children) = run_array_test_children();

        let err = ArrayDataBuilder::new(dt)
            .len(2)
            .offset(2)
            .child_data(children)
            .build()
            .unwrap_err();
        assert_eq!(
            err.to_string(),
            "Invalid argument error: The offset + length of array should be less or equal to last value in the run_ends array. The last value of run_ends array is 3 and offset + length of array is 4."
        );
    }

    /// A slice that stays within the last run end is valid:
    /// offset 1 + length 2 = 3 equals the last run end of 3.
    #[test]
    fn test_run_array_sliced_validates() {
        let (dt, children) = run_array_test_children();

        let data = ArrayDataBuilder::new(dt)
            .len(2)
            .offset(1)
            .child_data(children)
            .build()
            .unwrap();
        assert_eq!(data.len(), 2);
        assert_eq!(data.offset(), 1);
    }

    /// `validate_values` must report missing children rather than index out of bounds.
    #[test]
    #[cfg(not(feature = "force_validate"))]
    fn test_validate_values_rejects_missing_child_data() {
        let int32 = Box::new(DataType::Int32);
        let field = || Arc::new(Field::new("f", DataType::Int32, true));
        let data_types = [
            DataType::Dictionary(int32.clone(), int32.clone()),
            DataType::List(field()),
            DataType::LargeList(field()),
            DataType::RunEndEncoded(field(), field()),
        ];

        for data_type in data_types {
            let data = unsafe {
                ArrayData::builder(data_type.clone())
                    .len(1)
                    .build_unchecked()
            };
            let err = data.validate_values().expect_err("should get error");
            assert_eq!(
                err.to_string(),
                format!(
                    "Invalid argument error: {data_type} should contain at least 1 child data array(s), had 0"
                )
            );
        }
    }

    /// `validate_values` must report missing buffers rather than index out of bounds.
    #[test]
    #[cfg(not(feature = "force_validate"))]
    fn test_validate_values_rejects_missing_buffers() {
        // (data type, index of the first missing buffer)
        let cases = [
            (DataType::Utf8, 1),
            (DataType::LargeUtf8, 1),
            (DataType::Binary, 1),
            (DataType::LargeBinary, 1),
            (DataType::BinaryView, 0),
            (DataType::Utf8View, 0),
        ];

        for (data_type, missing) in cases {
            let data = unsafe {
                ArrayData::builder(data_type.clone())
                    .len(1)
                    .build_unchecked()
            };
            let err = data.validate_values().expect_err("should get error");
            assert_eq!(
                err.to_string(),
                format!(
                    "Invalid argument error: {data_type} should contain at least {} buffer(s), had 0",
                    missing + 1
                )
            );
        }
    }

    /// A dictionary whose keys buffer is too small must be reported, not asserted on.
    #[test]
    #[cfg(not(feature = "force_validate"))]
    fn test_validate_values_rejects_a_short_dictionary_keys_buffer() {
        let data_type = DataType::Dictionary(Box::new(DataType::Int32), Box::new(DataType::Int32));
        let dictionary = unsafe {
            ArrayData::builder(data_type)
                .len(4)
                .add_buffer(Buffer::from_slice_ref([1_i32, 0]))
                .add_child_data(valid_non_nullable_int32_array_data(2))
                .build_unchecked()
        };

        let err = dictionary.validate_values().expect_err("should get error");
        assert_eq!(
            err.to_string(),
            "Invalid argument error: Buffer 0 of Dictionary(Int32, Int32) isn't large enough. Expected 16 bytes got 8"
        );
    }

    #[test]
    fn should_fail_validation_when_having_map_field_type_is_not_struct() {
        let map_field = Field::new("key", DataType::Int32, false);

        let map_field_data = valid_non_nullable_int32_array_data(2);

        let results = test_both_builder_and_array_data(
            DataType::Map(map_field.into(), false),
            1,
            None,
            0,
            vec![
                OffsetBuffer::<i32>::from_lengths(vec![2])
                    .into_inner()
                    .into(),
            ],
            vec![map_field_data],
        );

        for result in results {
            let array_data_err = result.expect_err("should fail for non struct field");

            match array_data_err {
                ArrowError::InvalidArgumentError(msg) => {
                    assert_eq!(
                        msg,
                        "Map field should be a entries struct data type, got Int32 instead"
                    )
                }
                _ => panic!("unexpected error type {array_data_err}"),
            }
        }
    }

    #[test]
    fn should_fail_validation_when_having_map_entries_only_have_1_field() {
        let struct_data_type = DataType::Struct(Fields::from(vec![Field::new(
            Field::MAP_KEY_FIELD_DEFAULT_NAME,
            DataType::Int32,
            false,
        )]));

        let key_array_data = valid_non_nullable_int32_array_data(2);

        let struct_data = {
            let builder = ArrayDataBuilder::new(struct_data_type.clone())
                .len(2)
                .nulls(None)
                .child_data(vec![key_array_data]);

            builder.build().unwrap()
        };

        let results = test_both_builder_and_array_data(
            DataType::Map(
                Field::new(
                    Field::MAP_ENTRIES_FIELD_DEFAULT_NAME,
                    struct_data_type,
                    false,
                )
                .into(),
                false,
            ),
            1,
            None,
            0,
            vec![
                OffsetBuffer::<i32>::from_lengths(vec![2])
                    .into_inner()
                    .into(),
            ],
            vec![struct_data],
        );

        for result in results {
            let array_data_err = result.expect_err("should fail for nullable key");

            match array_data_err {
                ArrowError::InvalidArgumentError(msg) => {
                    assert_eq!(
                        msg,
                        "Map entries data type should be a struct containing 2 fields, got 1 fields"
                    )
                }
                _ => panic!("unexpected error type {array_data_err}"),
            }
        }
    }

    #[test]
    fn should_fail_validation_when_having_map_entries_have_3_fields() {
        let struct_data_type = DataType::Struct(Fields::from(vec![
            Field::new(Field::MAP_KEY_FIELD_DEFAULT_NAME, DataType::Int32, false),
            Field::new(Field::MAP_VALUE_FIELD_DEFAULT_NAME, DataType::Utf8, true),
            Field::new("other", DataType::Int32, true),
        ]));

        let key_array_data = valid_non_nullable_int32_array_data(2);

        let values_array_data = valid_string_array_data(2);

        let other_array_data = key_array_data.clone();

        let struct_data = {
            let builder = ArrayDataBuilder::new(struct_data_type.clone())
                .len(2)
                .nulls(None)
                .child_data(vec![key_array_data, values_array_data, other_array_data]);

            builder.build().unwrap()
        };

        let results = test_both_builder_and_array_data(
            DataType::Map(
                Field::new(
                    Field::MAP_ENTRIES_FIELD_DEFAULT_NAME,
                    struct_data_type,
                    false,
                )
                .into(),
                false,
            ),
            1,
            None,
            0,
            vec![
                OffsetBuffer::<i32>::from_lengths(vec![2])
                    .into_inner()
                    .into(),
            ],
            vec![struct_data],
        );

        for result in results {
            let array_data_err = result.expect_err("should fail for nullable key");

            match array_data_err {
                ArrowError::InvalidArgumentError(msg) => {
                    assert_eq!(
                        msg,
                        "Map entries data type should be a struct containing 2 fields, got 3 fields"
                    )
                }
                _ => panic!("unexpected error type {array_data_err}"),
            }
        }
    }

    #[test]
    fn should_fail_validation_when_having_nullable_map_keys() {
        let struct_data_type = DataType::Struct(Fields::from(vec![
            Field::new(Field::MAP_KEY_FIELD_DEFAULT_NAME, DataType::Int32, true),
            Field::new(Field::MAP_VALUE_FIELD_DEFAULT_NAME, DataType::Utf8, true),
        ]));

        let key_array_data = valid_non_nullable_int32_array_data(2);
        let values_array_data = valid_string_array_data(2);

        let struct_data = {
            let builder = ArrayDataBuilder::new(struct_data_type.clone())
                .len(2)
                .nulls(None)
                .child_data(vec![key_array_data, values_array_data]);

            builder.build().unwrap()
        };

        let results = test_both_builder_and_array_data(
            DataType::Map(
                Field::new(
                    Field::MAP_ENTRIES_FIELD_DEFAULT_NAME,
                    struct_data_type,
                    false,
                )
                .into(),
                false,
            ),
            1,
            None,
            0,
            vec![
                OffsetBuffer::<i32>::from_lengths(vec![2])
                    .into_inner()
                    .into(),
            ],
            vec![struct_data],
        );

        for result in results {
            let array_data_err = result.expect_err("should fail for nullable key");

            match array_data_err {
                ArrowError::InvalidArgumentError(msg) => {
                    assert_eq!(msg, "Map key field must not be nullable")
                }
                _ => panic!("unexpected error type {array_data_err}"),
            }
        }
    }

    #[test]
    fn should_fail_validation_when_having_entries_is_nullable_for_map() {
        let struct_data_type = DataType::Struct(Fields::from(vec![
            Field::new(Field::MAP_KEY_FIELD_DEFAULT_NAME, DataType::Int32, false),
            Field::new(Field::MAP_VALUE_FIELD_DEFAULT_NAME, DataType::Utf8, true),
        ]));

        let key_array_data = valid_non_nullable_int32_array_data(2);

        let values_array_data = valid_string_array_data(2);

        let struct_data = {
            let builder = ArrayDataBuilder::new(struct_data_type.clone())
                .len(2)
                .nulls(None)
                .child_data(vec![key_array_data, values_array_data]);

            builder.build().unwrap()
        };

        let results = test_both_builder_and_array_data(
            DataType::Map(
                Field::new(
                    Field::MAP_ENTRIES_FIELD_DEFAULT_NAME,
                    struct_data_type,
                    true,
                )
                .into(),
                false,
            ),
            1,
            None,
            0,
            vec![
                OffsetBuffer::<i32>::from_lengths(vec![2])
                    .into_inner()
                    .into(),
            ],
            vec![struct_data],
        );

        for result in results {
            let array_data_err = result.expect_err("should fail for nullable entries");

            match array_data_err {
                ArrowError::InvalidArgumentError(msg) => assert_eq!(
                    msg,
                    "The nullable should be set to false for the map entries field."
                ),
                _ => panic!("unexpected error type {array_data_err}"),
            }
        }
    }

    #[test]
    fn should_allow_to_create_map_from_data() {
        let struct_data_type = DataType::Struct(Fields::from(vec![
            Field::new(Field::MAP_KEY_FIELD_DEFAULT_NAME, DataType::Int32, false),
            Field::new(Field::MAP_VALUE_FIELD_DEFAULT_NAME, DataType::Utf8, true),
        ]));

        let key_array_data = valid_non_nullable_int32_array_data(2);
        let values_array_data = valid_string_array_data(2);

        let struct_data = {
            let builder = ArrayDataBuilder::new(struct_data_type.clone())
                .len(2)
                .nulls(None)
                .child_data(vec![key_array_data, values_array_data]);

            builder.build().unwrap()
        };

        let results = test_both_builder_and_array_data(
            DataType::Map(
                Field::new(
                    Field::MAP_ENTRIES_FIELD_DEFAULT_NAME,
                    struct_data_type,
                    false,
                )
                .into(),
                false,
            ),
            1,
            None,
            0,
            vec![
                OffsetBuffer::<i32>::from_lengths(vec![2])
                    .into_inner()
                    .into(),
            ],
            vec![struct_data],
        );

        for result in results {
            result.expect("should be able to create map ArrayData");
        }
    }

    fn valid_string_array_data(length: usize) -> ArrayData {
        let offsets = OffsetBuffer::<i32>::from_lengths(vec![0; length])
            .into_inner()
            .into_inner();
        let empty_bytes = Buffer::default();

        let builder = ArrayDataBuilder::new(DataType::Utf8)
            .len(length)
            .buffers(vec![offsets, empty_bytes])
            .nulls(None);

        builder.build().unwrap()
    }

    fn valid_non_nullable_int32_array_data(length: usize) -> ArrayData {
        let builder = ArrayDataBuilder::new(DataType::Int32)
            .len(length)
            .nulls(None)
            .buffers(vec![
                ScalarBuffer::<i32>::from(vec![1; length]).into_inner(),
            ]);

        builder.build().unwrap()
    }

    #[test]
    fn empty_and_null_map_array_should_pass_validation() {
        let dt = DataType::Map(
            Field::new(
                Field::MAP_ENTRIES_FIELD_DEFAULT_NAME,
                DataType::Struct(Fields::from(vec![
                    Field::new(Field::MAP_KEY_FIELD_DEFAULT_NAME, DataType::Int32, false),
                    Field::new(Field::MAP_VALUE_FIELD_DEFAULT_NAME, DataType::Utf8, true),
                ])),
                false,
            )
            .into(),
            false,
        );

        ArrayData::new_empty(&dt).validate_full().unwrap();
        ArrayData::new_null(&dt, 1).validate_full().unwrap();
    }

    #[test]
    fn null_buffer_offset_is_independent_of_data_offset() {
        // 100 values sliced down to the last 50, so the data has offset 50.
        let int_data = ArrayData::builder(DataType::UInt32)
            .offset(50)
            .len(50)
            .add_buffer(Buffer::from_vec(vec![0_u32; 100]))
            .build()
            .unwrap();
        int_data.validate().unwrap();

        // A null buffer that happens to share the data's offset.
        let nulls = NullBuffer::new(BooleanBuffer::from(vec![false; 100]).slice(0, 50));
        let with_sliced_nulls = int_data
            .clone()
            .into_builder()
            .nulls(Some(nulls))
            .build()
            .unwrap();
        with_sliced_nulls.validate().unwrap();

        // The same 50 nulls at offset 0. ArrayData::offset does not apply to the
        // null buffer, so this is just as valid and must not be rejected.
        let nulls = NullBuffer::new(BooleanBuffer::from(vec![false; 50]));
        let with_unsliced_nulls = int_data.into_builder().nulls(Some(nulls)).build().unwrap();
        with_unsliced_nulls.validate().unwrap();
        assert_eq!(with_unsliced_nulls.null_count(), 50);
    }

    fn test_both_builder_and_array_data(
        data_type: DataType,
        len: usize,
        null_bit_buffer: Option<Buffer>,
        offset: usize,
        buffers: Vec<Buffer>,
        child_data: Vec<ArrayData>,
    ) -> [Result<ArrayData, ArrowError>; 2] {
        let from_builder_res = ArrayData::builder(data_type.clone())
            .len(len)
            .add_buffers(buffers.clone())
            .null_bit_buffer(null_bit_buffer.clone())
            .offset(offset)
            .child_data(child_data.clone())
            .build();

        let from_try_new_res =
            ArrayData::try_new(data_type, len, null_bit_buffer, offset, buffers, child_data);

        [from_builder_res, from_try_new_res]
    }

    #[test]
    fn test_new_null_empty_union() {
        for mode in [UnionMode::Sparse, UnionMode::Dense] {
            let data_type = DataType::Union(UnionFields::empty(), mode);
            let data = ArrayData::new_null(&data_type, 0);
            data.validate_full()
                .expect("an empty union of length zero is valid");
            assert_eq!(data.len(), 0);
            assert!(data.child_data().is_empty());
        }
    }

    #[test]
    #[should_panic(expected = "cannot construct null data from an empty union")]
    fn test_new_null_empty_union_with_slots() {
        let data_type = DataType::Union(UnionFields::empty(), UnionMode::Dense);
        let _ = ArrayData::new_null(&data_type, 1);
    }

    #[test]
    fn test_union_validate_type_ids() {
        let dt = DataType::Union(
            UnionFields::try_new(
                vec![5],
                vec![Arc::new(Field::new("a", DataType::Int32, true))],
            )
            .unwrap(),
            UnionMode::Sparse,
        );

        let child = ArrayData::builder(DataType::Int32)
            .len(2)
            .add_buffer(make_i32_buffer(2))
            .build()
            .unwrap();

        let err = ArrayDataBuilder::new(dt)
            .len(2)
            .offset(0)
            .buffers(vec![vec![5_i8, 2].into()])
            .child_data(vec![child])
            .build()
            .unwrap_err();
        assert_eq!(
            err.to_string(),
            "Invalid argument error: Type id values must match one of the field type ids, got 2 at index 1"
        );
    }

    #[test]
    fn test_union_validate_dense_offsets() {
        let dt = DataType::Union(
            UnionFields::try_new(
                vec![5],
                vec![Arc::new(Field::new("a", DataType::Int32, true))],
            )
            .unwrap(),
            UnionMode::Dense,
        );

        let child = ArrayData::builder(DataType::Int32)
            .len(2)
            .add_buffer(make_i32_buffer(2))
            .build()
            .unwrap();

        let err = ArrayDataBuilder::new(dt.clone())
            .len(2)
            // 2nd offset '10' is out of bounds
            .buffers(vec![vec![5_i8, 5].into(), vec![0_i32, 10].into()])
            .child_data(vec![child.clone()])
            .build()
            .unwrap_err();
        assert_eq!(
            err.to_string(),
            "Invalid argument error: Offsets must be non-negative and within the length of the child array"
        );

        let err = ArrayDataBuilder::new(dt)
            .len(2)
            // 2nd offset is negative
            .buffers(vec![vec![5_i8, 5].into(), vec![0_i32, -1].into()])
            .child_data(vec![child])
            .build()
            .unwrap_err();
        assert_eq!(
            err.to_string(),
            "Invalid argument error: Offsets must be non-negative and within the length of the child array"
        );
    }
}