arrow_row/

lib.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.

//! A comparable row-oriented representation of a collection of [`Array`].
//!
//! [`Row`]s are [normalized for sorting], and can therefore be very efficiently [compared],
//! using [`memcmp`] under the hood, or used in [non-comparison sorts] such as [radix sort].
//! This makes the row format ideal for implementing efficient multi-column sorting,
//! grouping, aggregation, windowing and more, as described in more detail
//! [in this blog post](https://arrow.apache.org/blog/2022/11/07/multi-column-sorts-in-arrow-rust-part-1/).
//!
//! For example, given three input [`Array`], [`RowConverter`] creates byte
//! sequences that [compare] the same as when using [`lexsort`].
//!
//! ```text
//!    ┌─────┐   ┌─────┐   ┌─────┐
//!    │     │   │     │   │     │
//!    ├─────┤ ┌ ┼─────┼ ─ ┼─────┼ ┐              ┏━━━━━━━━━━━━━┓
//!    │     │   │     │   │     │  ─────────────▶┃             ┃
//!    ├─────┤ └ ┼─────┼ ─ ┼─────┼ ┘              ┗━━━━━━━━━━━━━┛
//!    │     │   │     │   │     │
//!    └─────┘   └─────┘   └─────┘
//!                ...
//!    ┌─────┐ ┌ ┬─────┬ ─ ┬─────┬ ┐              ┏━━━━━━━━┓
//!    │     │   │     │   │     │  ─────────────▶┃        ┃
//!    └─────┘ └ ┴─────┴ ─ ┴─────┴ ┘              ┗━━━━━━━━┛
//!     UInt64      Utf8     F64
//!
//!           Input Arrays                          Row Format
//!     (Columns)
//! ```
//!
//! _[`Rows`] must be generated by the same [`RowConverter`] for the comparison
//! to be meaningful._
//!
//! # Basic Example
//! ```
//! # use std::sync::Arc;
//! # use arrow_row::{RowConverter, SortField};
//! # use arrow_array::{ArrayRef, Int32Array, StringArray};
//! # use arrow_array::cast::{AsArray, as_string_array};
//! # use arrow_array::types::Int32Type;
//! # use arrow_schema::DataType;
//!
//! let a1 = Arc::new(Int32Array::from_iter_values([-1, -1, 0, 3, 3])) as ArrayRef;
//! let a2 = Arc::new(StringArray::from_iter_values(["a", "b", "c", "d", "d"])) as ArrayRef;
//! let arrays = vec![a1, a2];
//!
//! // Convert arrays to rows
//! let converter = RowConverter::new(vec![
//!     SortField::new(DataType::Int32),
//!     SortField::new(DataType::Utf8),
//! ]).unwrap();
//! let rows = converter.convert_columns(&arrays).unwrap();
//!
//! // Compare rows
//! for i in 0..4 {
//!     assert!(rows.row(i) <= rows.row(i + 1));
//! }
//! assert_eq!(rows.row(3), rows.row(4));
//!
//! // Convert rows back to arrays
//! let converted = converter.convert_rows(&rows).unwrap();
//! assert_eq!(arrays, converted);
//!
//! // Compare rows from different arrays
//! let a1 = Arc::new(Int32Array::from_iter_values([3, 4])) as ArrayRef;
//! let a2 = Arc::new(StringArray::from_iter_values(["e", "f"])) as ArrayRef;
//! let arrays = vec![a1, a2];
//! let rows2 = converter.convert_columns(&arrays).unwrap();
//!
//! assert!(rows.row(4) < rows2.row(0));
//! assert!(rows.row(4) < rows2.row(1));
//!
//! // Convert selection of rows back to arrays
//! let selection = [rows.row(0), rows2.row(1), rows.row(2), rows2.row(0)];
//! let converted = converter.convert_rows(selection).unwrap();
//! let c1 = converted[0].as_primitive::<Int32Type>();
//! assert_eq!(c1.values(), &[-1, 4, 0, 3]);
//!
//! let c2 = converted[1].as_string::<i32>();
//! let c2_values: Vec<_> = c2.iter().flatten().collect();
//! assert_eq!(&c2_values, &["a", "f", "c", "e"]);
//! ```
//!
//! # Lexicographic Sorts (lexsort)
//!
//! The row format can also be used to implement a fast multi-column / lexicographic sort
//!
//! ```
//! # use arrow_row::{RowConverter, SortField};
//! # use arrow_array::{ArrayRef, UInt32Array};
//! fn lexsort_to_indices(arrays: &[ArrayRef]) -> UInt32Array {
//!     let fields = arrays
//!         .iter()
//!         .map(|a| SortField::new(a.data_type().clone()))
//!         .collect();
//!     let converter = RowConverter::new(fields).unwrap();
//!     let rows = converter.convert_columns(arrays).unwrap();
//!     let mut sort: Vec<_> = rows.iter().enumerate().collect();
//!     sort.sort_unstable_by(|(_, a), (_, b)| a.cmp(b));
//!     UInt32Array::from_iter_values(sort.iter().map(|(i, _)| *i as u32))
//! }
//! ```
//!
//! # Flattening Dictionaries
//!
//! For performance reasons, dictionary arrays are flattened ("hydrated") to their
//! underlying values during row conversion. See [the issue] for more details.
//!
//! This means that the arrays that come out of [`RowConverter::convert_rows`]
//! may not have the same data types as the input arrays. For example, encoding
//! a `Dictionary<Int8, Utf8>` and then will come out as a `Utf8` array.
//!
//! ```
//! # use arrow_array::{Array, ArrayRef, DictionaryArray};
//! # use arrow_array::types::Int8Type;
//! # use arrow_row::{RowConverter, SortField};
//! # use arrow_schema::DataType;
//! # use std::sync::Arc;
//! // Input is a Dictionary array
//! let dict: DictionaryArray::<Int8Type> = ["a", "b", "c", "a", "b"].into_iter().collect();
//! let sort_fields = vec![SortField::new(dict.data_type().clone())];
//! let arrays = vec![Arc::new(dict) as ArrayRef];
//! let converter = RowConverter::new(sort_fields).unwrap();
//! // Convert to rows
//! let rows = converter.convert_columns(&arrays).unwrap();
//! let converted = converter.convert_rows(&rows).unwrap();
//! // result was a Utf8 array, not a Dictionary array
//! assert_eq!(converted[0].data_type(), &DataType::Utf8);
//! ```
//!
//! [non-comparison sorts]: https://en.wikipedia.org/wiki/Sorting_algorithm#Non-comparison_sorts
//! [radix sort]: https://en.wikipedia.org/wiki/Radix_sort
//! [normalized for sorting]: http://wwwlgis.informatik.uni-kl.de/archiv/wwwdvs.informatik.uni-kl.de/courses/DBSREAL/SS2005/Vorlesungsunterlagen/Implementing_Sorting.pdf
//! [`memcmp`]: https://www.man7.org/linux/man-pages/man3/memcmp.3.html
//! [`lexsort`]: https://docs.rs/arrow-ord/latest/arrow_ord/sort/fn.lexsort.html
//! [compared]: PartialOrd
//! [compare]: PartialOrd
//! [the issue]: https://github.com/apache/arrow-rs/issues/4811

#![doc(
    html_logo_url = "https://arrow.apache.org/img/arrow-logo_chevrons_black-txt_white-bg.svg",
    html_favicon_url = "https://arrow.apache.org/img/arrow-logo_chevrons_black-txt_transparent-bg.svg"
)]
#![cfg_attr(docsrs, feature(doc_cfg))]
#![warn(missing_docs)]
use std::cmp::Ordering;
use std::hash::{Hash, Hasher};
use std::iter::Map;
use std::slice::Windows;
use std::sync::Arc;

use arrow_array::cast::*;
use arrow_array::types::{ArrowDictionaryKeyType, ByteArrayType, ByteViewType};
use arrow_array::*;
use arrow_buffer::{ArrowNativeType, Buffer, OffsetBuffer, ScalarBuffer};
use arrow_schema::*;
use variable::{decode_binary_view, decode_string_view};

use crate::fixed::{decode_bool, decode_fixed_size_binary, decode_primitive};
use crate::list::{compute_lengths_fixed_size_list, encode_fixed_size_list};
use crate::variable::{decode_binary, decode_string};
use arrow_array::types::{Int16Type, Int32Type, Int64Type};

mod fixed;
mod list;
mod run;
mod variable;

/// Converts [`ArrayRef`] columns into a [row-oriented](self) format.
///
/// *Note: The encoding of the row format may change from release to release.*
///
/// ## Overview
///
/// The row format is a variable length byte sequence created by
/// concatenating the encoded form of each column. The encoding for
/// each column depends on its datatype (and sort options).
///
/// The encoding is carefully designed in such a way that escaping is
/// unnecessary: it is never ambiguous as to whether a byte is part of
/// a sentinel (e.g. null) or a value.
///
/// ## Unsigned Integer Encoding
///
/// A null integer is encoded as a `0_u8`, followed by a zero-ed number of bytes corresponding
/// to the integer's length.
///
/// A valid integer is encoded as `1_u8`, followed by the big-endian representation of the
/// integer.
///
/// ```text
///               ┌──┬──┬──┬──┐      ┌──┬──┬──┬──┬──┐
///    3          │03│00│00│00│      │01│00│00│00│03│
///               └──┴──┴──┴──┘      └──┴──┴──┴──┴──┘
///               ┌──┬──┬──┬──┐      ┌──┬──┬──┬──┬──┐
///   258         │02│01│00│00│      │01│00│00│01│02│
///               └──┴──┴──┴──┘      └──┴──┴──┴──┴──┘
///               ┌──┬──┬──┬──┐      ┌──┬──┬──┬──┬──┐
///  23423        │7F│5B│00│00│      │01│00│00│5B│7F│
///               └──┴──┴──┴──┘      └──┴──┴──┴──┴──┘
///               ┌──┬──┬──┬──┐      ┌──┬──┬──┬──┬──┐
///  NULL         │??│??│??│??│      │00│00│00│00│00│
///               └──┴──┴──┴──┘      └──┴──┴──┴──┴──┘
///
///              32-bit (4 bytes)        Row Format
///  Value        Little Endian
/// ```
///
/// ## Signed Integer Encoding
///
/// Signed integers have their most significant sign bit flipped, and are then encoded in the
/// same manner as an unsigned integer.
///
/// ```text
///        ┌──┬──┬──┬──┐       ┌──┬──┬──┬──┐       ┌──┬──┬──┬──┬──┐
///     5  │05│00│00│00│       │05│00│00│80│       │01│80│00│00│05│
///        └──┴──┴──┴──┘       └──┴──┴──┴──┘       └──┴──┴──┴──┴──┘
///        ┌──┬──┬──┬──┐       ┌──┬──┬──┬──┐       ┌──┬──┬──┬──┬──┐
///    -5  │FB│FF│FF│FF│       │FB│FF│FF│7F│       │01│7F│FF│FF│FB│
///        └──┴──┴──┴──┘       └──┴──┴──┴──┘       └──┴──┴──┴──┴──┘
///
///  Value  32-bit (4 bytes)    High bit flipped      Row Format
///          Little Endian
/// ```
///
/// ## Float Encoding
///
/// Floats are converted from IEEE 754 representation to a signed integer representation
/// by flipping all bar the sign bit if they are negative.
///
/// They are then encoded in the same manner as a signed integer.
///
/// ## Fixed Length Bytes Encoding
///
/// Fixed length bytes are encoded in the same fashion as primitive types above.
///
/// For a fixed length array of length `n`:
///
/// A null is encoded as `0_u8` null sentinel followed by `n` `0_u8` bytes
///
/// A valid value is encoded as `1_u8` followed by the value bytes
///
/// ## Variable Length Bytes (including Strings) Encoding
///
/// A null is encoded as a `0_u8`.
///
/// An empty byte array is encoded as `1_u8`.
///
/// A non-null, non-empty byte array is encoded as `2_u8` followed by the byte array
/// encoded using a block based scheme described below.
///
/// The byte array is broken up into fixed-width blocks, each block is written in turn
/// to the output, followed by `0xFF_u8`. The final block is padded to 32-bytes
/// with `0_u8` and written to the output, followed by the un-padded length in bytes
/// of this final block as a `u8`. The first 4 blocks have a length of 8, with subsequent
/// blocks using a length of 32, this is to reduce space amplification for small strings.
///
/// Note the following example encodings use a block size of 4 bytes for brevity:
///
/// ```text
///                       ┌───┬───┬───┬───┬───┬───┐
///  "MEEP"               │02 │'M'│'E'│'E'│'P'│04 │
///                       └───┴───┴───┴───┴───┴───┘
///
///                       ┌───┐
///  ""                   │01 |
///                       └───┘
///
///  NULL                 ┌───┐
///                       │00 │
///                       └───┘
///
/// "Defenestration"      ┌───┬───┬───┬───┬───┬───┐
///                       │02 │'D'│'e'│'f'│'e'│FF │
///                       └───┼───┼───┼───┼───┼───┤
///                           │'n'│'e'│'s'│'t'│FF │
///                           ├───┼───┼───┼───┼───┤
///                           │'r'│'a'│'t'│'r'│FF │
///                           ├───┼───┼───┼───┼───┤
///                           │'a'│'t'│'i'│'o'│FF │
///                           ├───┼───┼───┼───┼───┤
///                           │'n'│00 │00 │00 │01 │
///                           └───┴───┴───┴───┴───┘
/// ```
///
/// This approach is loosely inspired by [COBS] encoding, and chosen over more traditional
/// [byte stuffing] as it is more amenable to vectorisation, in particular AVX-256.
///
/// ## Dictionary Encoding
///
/// Dictionary encoded arrays are hydrated to their underlying values
///
/// ## REE Encoding
///
/// REE (Run End Encoding) arrays, A form of Run Length Encoding, are hydrated to their underlying values.
///
/// ## Struct Encoding
///
/// A null is encoded as a `0_u8`.
///
/// A valid value is encoded as `1_u8` followed by the row encoding of each child.
///
/// This encoding effectively flattens the schema in a depth-first fashion.
///
/// For example
///
/// ```text
/// ┌───────┬────────────────────────┬───────┐
/// │ Int32 │ Struct[Int32, Float32] │ Int32 │
/// └───────┴────────────────────────┴───────┘
/// ```
///
/// Is encoded as
///
/// ```text
/// ┌───────┬───────────────┬───────┬─────────┬───────┐
/// │ Int32 │ Null Sentinel │ Int32 │ Float32 │ Int32 │
/// └───────┴───────────────┴───────┴─────────┴───────┘
/// ```
///
/// ## List and Map Encoding
///
/// Lists are encoded by first encoding all child elements to the row format.
///
/// The Map encoding is the same with the only difference being that the child elements are key-value pairs.
///
/// A list/map value is then encoded as the concatenation of each of the child elements,
/// separately encoded using the variable length encoding described above, followed
/// by the variable length encoding of an empty byte array.
///
/// For example given:
///
/// ```text
/// [1_u8, 2_u8, 3_u8]
/// [1_u8, null]
/// []
/// null
/// ```
///
/// The elements would be converted to:
///
/// ```text
///     ┌──┬──┐     ┌──┬──┐     ┌──┬──┐     ┌──┬──┐        ┌──┬──┐
///  1  │01│01│  2  │01│02│  3  │01│03│  1  │01│01│  null  │00│00│
///     └──┴──┘     └──┴──┘     └──┴──┘     └──┴──┘        └──┴──┘
///```
///
/// Which would be encoded as
///
/// ```text
///                         ┌──┬──┬──┬──┬──┬──┬──┬──┬──┬──┬──┬──┬──┬──┬──┬──┬──┬──┬──┐
///  [1_u8, 2_u8, 3_u8]     │02│01│01│00│00│02│02│01│02│00│00│02│02│01│03│00│00│02│01│
///                         └──┴──┴──┴──┴──┴──┴──┴──┴──┴──┴──┴──┴──┴──┴──┴──┴──┴──┴──┘
///                          └──── 1_u8 ────┘   └──── 2_u8 ────┘  └──── 3_u8 ────┘
///
///                         ┌──┬──┬──┬──┬──┬──┬──┬──┬──┬──┬──┬──┬──┐
///  [1_u8, null]           │02│01│01│00│00│02│02│00│00│00│00│02│01│
///                         └──┴──┴──┴──┴──┴──┴──┴──┴──┴──┴──┴──┴──┘
///                          └──── 1_u8 ────┘   └──── null ────┘
///
///```
///
/// With `[]` represented by an empty byte array, and `null` a null byte array.
///
/// ### Map Equality
///
/// **Note:** Maps with different order of keys are not equal in row format
/// i.e. row_format({"hello": 1, "world": 2}) != row_format({"world": 2, "hello": 1})
///
/// ```
/// # use std::sync::Arc;
/// # use arrow_row::{RowConverter, SortField};
/// # use arrow_array::{Array, Int32Array, StringArray, MapArray};
///
/// // [{ "hello": 1, "world": 2 }, { "hey": 3, "you": 4 }]
/// let map_1 = MapArray::from_vec_of_maps::<StringArray, Int32Array, _, _>(vec![
///   Some(vec![("hello", Some(1)), ("world", Some(2))]),
///   Some(vec![("hey", Some(3)), ("you", Some(4))]),
/// ], false);
/// // [{ "world": 2, "hello": 1 }, { "hey": 3, "you": 4 }]
/// let map_2 = MapArray::from_vec_of_maps::<StringArray, Int32Array, _, _>(vec![
///   Some(vec![("world", Some(2)), ("hello", Some(1))]),
///   Some(vec![("hey", Some(3)), ("you", Some(4))]),
/// ], false);
///
/// let converter = RowConverter::new(vec![SortField::new(map_1.data_type().clone())]).unwrap();
///
/// let map_1_rows = converter.convert_columns(&[Arc::new(map_1)]).unwrap();
/// let map_2_rows = converter.convert_columns(&[Arc::new(map_2)]).unwrap();
///
/// // Attention! these maps are NOT equal since the order of the keys are different
/// assert_ne!(map_1_rows.row(0), map_2_rows.row(0));
///
/// // These maps ARE equal since the order and the content of the keys and values are the same
/// assert_eq!(map_1_rows.row(1), map_2_rows.row(1));
/// ```
///
/// If you DO want to treat maps with different order of keys as the same in row format you should canonicalize them first.
///
/// ## Fixed Size List Encoding
///
/// Fixed Size Lists are encoded by first encoding all child elements to the row format.
///
/// A non-null list value is then encoded as 0x01 followed by the concatenation of each
/// of the child elements. A null list value is encoded as a null marker.
///
/// For example given:
///
/// ```text
/// [1_u8, 2_u8]
/// [3_u8, null]
/// null
/// ```
///
/// The elements would be converted to:
///
/// ```text
///     ┌──┬──┐     ┌──┬──┐     ┌──┬──┐        ┌──┬──┐
///  1  │01│01│  2  │01│02│  3  │01│03│  null  │00│00│
///     └──┴──┘     └──┴──┘     └──┴──┘        └──┴──┘
///```
///
/// Which would be encoded as
///
/// ```text
///                 ┌──┬──┬──┬──┬──┐
///  [1_u8, 2_u8]   │01│01│01│01│02│
///                 └──┴──┴──┴──┴──┘
///                     └ 1 ┘ └ 2 ┘
///                 ┌──┬──┬──┬──┬──┐
///  [3_u8, null]   │01│01│03│00│00│
///                 └──┴──┴──┴──┴──┘
///                     └ 1 ┘ └null┘
///                 ┌──┐
///  null           │00│
///                 └──┘
///
///```
///
/// ## ListView Encoding
///
/// ListView arrays differ from List arrays in their representation: instead of using
/// consecutive offset pairs to define each list, ListView uses explicit offset and size
/// pairs for each element. This allows ListView elements to reference arbitrary (potentially
/// overlapping) regions of the child array.
///
/// Despite this structural difference, ListView uses the **same row encoding as List**.
/// Each list value is encoded as the concatenation of its child elements (each separately
/// variable-length encoded), followed by a variable-length encoded empty byte array terminator.
///
/// **Important**: When a ListView is decoded back from row format, it is still a
/// ListView, but any child element sharing that may have existed in the original
/// (where multiple list entries could reference overlapping regions of the child
/// array) is **not preserved** - each list's children are decoded independently
/// with sequential offsets.
///
/// For example, given a ListView with offset/size pairs:
///
/// ```text
/// offsets: [0, 1, 0]
/// sizes:   [2, 2, 0]
/// values:  [1_u8, 2_u8, 3_u8]
///
/// Resulting lists:
/// [1_u8, 2_u8]  (offset=0, size=2 -> values[0..2])
/// [2_u8, 3_u8]  (offset=1, size=2 -> values[1..3])
/// []            (offset=0, size=0 -> empty)
/// ```
///
/// The elements would be encoded identically to List encoding:
///
/// ```text
///                         ┌──┬──┬──┬──┬──┬──┬──┬──┬──┬──┬──┬──┬──┐
///  [1_u8, 2_u8]           │02│01│01│00│00│02│02│01│02│00│00│02│01│
///                         └──┴──┴──┴──┴──┴──┴──┴──┴──┴──┴──┴──┴──┘
///                          └──── 1_u8 ────┘   └──── 2_u8 ────┘
///
///                         ┌──┬──┬──┬──┬──┬──┬──┬──┬──┬──┬──┬──┬──┐
///  [2_u8, 3_u8]           │02│01│02│00│00│02│02│01│03│00│00│02│01│
///                         └──┴──┴──┴──┴──┴──┴──┴──┴──┴──┴──┴──┴──┘
///                          └──── 2_u8 ────┘   └──── 3_u8 ────┘
///
///```
///
/// Note that element `2_u8` appears in both encoded rows, even though it was shared
/// in the original ListView, and `[]` is represented by an empty byte array.
///
/// ## Union Encoding
///
/// A union value is encoded as a single type-id byte followed by the row encoding of the selected child value.
/// The type-id byte is always present; union arrays have no top-level null marker, so nulls are represented by the child encoding.
///
/// For example, given a union of Int32 (type_id = 0) and Utf8 (type_id = 1):
///
/// ```text
///                           ┌──┬──────────────┐
///  3                        │00│01│80│00│00│03│
///                           └──┴──────────────┘
///                            │  └─ signed integer encoding (non-null)
///                            └──── type_id
///
///                           ┌──┬────────────────────────────────┐
/// "abc"                     │01│02│'a'│'b'│'c'│00│00│00│00│00│03│
///                           └──┴────────────────────────────────┘
///                            │  └─ string encoding (non-null)
///                            └──── type_id
///
///                           ┌──┬──────────────┐
/// null Int32                │00│00│00│00│00│00│
///                           └──┴──────────────┘
///                            │  └─ signed integer encoding (null)
///                            └──── type_id
///
///                           ┌──┬──┐
/// null Utf8                 │01│00│
///                           └──┴──┘
///                            │  └─ string encoding (null)
///                            └──── type_id
/// ```
///
/// See [`UnionArray`] for more details on union types.
///
/// # Ordering
///
/// ## Float Ordering
///
/// Floats are totally ordered in accordance to the `totalOrder` predicate as defined
/// in the IEEE 754 (2008 revision) floating point standard.
///
/// The ordering established by this does not always agree with the
/// [`PartialOrd`] and [`PartialEq`] implementations of `f32`. For example,
/// they consider negative and positive zero equal, while this does not
///
/// ## Null Ordering
///
/// The encoding described above will order nulls first, this can be inverted by representing
/// nulls as `0xFF_u8` instead of `0_u8`
///
/// ## Union Ordering
///
/// Values of the same type are ordered according to the ordering of that type.
/// Values of different types are ordered by their type id.
/// The type_id is negated when descending order is specified.
///
/// ## Reverse Column Ordering
///
/// The order of a given column can be reversed by negating the encoded bytes of non-null values
///
/// [COBS]: https://en.wikipedia.org/wiki/Consistent_Overhead_Byte_Stuffing
/// [byte stuffing]: https://en.wikipedia.org/wiki/High-Level_Data_Link_Control#Asynchronous_framing
#[derive(Debug)]
pub struct RowConverter {
    fields: Arc<[SortField]>,
    /// State for codecs
    codecs: Vec<Codec>,
}

#[derive(Debug)]
enum Codec {
    /// No additional codec state is necessary
    Stateless,
    /// A row converter for the dictionary values
    /// and the encoding of a row containing only nulls
    Dictionary(RowConverter, OwnedRow),
    /// A row converter for the child fields
    /// and the encoding of a row containing only nulls
    Struct(RowConverter, OwnedRow),
    /// A row converter for the child field
    List(RowConverter),
    /// A row converter for the entries in map (keys and values and not the struct)
    Map(RowConverter),
    /// A row converter for the values array of a run-end encoded array
    RunEndEncoded(RowConverter),
    /// Row converters for each union field (indexed by field position)
    /// the type_ids for each field position, and the encoding of null rows for each field
    Union(Vec<RowConverter>, Vec<i8>, Vec<OwnedRow>),
}

/// Computes the minimum offset and maximum end (offset + size) for a ListView array.
/// Returns (min_offset, max_end) which can be used to slice the values array.
fn compute_list_view_bounds<O: OffsetSizeTrait>(array: &GenericListViewArray<O>) -> (usize, usize) {
    if array.is_empty() {
        return (0, 0);
    }

    let offsets = array.value_offsets();
    let sizes = array.value_sizes();
    let values_len = array.values().len();

    let mut min_offset = usize::MAX;
    let mut max_end = 0usize;

    for i in 0..array.len() {
        let offset = offsets[i].as_usize();
        let size = sizes[i].as_usize();
        let end = offset + size;

        if size > 0 {
            min_offset = min_offset.min(offset);
            max_end = max_end.max(end);
        }

        // Early exit if we've found the full range of the values array. This is possible with
        // ListViews since offsets and sizes are arbitrary and the full range can be covered early
        // in the iteration contrary to regular Lists.
        if min_offset == 0 && max_end == values_len {
            break;
        }
    }

    if min_offset == usize::MAX {
        // All lists are empty
        (0, 0)
    } else {
        (min_offset, max_end)
    }
}

impl Codec {
    fn new(sort_field: &SortField) -> Result<Self, ArrowError> {
        match &sort_field.data_type {
            DataType::Dictionary(_, values) => {
                let sort_field =
                    SortField::new_with_options(values.as_ref().clone(), sort_field.options);

                let converter = RowConverter::new(vec![sort_field])?;
                let null_array = new_null_array(values.as_ref(), 1);
                let nulls = converter.convert_columns(&[null_array])?;

                let owned = OwnedRow {
                    data: nulls.buffer.into(),
                    config: nulls.config,
                };
                Ok(Self::Dictionary(converter, owned))
            }
            DataType::RunEndEncoded(_, values) => {
                // Similar to List implementation
                let options = SortOptions {
                    descending: false,
                    nulls_first: sort_field.options.nulls_first != sort_field.options.descending,
                };

                let field = SortField::new_with_options(values.data_type().clone(), options);
                let converter = RowConverter::new(vec![field])?;
                Ok(Self::RunEndEncoded(converter))
            }
            d if !d.is_nested() => Ok(Self::Stateless),
            DataType::List(f)
            | DataType::LargeList(f)
            | DataType::ListView(f)
            | DataType::LargeListView(f) => {
                // The encoded contents will be inverted if descending is set to true
                // As such we set `descending` to false and negate nulls first if it
                // it set to true
                let options = SortOptions {
                    descending: false,
                    nulls_first: sort_field.options.nulls_first != sort_field.options.descending,
                };

                let field = SortField::new_with_options(f.data_type().clone(), options);
                let converter = RowConverter::new(vec![field])?;
                Ok(Self::List(converter))
            }
            DataType::Map(f, _) => {
                // The encoded contents will be inverted if `descending` is set to true
                // As such we set `descending` to false and negate nulls first if it
                // `descending` set to true.
                let options = SortOptions {
                    descending: false,
                    nulls_first: sort_field.options.nulls_first != sort_field.options.descending,
                };

                let DataType::Struct(fields) = f.data_type() else {
                    return Err(ArrowError::InvalidArgumentError(format!(
                        "expected struct field in map, got {:?}",
                        f.data_type()
                    )));
                };

                // For Map type we unwrap the intermediate struct type to avoid going through Struct codec to improve performance
                let fields = fields
                    .iter()
                    .map(|struct_field| {
                        SortField::new_with_options(struct_field.data_type().clone(), options)
                    })
                    .collect::<Vec<_>>();
                assert_eq!(fields.len(), 2);
                let converter = RowConverter::new(fields)?;
                Ok(Self::Map(converter))
            }
            DataType::FixedSizeList(f, _) => {
                let field = SortField::new_with_options(f.data_type().clone(), sort_field.options);
                let converter = RowConverter::new(vec![field])?;
                Ok(Self::List(converter))
            }
            DataType::Struct(f) => {
                let sort_fields = f
                    .iter()
                    .map(|x| SortField::new_with_options(x.data_type().clone(), sort_field.options))
                    .collect();

                let converter = RowConverter::new(sort_fields)?;
                let nulls: Vec<_> = f.iter().map(|x| new_null_array(x.data_type(), 1)).collect();

                let nulls = converter.convert_columns(&nulls)?;
                let owned = OwnedRow {
                    data: nulls.buffer.into(),
                    config: nulls.config,
                };

                Ok(Self::Struct(converter, owned))
            }
            DataType::Union(fields, _mode) => {
                // similar to dictionaries and lists, we set descending to false and negate nulls_first
                // since the encoded contents will be inverted if descending is set
                let options = SortOptions {
                    descending: false,
                    nulls_first: sort_field.options.nulls_first != sort_field.options.descending,
                };

                let mut converters = Vec::with_capacity(fields.len());
                let mut type_ids = Vec::with_capacity(fields.len());
                let mut null_rows = Vec::with_capacity(fields.len());

                for (type_id, field) in fields.iter() {
                    let sort_field =
                        SortField::new_with_options(field.data_type().clone(), options);
                    let converter = RowConverter::new(vec![sort_field])?;

                    let null_array = new_null_array(field.data_type(), 1);
                    let nulls = converter.convert_columns(&[null_array])?;
                    let owned = OwnedRow {
                        data: nulls.buffer.into(),
                        config: nulls.config,
                    };

                    converters.push(converter);
                    type_ids.push(type_id);
                    null_rows.push(owned);
                }

                Ok(Self::Union(converters, type_ids, null_rows))
            }
            _ => Err(ArrowError::NotYetImplemented(format!(
                "not yet implemented: {:?}",
                sort_field.data_type
            ))),
        }
    }

    fn encoder(&self, array: &dyn Array) -> Result<Encoder<'_>, ArrowError> {
        match self {
            Codec::Stateless => Ok(Encoder::Stateless),
            Codec::Dictionary(converter, nulls) => {
                let values = array.as_any_dictionary().values().clone();
                let rows = converter.convert_columns(&[values])?;
                Ok(Encoder::Dictionary(rows, nulls.row()))
            }
            Codec::Struct(converter, null) => {
                let v = as_struct_array(array);
                let rows = converter.convert_columns(v.columns())?;
                Ok(Encoder::Struct(rows, null.row()))
            }
            Codec::List(converter) => {
                let values = match array.data_type() {
                    DataType::List(_) => {
                        let list_array = as_list_array(array);
                        let first_offset = list_array.offsets()[0] as usize;
                        let last_offset =
                            list_array.offsets()[list_array.offsets().len() - 1] as usize;

                        // values can include more data than referenced in the ListArray, only encode
                        // the referenced values.
                        list_array
                            .values()
                            .slice(first_offset, last_offset - first_offset)
                    }
                    DataType::LargeList(_) => {
                        let list_array = as_large_list_array(array);

                        let first_offset = list_array.offsets()[0] as usize;
                        let last_offset =
                            list_array.offsets()[list_array.offsets().len() - 1] as usize;

                        // values can include more data than referenced in the LargeListArray, only encode
                        // the referenced values.
                        list_array
                            .values()
                            .slice(first_offset, last_offset - first_offset)
                    }
                    DataType::ListView(_) => {
                        let list_view_array = array.as_list_view::<i32>();
                        let (min_offset, max_end) = compute_list_view_bounds(list_view_array);
                        list_view_array
                            .values()
                            .slice(min_offset, max_end - min_offset)
                    }
                    DataType::LargeListView(_) => {
                        let list_view_array = array.as_list_view::<i64>();
                        let (min_offset, max_end) = compute_list_view_bounds(list_view_array);
                        list_view_array
                            .values()
                            .slice(min_offset, max_end - min_offset)
                    }
                    DataType::FixedSizeList(_, _) => {
                        as_fixed_size_list_array(array).values().clone()
                    }
                    _ => unreachable!(),
                };
                let rows = converter.convert_columns(&[values])?;
                Ok(Encoder::List(rows))
            }
            Codec::Map(converter) => {
                let map_array = as_map_array(array);

                let first_offset = map_array.offsets()[0] as usize;
                let last_offset = map_array.offsets()[map_array.offsets().len() - 1] as usize;

                // entries can include more data than referenced in the MapArray, only encode
                // the referenced entries.
                let sliced_entries = map_array
                    .entries()
                    .slice(first_offset, last_offset - first_offset);

                // the converter for the map is the keys and values and not the wrapping struct
                let rows = converter.convert_columns(sliced_entries.columns())?;
                Ok(Encoder::Map(rows))
            }
            Codec::RunEndEncoded(converter) => {
                let values = match array.data_type() {
                    DataType::RunEndEncoded(r, _) => match r.data_type() {
                        DataType::Int16 => array.as_run::<Int16Type>().values_slice(),
                        DataType::Int32 => array.as_run::<Int32Type>().values_slice(),
                        DataType::Int64 => array.as_run::<Int64Type>().values_slice(),
                        _ => unreachable!("Unsupported run end index type: {r:?}"),
                    },
                    _ => unreachable!(),
                };
                let rows = converter.convert_columns(std::slice::from_ref(&values))?;
                Ok(Encoder::RunEndEncoded(rows))
            }
            Codec::Union(converters, field_to_type_ids, _) => {
                let union_array = array
                    .as_any()
                    .downcast_ref::<UnionArray>()
                    .expect("expected Union array");

                let type_ids = union_array.type_ids().clone();
                let offsets = union_array.offsets().cloned();

                let mut child_rows = Vec::with_capacity(converters.len());
                for (field_idx, converter) in converters.iter().enumerate() {
                    let type_id = field_to_type_ids[field_idx];
                    let child_array = union_array.child(type_id);
                    let rows = converter.convert_columns(std::slice::from_ref(child_array))?;
                    child_rows.push(rows);
                }

                Ok(Encoder::Union {
                    child_rows,
                    field_to_type_ids: field_to_type_ids.clone(),
                    type_ids,
                    offsets,
                })
            }
        }
    }

    fn size(&self) -> usize {
        match self {
            Codec::Stateless => 0,
            Codec::Dictionary(converter, nulls) => converter.size() + nulls.data.len(),
            Codec::Struct(converter, nulls) => converter.size() + nulls.data.len(),
            Codec::List(converter) => converter.size(),
            Codec::Map(converter) => converter.size(),
            Codec::RunEndEncoded(converter) => converter.size(),
            Codec::Union(converters, _, null_rows) => {
                converters.iter().map(|c| c.size()).sum::<usize>()
                    + null_rows.iter().map(|n| n.data.len()).sum::<usize>()
            }
        }
    }
}

#[derive(Debug)]
enum Encoder<'a> {
    /// No additional encoder state is necessary
    Stateless,
    /// The encoding of the child array and the encoding of a null row
    Dictionary(Rows, Row<'a>),
    /// The row encoding of the child arrays and the encoding of a null row
    ///
    /// It is necessary to encode to a temporary [`Rows`] to avoid serializing
    /// values that are masked by a null in the parent StructArray, otherwise
    /// this would establish an ordering between semantically null values
    Struct(Rows, Row<'a>),
    /// The row encoding of the child array
    List(Rows),
    /// The row encoding of the map entries
    Map(Rows),
    /// The row encoding of the values array
    RunEndEncoded(Rows),
    /// The row encoding of each union field's child array, type_ids buffer, offsets buffer (for Dense), and mode
    Union {
        child_rows: Vec<Rows>,
        field_to_type_ids: Vec<i8>,
        type_ids: ScalarBuffer<i8>,
        offsets: Option<ScalarBuffer<i32>>,
    },
}

/// Configure the data type and sort order for a given column
#[derive(Debug, Clone, PartialEq, Eq)]
pub struct SortField {
    /// Sort options
    options: SortOptions,
    /// Data type
    data_type: DataType,
}

impl SortField {
    /// Create a new column with the given data type
    pub fn new(data_type: DataType) -> Self {
        Self::new_with_options(data_type, Default::default())
    }

    /// Create a new column with the given data type and [`SortOptions`]
    pub fn new_with_options(data_type: DataType, options: SortOptions) -> Self {
        Self { options, data_type }
    }

    /// Return size of this instance in bytes.
    ///
    /// Includes the size of `Self`.
    pub fn size(&self) -> usize {
        self.data_type.size() + std::mem::size_of::<Self>() - std::mem::size_of::<DataType>()
    }
}

impl RowConverter {
    /// Create a new [`RowConverter`] with the provided schema
    pub fn new(fields: Vec<SortField>) -> Result<Self, ArrowError> {
        if !Self::supports_fields(&fields) {
            return Err(ArrowError::NotYetImplemented(format!(
                "Row format support not yet implemented for: {fields:?}"
            )));
        }

        let codecs = fields.iter().map(Codec::new).collect::<Result<_, _>>()?;
        Ok(Self {
            fields: fields.into(),
            codecs,
        })
    }

    /// Check if the given fields are supported by the row format.
    pub fn supports_fields(fields: &[SortField]) -> bool {
        fields.iter().all(|x| Self::supports_datatype(&x.data_type))
    }

    fn supports_datatype(d: &DataType) -> bool {
        match d {
            _ if !d.is_nested() => true,
            DataType::List(f)
            | DataType::LargeList(f)
            | DataType::ListView(f)
            | DataType::LargeListView(f)
            | DataType::FixedSizeList(f, _)
            | DataType::Map(f, _) => Self::supports_datatype(f.data_type()),
            DataType::Struct(f) => f.iter().all(|x| Self::supports_datatype(x.data_type())),
            DataType::RunEndEncoded(_, values) => Self::supports_datatype(values.data_type()),
            DataType::Union(fs, _mode) => fs
                .iter()
                .all(|(_, f)| Self::supports_datatype(f.data_type())),
            _ => false,
        }
    }

    /// Convert [`ArrayRef`] columns into [`Rows`]
    ///
    /// See [`Row`] for information on when [`Row`] can be compared
    ///
    /// See [`Self::convert_rows`] for converting [`Rows`] back into [`ArrayRef`]
    ///
    /// # Panics
    ///
    /// Panics if the schema of `columns` does not match that provided to [`RowConverter::new`]
    pub fn convert_columns(&self, columns: &[ArrayRef]) -> Result<Rows, ArrowError> {
        let num_rows = columns.first().map(|x| x.len()).unwrap_or(0);
        let mut rows = self.empty_rows(num_rows, 0);
        self.append(&mut rows, columns)?;
        Ok(rows)
    }

    /// Convert [`ArrayRef`] columns appending to an existing [`Rows`]
    ///
    /// See [`Row`] for information on when [`Row`] can be compared
    ///
    /// # Panics
    ///
    /// Panics if
    /// * The schema of `columns` does not match that provided to [`RowConverter::new`]
    /// * The provided [`Rows`] were not created by this [`RowConverter`]
    ///
    /// ```
    /// # use std::sync::Arc;
    /// # use std::collections::HashSet;
    /// # use arrow_array::cast::AsArray;
    /// # use arrow_array::StringArray;
    /// # use arrow_row::{Row, RowConverter, SortField};
    /// # use arrow_schema::DataType;
    /// #
    /// let converter = RowConverter::new(vec![SortField::new(DataType::Utf8)]).unwrap();
    /// let a1 = StringArray::from(vec!["hello", "world"]);
    /// let a2 = StringArray::from(vec!["a", "a", "hello"]);
    ///
    /// let mut rows = converter.empty_rows(5, 128);
    /// converter.append(&mut rows, &[Arc::new(a1)]).unwrap();
    /// converter.append(&mut rows, &[Arc::new(a2)]).unwrap();
    ///
    /// let back = converter.convert_rows(&rows).unwrap();
    /// let values: Vec<_> = back[0].as_string::<i32>().iter().map(Option::unwrap).collect();
    /// assert_eq!(&values, &["hello", "world", "a", "a", "hello"]);
    /// ```
    pub fn append(&self, rows: &mut Rows, columns: &[ArrayRef]) -> Result<(), ArrowError> {
        assert!(
            Arc::ptr_eq(&rows.config.fields, &self.fields),
            "rows were not produced by this RowConverter"
        );

        if columns.len() != self.fields.len() {
            return Err(ArrowError::InvalidArgumentError(format!(
                "Incorrect number of arrays provided to RowConverter, expected {} got {}",
                self.fields.len(),
                columns.len()
            )));
        }
        for column in columns.iter().skip(1) {
            if column.len() != columns[0].len() {
                return Err(ArrowError::InvalidArgumentError(format!(
                    "RowConverter columns must all have the same length, expected {} got {}",
                    columns[0].len(),
                    column.len()
                )));
            }
        }

        let encoders = columns
            .iter()
            .zip(&self.codecs)
            .zip(self.fields.iter())
            .map(|((column, codec), field)| {
                if !column.data_type().equals_datatype(&field.data_type) {
                    return Err(ArrowError::InvalidArgumentError(format!(
                        "RowConverter column schema mismatch, expected {} got {}",
                        field.data_type,
                        column.data_type()
                    )));
                }
                codec.encoder(column.as_ref())
            })
            .collect::<Result<Vec<_>, _>>()?;

        let write_offset = rows.num_rows();
        let lengths = row_lengths(columns, &encoders);
        let total = lengths.extend_offsets(rows.offsets[write_offset], &mut rows.offsets);
        rows.buffer.resize(total, 0);

        for ((column, field), encoder) in columns.iter().zip(self.fields.iter()).zip(encoders) {
            // We encode a column at a time to minimise dispatch overheads
            encode_column(
                &mut rows.buffer,
                &mut rows.offsets[write_offset..],
                column.as_ref(),
                field.options,
                &encoder,
            )
        }

        if cfg!(debug_assertions) {
            assert_eq!(*rows.offsets.last().unwrap(), rows.buffer.len());
            rows.offsets
                .windows(2)
                .for_each(|w| assert!(w[0] <= w[1], "offsets should be monotonic"));
        }

        Ok(())
    }

    /// Convert [`Rows`] columns into [`ArrayRef`]
    ///
    /// See [`Self::convert_columns`] for converting [`ArrayRef`] into [`Rows`]
    ///
    /// # Panics
    ///
    /// Panics if the rows were not produced by this [`RowConverter`]
    pub fn convert_rows<'a, I>(&self, rows: I) -> Result<Vec<ArrayRef>, ArrowError>
    where
        I: IntoIterator<Item = Row<'a>>,
    {
        let mut validate_utf8 = false;
        let mut rows: Vec<_> = rows
            .into_iter()
            .map(|row| {
                assert!(
                    Arc::ptr_eq(&row.config.fields, &self.fields),
                    "rows were not produced by this RowConverter"
                );
                validate_utf8 |= row.config.validate_utf8;
                row.data
            })
            .collect();

        // SAFETY
        // We have validated that the rows came from this [`RowConverter`]
        // and therefore must be valid
        let result = unsafe { self.convert_raw(&mut rows, validate_utf8) }?;

        if cfg!(debug_assertions) {
            for (i, row) in rows.iter().enumerate() {
                if !row.is_empty() {
                    return Err(ArrowError::InvalidArgumentError(format!(
                        "Codecs {codecs:?} did not consume all bytes for row {i}, remaining bytes: {row:?}",
                        codecs = self.codecs
                    )));
                }
            }
        }

        Ok(result)
    }

    /// Returns an empty [`Rows`] with capacity for `row_capacity` rows with
    /// a total length of `data_capacity`
    ///
    /// This can be used to buffer a selection of [`Row`]
    ///
    /// ```
    /// # use std::sync::Arc;
    /// # use std::collections::HashSet;
    /// # use arrow_array::cast::AsArray;
    /// # use arrow_array::StringArray;
    /// # use arrow_row::{Row, RowConverter, SortField};
    /// # use arrow_schema::DataType;
    /// #
    /// let converter = RowConverter::new(vec![SortField::new(DataType::Utf8)]).unwrap();
    /// let array = StringArray::from(vec!["hello", "world", "a", "a", "hello"]);
    ///
    /// // Convert to row format and deduplicate
    /// let converted = converter.convert_columns(&[Arc::new(array)]).unwrap();
    /// let mut distinct_rows = converter.empty_rows(3, 100);
    /// let mut dedup: HashSet<Row> = HashSet::with_capacity(3);
    /// converted.iter().filter(|row| dedup.insert(*row)).for_each(|row| distinct_rows.push(row));
    ///
    /// // Note: we could skip buffering and feed the filtered iterator directly
    /// // into convert_rows, this is done for demonstration purposes only
    /// let distinct = converter.convert_rows(&distinct_rows).unwrap();
    /// let values: Vec<_> = distinct[0].as_string::<i32>().iter().map(Option::unwrap).collect();
    /// assert_eq!(&values, &["hello", "world", "a"]);
    /// ```
    pub fn empty_rows(&self, row_capacity: usize, data_capacity: usize) -> Rows {
        let mut offsets = Vec::with_capacity(row_capacity.saturating_add(1));
        offsets.push(0);

        Rows {
            offsets,
            buffer: Vec::with_capacity(data_capacity),
            config: RowConfig {
                fields: self.fields.clone(),
                validate_utf8: false,
            },
        }
    }

    /// Create a new [Rows] instance from the given binary data.
    ///
    /// ```
    /// # use std::sync::Arc;
    /// # use std::collections::HashSet;
    /// # use arrow_array::cast::AsArray;
    /// # use arrow_array::StringArray;
    /// # use arrow_row::{OwnedRow, Row, RowConverter, RowParser, SortField};
    /// # use arrow_schema::DataType;
    /// #
    /// let converter = RowConverter::new(vec![SortField::new(DataType::Utf8)]).unwrap();
    /// let array = StringArray::from(vec!["hello", "world", "a", "a", "hello"]);
    /// let rows = converter.convert_columns(&[Arc::new(array)]).unwrap();
    ///
    /// // We can convert rows into binary format and back in batch.
    /// let values: Vec<OwnedRow> = rows.iter().map(|r| r.owned()).collect();
    /// let binary = rows.try_into_binary().expect("known-small array");
    /// let converted = converter.from_binary(binary.clone());
    /// assert!(converted.iter().eq(values.iter().map(|r| r.row())));
    /// ```
    ///
    /// # Panics
    ///
    /// This function expects the passed [BinaryArray] to contain valid row data as produced by this
    /// [RowConverter]. It will panic if any rows are null. Operations on the returned [Rows] may
    /// panic if the data is malformed.
    pub fn from_binary(&self, array: BinaryArray) -> Rows {
        assert_eq!(
            array.null_count(),
            0,
            "can't construct Rows instance from array with nulls"
        );
        let (offsets, values, _) = array.into_parts();
        let offsets = offsets.iter().map(|&i| i.as_usize()).collect();
        // Try zero-copy, if it does not succeed, fall back to copying the values.
        let buffer = values.into_vec().unwrap_or_else(|values| values.to_vec());
        Rows {
            buffer,
            offsets,
            config: RowConfig {
                fields: Arc::clone(&self.fields),
                validate_utf8: true,
            },
        }
    }

    /// Convert raw bytes into [`ArrayRef`]
    ///
    /// # Safety
    ///
    /// `rows` must contain valid data for this [`RowConverter`]
    unsafe fn convert_raw(
        &self,
        rows: &mut [&[u8]],
        validate_utf8: bool,
    ) -> Result<Vec<ArrayRef>, ArrowError> {
        self.fields
            .iter()
            .zip(&self.codecs)
            .map(|(field, codec)| unsafe { decode_column(field, rows, codec, validate_utf8) })
            .collect()
    }

    /// Returns a [`RowParser`] that can be used to parse [`Row`] from bytes
    pub fn parser(&self) -> RowParser {
        RowParser::new(Arc::clone(&self.fields))
    }

    /// Like [`Self::parser`] but skips UTF-8 validation on decode.
    ///
    /// # Safety
    /// The caller must ensure all row bytes contain valid UTF-8 for string columns.
    pub unsafe fn parser_skip_utf8_validation(&self) -> RowParser {
        unsafe { RowParser::with_skip_utf8_validate(Arc::clone(&self.fields)) }
    }

    /// Returns the size of this instance in bytes
    ///
    /// Includes the size of `Self`.
    pub fn size(&self) -> usize {
        std::mem::size_of::<Self>()
            + self.fields.iter().map(|x| x.size()).sum::<usize>()
            + self.codecs.capacity() * std::mem::size_of::<Codec>()
            + self.codecs.iter().map(Codec::size).sum::<usize>()
    }
}

/// A [`RowParser`] can be created from a [`RowConverter`] and used to parse bytes to [`Row`]
#[derive(Debug)]
pub struct RowParser {
    config: RowConfig,
}

impl RowParser {
    fn new(fields: Arc<[SortField]>) -> Self {
        Self {
            config: RowConfig {
                fields,
                validate_utf8: true,
            },
        }
    }
    /// Like [`RowConverter::parser`] but skips UTF-8 validation on decode.
    ///
    /// # Safety
    /// The caller must ensure all row bytes contain valid UTF-8 for string columns.
    unsafe fn with_skip_utf8_validate(fields: Arc<[SortField]>) -> Self {
        Self {
            config: RowConfig {
                fields,
                validate_utf8: false,
            },
        }
    }

    /// Creates a [`Row`] from the provided `bytes`.
    ///
    /// `bytes` must be a [`Row`] produced by the [`RowConverter`] associated with
    /// this [`RowParser`], otherwise subsequent operations with the produced [`Row`] may panic
    pub fn parse<'a>(&'a self, bytes: &'a [u8]) -> Row<'a> {
        Row {
            data: bytes,
            config: &self.config,
        }
    }
}

/// The config of a given set of [`Row`]
#[derive(Debug, Clone)]
struct RowConfig {
    /// The schema for these rows
    fields: Arc<[SortField]>,
    /// Whether to run UTF-8 validation when converting to arrow arrays
    validate_utf8: bool,
}

/// A row-oriented representation of arrow data, that is normalized for comparison.
///
/// See the [module level documentation](self) and [`RowConverter`] for more details.
#[derive(Debug, Clone)]
pub struct Rows {
    /// Underlying row bytes
    buffer: Vec<u8>,
    /// Row `i` has data `&buffer[offsets[i]..offsets[i+1]]`
    offsets: Vec<usize>,
    /// The config for these rows
    config: RowConfig,
}

/// The iterator type for [`Rows::lengths`]
pub type RowLengthIter<'a> = Map<Windows<'a, usize>, fn(&'a [usize]) -> usize>;

impl Rows {
    /// Append a [`Row`] to this [`Rows`]
    ///
    /// # Panics
    ///
    /// Panics if `row` was not produced by the same [`RowConverter`] as `self`
    pub fn push(&mut self, row: Row<'_>) {
        assert!(
            Arc::ptr_eq(&row.config.fields, &self.config.fields),
            "row was not produced by this RowConverter"
        );
        self.config.validate_utf8 |= row.config.validate_utf8;
        self.buffer.extend_from_slice(row.data);
        self.offsets.push(self.buffer.len())
    }

    /// Reserve capacity for `row_capacity` rows with a total length of `data_capacity`
    pub fn reserve(&mut self, row_capacity: usize, data_capacity: usize) {
        self.buffer.reserve(data_capacity);
        self.offsets.reserve(row_capacity);
    }

    /// Returns the row at index `row`
    ///
    /// # Panics
    ///
    /// Panics if `row >= self.num_rows()`
    pub fn row(&self, row: usize) -> Row<'_> {
        self.checked_row_end(row);
        unsafe { self.row_unchecked(row) }
    }

    fn checked_row_end(&self, row: usize) -> usize {
        row.checked_add(1)
            .filter(|end| *end < self.offsets.len())
            .expect("row index out of bounds")
    }

    /// Returns the row at `index` without bounds checking
    ///
    /// # Safety
    /// Caller must ensure that `index + 1` is less than the number of offsets (#rows + 1)
    pub unsafe fn row_unchecked(&self, index: usize) -> Row<'_> {
        let end = unsafe { self.offsets.get_unchecked(index + 1) };
        let start = unsafe { self.offsets.get_unchecked(index) };
        let data = unsafe { self.buffer.get_unchecked(*start..*end) };
        Row {
            data,
            config: &self.config,
        }
    }

    /// Returns the number of bytes the row at index `row` is occupying,
    /// that is, what is the length of the returned [`Row::data`] will be.
    pub fn row_len(&self, row: usize) -> usize {
        let end = self.checked_row_end(row);

        self.offsets[end] - self.offsets[row]
    }

    /// Get an iterator over the lengths of each row in this [`Rows`]
    pub fn lengths(&self) -> RowLengthIter<'_> {
        self.offsets.windows(2).map(|w| w[1] - w[0])
    }

    /// Sets the length of this [`Rows`] to 0
    pub fn clear(&mut self) {
        self.offsets.truncate(1);
        self.buffer.clear();
    }

    /// Returns the number of [`Row`] in this [`Rows`]
    pub fn num_rows(&self) -> usize {
        self.offsets.len() - 1
    }

    /// Returns an iterator over the [`Row`] in this [`Rows`]
    pub fn iter(&self) -> RowsIter<'_> {
        self.into_iter()
    }

    /// Returns the size of this instance in bytes
    ///
    /// Includes the size of `Self`.
    pub fn size(&self) -> usize {
        // Size of fields is accounted for as part of RowConverter
        std::mem::size_of::<Self>()
            + self.buffer.capacity()
            + self.offsets.capacity() * std::mem::size_of::<usize>()
    }

    /// Create a [BinaryArray] from the [Rows] data without reallocating the
    /// underlying bytes.
    ///
    ///
    /// ```
    /// # use std::sync::Arc;
    /// # use std::collections::HashSet;
    /// # use arrow_array::cast::AsArray;
    /// # use arrow_array::StringArray;
    /// # use arrow_row::{OwnedRow, Row, RowConverter, RowParser, SortField};
    /// # use arrow_schema::DataType;
    /// #
    /// let converter = RowConverter::new(vec![SortField::new(DataType::Utf8)]).unwrap();
    /// let array = StringArray::from(vec!["hello", "world", "a", "a", "hello"]);
    /// let rows = converter.convert_columns(&[Arc::new(array)]).unwrap();
    ///
    /// // We can convert rows into binary format and back.
    /// let values: Vec<OwnedRow> = rows.iter().map(|r| r.owned()).collect();
    /// let binary = rows.try_into_binary().expect("known-small array");
    /// let parser = converter.parser();
    /// let parsed: Vec<OwnedRow> =
    ///   binary.iter().flatten().map(|b| parser.parse(b).owned()).collect();
    /// assert_eq!(values, parsed);
    /// ```
    ///
    /// # Errors
    ///
    /// This function will return an error if there is more data than can be stored in
    /// a [BinaryArray] -- i.e. if the total data size is more than 2GiB.
    pub fn try_into_binary(self) -> Result<BinaryArray, ArrowError> {
        if self.buffer.len() > i32::MAX as usize {
            return Err(ArrowError::InvalidArgumentError(format!(
                "{}-byte rows buffer too long to convert into a i32-indexed BinaryArray",
                self.buffer.len()
            )));
        }
        // We've checked that the buffer length fits in an i32; so all offsets into that buffer should fit as well.
        let offsets_scalar = ScalarBuffer::from_iter(self.offsets.into_iter().map(i32::usize_as));
        // SAFETY: offsets buffer is nonempty, monotonically increasing, and all represent valid indexes into buffer.
        let array = unsafe {
            BinaryArray::new_unchecked(
                OffsetBuffer::new_unchecked(offsets_scalar),
                Buffer::from_vec(self.buffer),
                None,
            )
        };
        Ok(array)
    }
}

impl<'a> IntoIterator for &'a Rows {
    type Item = Row<'a>;
    type IntoIter = RowsIter<'a>;

    fn into_iter(self) -> Self::IntoIter {
        RowsIter {
            rows: self,
            start: 0,
            end: self.num_rows(),
        }
    }
}

/// An iterator over [`Rows`]
#[derive(Debug)]
pub struct RowsIter<'a> {
    rows: &'a Rows,
    start: usize,
    end: usize,
}

impl<'a> Iterator for RowsIter<'a> {
    type Item = Row<'a>;

    fn next(&mut self) -> Option<Self::Item> {
        if self.end == self.start {
            return None;
        }

        // SAFETY: We have checked that `start` is less than `end`
        let row = unsafe { self.rows.row_unchecked(self.start) };
        self.start += 1;
        Some(row)
    }

    fn size_hint(&self) -> (usize, Option<usize>) {
        let len = self.len();
        (len, Some(len))
    }
}

impl ExactSizeIterator for RowsIter<'_> {
    fn len(&self) -> usize {
        self.end - self.start
    }
}

impl DoubleEndedIterator for RowsIter<'_> {
    fn next_back(&mut self) -> Option<Self::Item> {
        if self.end == self.start {
            return None;
        }

        self.end -= 1;

        // Safety: By construction we create `end >= start`, so if `end` is not equal to `start` it cannot be less than `start`
        //          therefore `end - 1` is within range
        let row = unsafe { self.rows.row_unchecked(self.end) };
        Some(row)
    }
}

/// A comparable representation of a row.
///
/// See the [module level documentation](self) for more details.
///
/// Two [`Row`] can only be compared if they both belong to [`Rows`]
/// returned by calls to [`RowConverter::convert_columns`] on the same
/// [`RowConverter`]. If different [`RowConverter`]s are used, any
/// ordering established by comparing the [`Row`] is arbitrary.
#[derive(Debug, Copy, Clone)]
pub struct Row<'a> {
    data: &'a [u8],
    config: &'a RowConfig,
}

impl<'a> Row<'a> {
    /// Create owned version of the row to detach it from the shared [`Rows`].
    pub fn owned(&self) -> OwnedRow {
        OwnedRow {
            data: self.data.into(),
            config: self.config.clone(),
        }
    }

    /// The row's bytes, with the lifetime of the underlying data.
    pub fn data(&self) -> &'a [u8] {
        self.data
    }
}

// Manually derive these as don't wish to include `fields`

impl PartialEq for Row<'_> {
    #[inline]
    fn eq(&self, other: &Self) -> bool {
        self.data.eq(other.data)
    }
}

impl Eq for Row<'_> {}

impl PartialOrd for Row<'_> {
    #[inline]
    fn partial_cmp(&self, other: &Self) -> Option<Ordering> {
        Some(self.cmp(other))
    }
}

impl Ord for Row<'_> {
    #[inline]
    fn cmp(&self, other: &Self) -> Ordering {
        self.data.cmp(other.data)
    }
}

impl Hash for Row<'_> {
    #[inline]
    fn hash<H: Hasher>(&self, state: &mut H) {
        self.data.hash(state)
    }
}

impl AsRef<[u8]> for Row<'_> {
    #[inline]
    fn as_ref(&self) -> &[u8] {
        self.data
    }
}

/// Owned version of a [`Row`] that can be moved/cloned freely.
///
/// This contains the data for the one specific row (not the entire buffer of all rows).
#[derive(Debug, Clone)]
pub struct OwnedRow {
    data: Box<[u8]>,
    config: RowConfig,
}

impl OwnedRow {
    /// Get borrowed [`Row`] from owned version.
    ///
    /// This is helpful if you want to compare an [`OwnedRow`] with a [`Row`].
    pub fn row(&self) -> Row<'_> {
        Row {
            data: &self.data,
            config: &self.config,
        }
    }
}

// Manually derive these as don't wish to include `fields`. Also we just want to use the same `Row` implementations here.

impl PartialEq for OwnedRow {
    #[inline]
    fn eq(&self, other: &Self) -> bool {
        self.row().eq(&other.row())
    }
}

impl Eq for OwnedRow {}

impl PartialOrd for OwnedRow {
    #[inline]
    fn partial_cmp(&self, other: &Self) -> Option<Ordering> {
        Some(self.cmp(other))
    }
}

impl Ord for OwnedRow {
    #[inline]
    fn cmp(&self, other: &Self) -> Ordering {
        self.row().cmp(&other.row())
    }
}

impl Hash for OwnedRow {
    #[inline]
    fn hash<H: Hasher>(&self, state: &mut H) {
        self.row().hash(state)
    }
}

impl AsRef<[u8]> for OwnedRow {
    #[inline]
    fn as_ref(&self) -> &[u8] {
        &self.data
    }
}

/// Returns the null sentinel, negated if `invert` is true
#[inline]
fn null_sentinel(options: SortOptions) -> u8 {
    match options.nulls_first {
        true => 0,
        false => 0xFF,
    }
}

/// Stores the lengths of the rows. Lazily materializes lengths for columns with fixed-size types.
enum LengthTracker {
    /// Fixed state: All rows have length `length`
    Fixed { length: usize, num_rows: usize },
    /// Variable state: The length of row `i` is `lengths[i] + fixed_length`
    Variable {
        fixed_length: usize,
        lengths: Vec<usize>,
    },
}

impl LengthTracker {
    fn new(num_rows: usize) -> Self {
        Self::Fixed {
            length: 0,
            num_rows,
        }
    }

    /// Adds a column of fixed-length elements, each of size `new_length` to the LengthTracker
    fn push_fixed(&mut self, new_length: usize) {
        match self {
            LengthTracker::Fixed { length, .. } => *length += new_length,
            LengthTracker::Variable { fixed_length, .. } => *fixed_length += new_length,
        }
    }

    /// Adds a column of possibly variable-length elements, element `i` has length `new_lengths.nth(i)`
    fn push_variable(&mut self, new_lengths: impl ExactSizeIterator<Item = usize>) {
        match self {
            LengthTracker::Fixed { length, .. } => {
                *self = LengthTracker::Variable {
                    fixed_length: *length,
                    lengths: new_lengths.collect(),
                }
            }
            LengthTracker::Variable { lengths, .. } => {
                assert_eq!(lengths.len(), new_lengths.len());
                lengths
                    .iter_mut()
                    .zip(new_lengths)
                    .for_each(|(length, new_length)| *length += new_length);
            }
        }
    }

    /// Returns the tracked row lengths as a slice
    fn materialized(&mut self) -> &mut [usize] {
        if let LengthTracker::Fixed { length, num_rows } = *self {
            *self = LengthTracker::Variable {
                fixed_length: length,
                lengths: vec![0; num_rows],
            };
        }

        match self {
            LengthTracker::Variable { lengths, .. } => lengths,
            LengthTracker::Fixed { .. } => unreachable!(),
        }
    }

    /// Initializes the offsets using the tracked lengths. Returns the sum of the
    /// lengths of the rows added.
    ///
    /// We initialize the offsets shifted down by one row index.
    ///
    /// As the rows are appended to the offsets will be incremented to match
    ///
    /// For example, consider the case of 3 rows of length 3, 4, and 6 respectively.
    /// The offsets would be initialized to `0, 0, 3, 7`
    ///
    /// Writing the first row entirely would yield `0, 3, 3, 7`
    /// The second, `0, 3, 7, 7`
    /// The third, `0, 3, 7, 13`
    //
    /// This would be the final offsets for reading
    //
    /// In this way offsets tracks the position during writing whilst eventually serving
    fn extend_offsets(&self, initial_offset: usize, offsets: &mut Vec<usize>) -> usize {
        match self {
            LengthTracker::Fixed { length, num_rows } => {
                offsets.extend((0..*num_rows).map(|i| initial_offset + i * length));

                initial_offset + num_rows * length
            }
            LengthTracker::Variable {
                fixed_length,
                lengths,
            } => {
                let mut acc = initial_offset;

                offsets.extend(lengths.iter().map(|length| {
                    let current = acc;
                    acc += length + fixed_length;
                    current
                }));

                acc
            }
        }
    }
}

/// Computes the length of each encoded [`Rows`] and returns an empty [`Rows`]
fn row_lengths(cols: &[ArrayRef], encoders: &[Encoder]) -> LengthTracker {
    use fixed::FixedLengthEncoding;

    let num_rows = cols.first().map(|x| x.len()).unwrap_or(0);
    let mut tracker = LengthTracker::new(num_rows);

    for (array, encoder) in cols.iter().zip(encoders) {
        match encoder {
            Encoder::Stateless => {
                downcast_primitive_array! {
                    array => tracker.push_fixed(fixed::encoded_len(array)),
                    DataType::Null => tracker.push_fixed(2)
                    DataType::Boolean => tracker.push_fixed(bool::ENCODED_LEN),
                    DataType::Binary => push_generic_byte_array_lengths(&mut tracker, as_generic_binary_array::<i32>(array)),
                    DataType::LargeBinary => push_generic_byte_array_lengths(&mut tracker, as_generic_binary_array::<i64>(array)),
                    DataType::BinaryView => push_byte_view_array_lengths(&mut tracker, array.as_binary_view()),
                    DataType::Utf8 => push_generic_byte_array_lengths(&mut tracker, array.as_string::<i32>()),
                    DataType::LargeUtf8 => push_generic_byte_array_lengths(&mut tracker, array.as_string::<i64>()),
                    DataType::Utf8View => push_byte_view_array_lengths(&mut tracker, array.as_string_view()),
                    DataType::FixedSizeBinary(len) => {
                        let len = len.to_usize().unwrap();
                        tracker.push_fixed(1 + len)
                    }
                    _ => unimplemented!("unsupported data type: {}", array.data_type()),
                }
            }
            Encoder::Dictionary(values, null) => {
                downcast_dictionary_array! {
                    array => {
                        tracker.push_variable(
                            array.keys().iter().map(|v| match v {
                                Some(k) => values.row_len(k.as_usize()),
                                None => null.data.len(),
                            })
                        )
                    }
                    _ => unreachable!(),
                }
            }
            Encoder::Struct(rows, null) => {
                let array = as_struct_array(array);
                if rows.num_rows() > 0 {
                    // Only calculate row length if there are rows
                    tracker.push_variable((0..array.len()).map(|idx| match array.is_valid(idx) {
                        true => 1 + rows.row_len(idx),
                        false => 1 + null.data.len(),
                    }));
                } else {
                    // Edge case for Struct([]) arrays (no child fields)
                    tracker.push_variable((0..array.len()).map(|idx| match array.is_valid(idx) {
                        true => 1,
                        false => 1 + null.data.len(),
                    }));
                }
            }
            Encoder::List(rows) => match array.data_type() {
                DataType::List(_) => {
                    list::compute_lengths(tracker.materialized(), rows, as_list_array(array))
                }
                DataType::LargeList(_) => {
                    list::compute_lengths(tracker.materialized(), rows, as_large_list_array(array))
                }
                DataType::ListView(_) => {
                    let list_view = array.as_list_view::<i32>();
                    let (min_offset, _) = compute_list_view_bounds(list_view);
                    list::compute_lengths_list_view(
                        tracker.materialized(),
                        rows,
                        list_view,
                        min_offset,
                    )
                }
                DataType::LargeListView(_) => {
                    let list_view = array.as_list_view::<i64>();
                    let (min_offset, _) = compute_list_view_bounds(list_view);
                    list::compute_lengths_list_view(
                        tracker.materialized(),
                        rows,
                        list_view,
                        min_offset,
                    )
                }
                DataType::FixedSizeList(_, _) => compute_lengths_fixed_size_list(
                    &mut tracker,
                    rows,
                    as_fixed_size_list_array(array),
                ),
                _ => unreachable!(),
            },
            Encoder::Map(rows) => {
                list::compute_lengths(tracker.materialized(), rows, as_map_array(array))
            }
            Encoder::RunEndEncoded(rows) => match array.data_type() {
                DataType::RunEndEncoded(r, _) => match r.data_type() {
                    DataType::Int16 => run::compute_lengths(
                        tracker.materialized(),
                        rows,
                        array.as_run::<Int16Type>(),
                    ),
                    DataType::Int32 => run::compute_lengths(
                        tracker.materialized(),
                        rows,
                        array.as_run::<Int32Type>(),
                    ),
                    DataType::Int64 => run::compute_lengths(
                        tracker.materialized(),
                        rows,
                        array.as_run::<Int64Type>(),
                    ),
                    _ => unreachable!("Unsupported run end index type: {r:?}"),
                },
                _ => unreachable!(),
            },
            Encoder::Union {
                child_rows,
                field_to_type_ids,
                type_ids,
                offsets,
            } => {
                let union_array = array
                    .as_any()
                    .downcast_ref::<UnionArray>()
                    .expect("expected UnionArray");

                let mut type_id_to_field_idx = [0usize; 128];
                for (field_idx, &type_id) in field_to_type_ids.iter().enumerate() {
                    type_id_to_field_idx[type_id as usize] = field_idx;
                }

                let lengths = (0..union_array.len()).map(|i| {
                    let type_id = type_ids[i];
                    let field_idx = type_id_to_field_idx[type_id as usize];
                    let child_row_i = offsets.as_ref().map(|o| o[i] as usize).unwrap_or(i);
                    let child_row_len = child_rows[field_idx].row_len(child_row_i);

                    // length: 1 byte type_id + child row bytes
                    1 + child_row_len
                });

                tracker.push_variable(lengths);
            }
        }
    }

    tracker
}

/// Add to [`LengthTracker`] the encoded length of each item in the [`GenericByteArray`]
fn push_generic_byte_array_lengths<T: ByteArrayType>(
    tracker: &mut LengthTracker,
    array: &GenericByteArray<T>,
) {
    if let Some(nulls) = array.nulls().filter(|n| n.null_count() > 0) {
        tracker.push_variable(
            array
                .offsets()
                .lengths()
                .zip(nulls.iter())
                .map(|(length, is_valid)| if is_valid { Some(length) } else { None })
                .map(variable::padded_length),
        )
    } else {
        tracker.push_variable(
            array
                .offsets()
                .lengths()
                .map(variable::non_null_padded_length),
        )
    }
}

/// Add to [`LengthTracker`] the encoded length of each item in the [`GenericByteViewArray`]
fn push_byte_view_array_lengths<T: ByteViewType>(
    tracker: &mut LengthTracker,
    array: &GenericByteViewArray<T>,
) {
    if let Some(nulls) = array.nulls().filter(|n| n.null_count() > 0) {
        tracker.push_variable(
            array
                .lengths()
                .zip(nulls.iter())
                .map(|(length, is_valid)| {
                    if is_valid {
                        Some(length as usize)
                    } else {
                        None
                    }
                })
                .map(variable::padded_length),
        )
    } else {
        tracker.push_variable(
            array
                .lengths()
                .map(|len| variable::padded_length(Some(len as usize))),
        )
    }
}

/// Encodes a column to the provided [`Rows`] incrementing the offsets as it progresses
fn encode_column(
    data: &mut [u8],
    offsets: &mut [usize],
    column: &dyn Array,
    opts: SortOptions,
    encoder: &Encoder<'_>,
) {
    match encoder {
        Encoder::Stateless => {
            downcast_primitive_array! {
                column => {
                    if let Some(nulls) = column.nulls().filter(|n| n.null_count() > 0){
                        fixed::encode(data, offsets, column.values(), nulls, opts)
                    } else {
                        fixed::encode_not_null(data, offsets, column.values(), opts)
                    }
                }
                DataType::Null => {
                    for offset in offsets.iter_mut().skip(1) {
                        variable::encode_null_value(&mut data[*offset..], opts);
                        *offset += 2;
                    }
                }
                DataType::Boolean => {
                    if let Some(nulls) = column.nulls().filter(|n| n.null_count() > 0){
                        fixed::encode_boolean(data, offsets, column.as_boolean().values(), nulls, opts)
                    } else {
                        fixed::encode_boolean_not_null(data, offsets, column.as_boolean().values(), opts)
                    }
                }
                DataType::Binary => {
                    variable::encode_generic_byte_array(data, offsets, as_generic_binary_array::<i32>(column), opts)
                }
                DataType::BinaryView => {
                    variable::encode(data, offsets, column.as_binary_view().iter(), opts)
                }
                DataType::LargeBinary => {
                    variable::encode_generic_byte_array(data, offsets, as_generic_binary_array::<i64>(column), opts)
                }
                DataType::Utf8 => variable::encode_generic_byte_array(
                    data, offsets,
                    column.as_string::<i32>(),
                    opts,
                ),
                DataType::LargeUtf8 => variable::encode_generic_byte_array(
                    data, offsets,
                    column.as_string::<i64>(),
                    opts,
                ),
                DataType::Utf8View => variable::encode(
                    data, offsets,
                    column.as_string_view().iter().map(|x| x.map(|x| x.as_bytes())),
                    opts,
                ),
                DataType::FixedSizeBinary(_) => {
                    let array = column.as_any().downcast_ref().unwrap();
                    fixed::encode_fixed_size_binary(data, offsets, array, opts)
                }
                _ => unimplemented!("unsupported data type: {}", column.data_type()),
            }
        }
        Encoder::Dictionary(values, nulls) => {
            downcast_dictionary_array! {
                column => encode_dictionary_values(data, offsets, column, values, nulls),
                _ => unreachable!()
            }
        }
        Encoder::Struct(rows, null) => {
            fn struct_encode_helper<const NO_CHILD_FIELDS: bool>(
                array: &StructArray,
                offsets: &mut [usize],
                null_sentinel: u8,
                rows: &Rows,
                null: &Row<'_>,
                data: &mut [u8],
            ) {
                let empty_row = Row {
                    data: &[],
                    config: &rows.config,
                };

                offsets
                    .iter_mut()
                    .skip(1)
                    .enumerate()
                    .for_each(|(idx, offset)| {
                        let (row, sentinel) = match array.is_valid(idx) {
                            true => (
                                if NO_CHILD_FIELDS {
                                    empty_row
                                } else {
                                    rows.row(idx)
                                },
                                0x01,
                            ),
                            false => (*null, null_sentinel),
                        };
                        let end_offset = *offset + 1 + row.as_ref().len();
                        data[*offset] = sentinel;
                        data[*offset + 1..end_offset].copy_from_slice(row.as_ref());
                        *offset = end_offset;
                    })
            }

            let array = as_struct_array(column);
            let null_sentinel = null_sentinel(opts);
            if rows.num_rows() == 0 {
                // Edge case for Struct([]) arrays (no child fields)
                struct_encode_helper::<true>(array, offsets, null_sentinel, rows, null, data);
            } else {
                struct_encode_helper::<false>(array, offsets, null_sentinel, rows, null, data);
            }
        }
        Encoder::List(rows) => match column.data_type() {
            DataType::List(_) => list::encode(data, offsets, rows, opts, as_list_array(column)),
            DataType::LargeList(_) => {
                list::encode(data, offsets, rows, opts, as_large_list_array(column))
            }
            DataType::ListView(_) => {
                let list_view = column.as_list_view::<i32>();
                let (min_offset, _) = compute_list_view_bounds(list_view);
                list::encode_list_view(data, offsets, rows, opts, list_view, min_offset)
            }
            DataType::LargeListView(_) => {
                let list_view = column.as_list_view::<i64>();
                let (min_offset, _) = compute_list_view_bounds(list_view);
                list::encode_list_view(data, offsets, rows, opts, list_view, min_offset)
            }
            DataType::FixedSizeList(_, _) => {
                encode_fixed_size_list(data, offsets, rows, opts, as_fixed_size_list_array(column))
            }
            _ => unreachable!(),
        },
        Encoder::Map(rows) => list::encode(data, offsets, rows, opts, as_map_array(column)),
        Encoder::RunEndEncoded(rows) => match column.data_type() {
            DataType::RunEndEncoded(r, _) => match r.data_type() {
                DataType::Int16 => {
                    run::encode(data, offsets, rows, opts, column.as_run::<Int16Type>())
                }
                DataType::Int32 => {
                    run::encode(data, offsets, rows, opts, column.as_run::<Int32Type>())
                }
                DataType::Int64 => {
                    run::encode(data, offsets, rows, opts, column.as_run::<Int64Type>())
                }
                _ => unreachable!("Unsupported run end index type: {r:?}"),
            },
            _ => unreachable!(),
        },
        Encoder::Union {
            child_rows,
            field_to_type_ids,
            type_ids,
            offsets: offsets_buf,
        } => {
            let mut type_id_to_field_idx = [0usize; 128];
            for (field_idx, &type_id) in field_to_type_ids.iter().enumerate() {
                type_id_to_field_idx[type_id as usize] = field_idx;
            }

            offsets
                .iter_mut()
                .skip(1)
                .enumerate()
                .for_each(|(i, offset)| {
                    let type_id = type_ids[i];
                    let field_idx = type_id_to_field_idx[type_id as usize];

                    let child_row_idx = offsets_buf.as_ref().map(|o| o[i] as usize).unwrap_or(i);
                    let child_row = child_rows[field_idx].row(child_row_idx);
                    let child_bytes = child_row.as_ref();

                    let type_id_byte = if opts.descending {
                        !(type_id as u8)
                    } else {
                        type_id as u8
                    };
                    data[*offset] = type_id_byte;

                    let child_start = *offset + 1;
                    let child_end = child_start + child_bytes.len();
                    data[child_start..child_end].copy_from_slice(child_bytes);

                    *offset = child_end;
                });
        }
    }
}

/// Encode dictionary values not preserving the dictionary encoding
pub fn encode_dictionary_values<K: ArrowDictionaryKeyType>(
    data: &mut [u8],
    offsets: &mut [usize],
    column: &DictionaryArray<K>,
    values: &Rows,
    null: &Row<'_>,
) {
    for (offset, k) in offsets.iter_mut().skip(1).zip(column.keys()) {
        let row = match k {
            Some(k) => values.row(k.as_usize()).data,
            None => null.data,
        };
        let end_offset = *offset + row.len();
        data[*offset..end_offset].copy_from_slice(row);
        *offset = end_offset;
    }
}

macro_rules! decode_primitive_helper {
    ($t:ty, $rows:ident, $data_type:ident, $options:ident) => {
        Arc::new(decode_primitive::<$t>($rows, $data_type, $options))
    };
}

/// Decodes a the provided `field` from `rows`
///
/// # Safety
///
/// Rows must contain valid data for the provided field
unsafe fn decode_column(
    field: &SortField,
    rows: &mut [&[u8]],
    codec: &Codec,
    validate_utf8: bool,
) -> Result<ArrayRef, ArrowError> {
    let options = field.options;

    let array: ArrayRef = match codec {
        Codec::Stateless => {
            let data_type = field.data_type.clone();
            downcast_primitive! {
                data_type => (decode_primitive_helper, rows, data_type, options),
                DataType::Null => {
                    variable::decode_null_value(rows, options);
                    Arc::new(NullArray::new(rows.len()))
                }
                DataType::Boolean => Arc::new(decode_bool(rows, options)),
                DataType::Binary => Arc::new(decode_binary::<i32>(rows, options)),
                DataType::LargeBinary => Arc::new(decode_binary::<i64>(rows, options)),
                DataType::BinaryView => Arc::new(decode_binary_view(rows, options)),
                DataType::FixedSizeBinary(size) => Arc::new(decode_fixed_size_binary(rows, size, options)),
                DataType::Utf8 => Arc::new(unsafe{ decode_string::<i32>(rows, options, validate_utf8) }),
                DataType::LargeUtf8 => Arc::new(unsafe { decode_string::<i64>(rows, options, validate_utf8) }),
                DataType::Utf8View => Arc::new(unsafe { decode_string_view(rows, options, validate_utf8) }),
                _ => return Err(ArrowError::NotYetImplemented(format!("unsupported data type: {data_type}" )))
            }
        }
        Codec::Dictionary(converter, _) => {
            let cols = unsafe { converter.convert_raw(rows, validate_utf8) }?;
            cols.into_iter().next().unwrap()
        }
        Codec::Struct(converter, _) => {
            let nulls = fixed::decode_nulls(rows);
            rows.iter_mut().for_each(|row| *row = &row[1..]);
            let children = unsafe { converter.convert_raw(rows, validate_utf8) }?;

            // Since RowConverter flattens certain data types (i.e. Dictionary),
            // we need to use updated data type instead of original field
            let corrected_fields: Vec<Field> = match &field.data_type {
                DataType::Struct(struct_fields) => struct_fields
                    .iter()
                    .zip(children.iter())
                    .map(|(orig_field, child_array)| {
                        orig_field
                            .as_ref()
                            .clone()
                            .with_data_type(child_array.data_type().clone())
                    })
                    .collect(),
                _ => unreachable!("Only Struct types should be corrected here"),
            };

            Arc::new(unsafe {
                StructArray::new_unchecked_with_length(
                    corrected_fields.into(),
                    children,
                    nulls,
                    rows.len(),
                )
            })
        }
        Codec::List(converter) => match &field.data_type {
            DataType::List(_) => Arc::new(unsafe {
                list::decode::<GenericListArray<i32>>(converter, rows, field, validate_utf8)
            }?),
            DataType::LargeList(_) => Arc::new(unsafe {
                list::decode::<GenericListArray<i64>>(converter, rows, field, validate_utf8)
            }?),
            DataType::ListView(_) => Arc::new(unsafe {
                list::decode_list_view::<i32>(converter, rows, field, validate_utf8)
            }?),
            DataType::LargeListView(_) => Arc::new(unsafe {
                list::decode_list_view::<i64>(converter, rows, field, validate_utf8)
            }?),
            DataType::FixedSizeList(_, value_length) => Arc::new(unsafe {
                list::decode_fixed_size_list(
                    converter,
                    rows,
                    field,
                    validate_utf8,
                    value_length.as_usize(),
                )
            }?),
            _ => unreachable!(),
        },
        Codec::Map(converter) => {
            Arc::new(unsafe { list::decode::<MapArray>(converter, rows, field, validate_utf8) }?)
        }
        Codec::RunEndEncoded(converter) => match &field.data_type {
            DataType::RunEndEncoded(run_ends, _) => match run_ends.data_type() {
                DataType::Int16 => Arc::new(unsafe {
                    run::decode::<Int16Type>(converter, rows, field, validate_utf8)
                }?),
                DataType::Int32 => Arc::new(unsafe {
                    run::decode::<Int32Type>(converter, rows, field, validate_utf8)
                }?),
                DataType::Int64 => Arc::new(unsafe {
                    run::decode::<Int64Type>(converter, rows, field, validate_utf8)
                }?),
                _ => unreachable!(),
            },
            _ => unreachable!(),
        },
        Codec::Union(converters, field_to_type_ids, null_rows) => {
            let len = rows.len();

            let DataType::Union(union_fields, mode) = &field.data_type else {
                unreachable!()
            };

            let mut type_id_to_field_idx = [0usize; 128];
            for (field_idx, &type_id) in field_to_type_ids.iter().enumerate() {
                type_id_to_field_idx[type_id as usize] = field_idx;
            }

            let mut type_ids = Vec::with_capacity(len);
            let mut rows_by_field: Vec<Vec<(usize, &[u8])>> = vec![Vec::new(); converters.len()];

            for (idx, row) in rows.iter_mut().enumerate() {
                let type_id_byte = {
                    let id = row[0];
                    if options.descending { !id } else { id }
                };

                let type_id = type_id_byte as i8;
                type_ids.push(type_id);

                let field_idx = type_id_to_field_idx[type_id as usize];

                let child_row = &row[1..];
                rows_by_field[field_idx].push((idx, child_row));
            }

            let mut child_arrays: Vec<ArrayRef> = Vec::with_capacity(converters.len());
            let mut offsets = (*mode == UnionMode::Dense).then(|| Vec::with_capacity(len));

            for (field_idx, converter) in converters.iter().enumerate() {
                let field_rows = &rows_by_field[field_idx];

                match &mode {
                    UnionMode::Dense => {
                        if field_rows.is_empty() {
                            let (_, field) = union_fields.iter().nth(field_idx).unwrap();
                            child_arrays.push(arrow_array::new_empty_array(field.data_type()));
                            continue;
                        }

                        let mut child_data = field_rows
                            .iter()
                            .map(|(_, bytes)| *bytes)
                            .collect::<Vec<_>>();

                        let child_array =
                            unsafe { converter.convert_raw(&mut child_data, validate_utf8) }?;

                        // advance row slices by the bytes consumed
                        for ((row_idx, original_bytes), remaining_bytes) in
                            field_rows.iter().zip(child_data)
                        {
                            let consumed_length = 1 + original_bytes.len() - remaining_bytes.len();
                            rows[*row_idx] = &rows[*row_idx][consumed_length..];
                        }

                        child_arrays.push(child_array.into_iter().next().unwrap());
                    }
                    UnionMode::Sparse => {
                        let mut sparse_data: Vec<&[u8]> = Vec::with_capacity(len);
                        let mut field_row_iter = field_rows.iter().peekable();
                        let null_row_bytes: &[u8] = &null_rows[field_idx].data;

                        for idx in 0..len {
                            if let Some((next_idx, bytes)) = field_row_iter.peek()
                                && *next_idx == idx
                            {
                                sparse_data.push(*bytes);

                                field_row_iter.next();
                                continue;
                            }
                            sparse_data.push(null_row_bytes);
                        }

                        let child_array =
                            unsafe { converter.convert_raw(&mut sparse_data, validate_utf8) }?;

                        // advance row slices by the bytes consumed for rows that belong to this field
                        for (row_idx, child_row) in field_rows {
                            let remaining_len = sparse_data[*row_idx].len();
                            let consumed_length = 1 + child_row.len() - remaining_len;
                            rows[*row_idx] = &rows[*row_idx][consumed_length..];
                        }

                        child_arrays.push(child_array.into_iter().next().unwrap());
                    }
                }
            }

            // build offsets for dense unions
            if let Some(ref mut offsets_vec) = offsets {
                let mut count = vec![0i32; converters.len()];
                for type_id in &type_ids {
                    let field_idx = *type_id as usize;
                    offsets_vec.push(count[field_idx]);

                    count[field_idx] += 1;
                }
            }

            let type_ids_buffer = ScalarBuffer::from(type_ids);
            let offsets_buffer = offsets.map(ScalarBuffer::from);

            let union_array = UnionArray::try_new(
                union_fields.clone(),
                type_ids_buffer,
                offsets_buffer,
                child_arrays,
            )?;

            // note: union arrays don't support physical null buffers
            // nulls are represented logically though child arrays
            Arc::new(union_array)
        }
    };
    Ok(array)
}

#[cfg(test)]
mod tests {
    use arrow_array::builder::*;
    use arrow_array::types::*;
    use arrow_buffer::{Buffer, OffsetBuffer};
    use arrow_buffer::{NullBuffer, i256};
    use arrow_cast::display::{ArrayFormatter, FormatOptions};
    use arrow_ord::sort::{LexicographicalComparator, SortColumn};
    use rand::distr::uniform::SampleUniform;
    use rand::distr::{Distribution, StandardUniform};
    use rand::prelude::StdRng;
    use rand::{RngExt, SeedableRng};

    use super::*;

    fn all_sort_options() -> [SortOptions; 4] {
        [
            SortOptions {
                descending: false,
                nulls_first: false,
            },
            SortOptions {
                descending: false,
                nulls_first: true,
            },
            SortOptions {
                descending: true,
                nulls_first: false,
            },
            SortOptions {
                descending: true,
                nulls_first: true,
            },
        ]
    }

    #[test]
    fn test_fixed_width() {
        let cols = [
            Arc::new(Int16Array::from_iter([
                Some(1),
                Some(2),
                None,
                Some(-5),
                Some(2),
                Some(2),
                Some(0),
            ])) as ArrayRef,
            Arc::new(Float32Array::from_iter([
                Some(1.3),
                Some(2.5),
                None,
                Some(4.),
                Some(0.1),
                Some(-4.),
                Some(-0.),
            ])) as ArrayRef,
        ];

        let converter = RowConverter::new(vec![
            SortField::new(DataType::Int16),
            SortField::new(DataType::Float32),
        ])
        .unwrap();
        let rows = converter.convert_columns(&cols).unwrap();

        assert_eq!(rows.offsets, &[0, 8, 16, 24, 32, 40, 48, 56]);
        assert_eq!(
            rows.buffer,
            &[
                1, 128, 1, //
                1, 191, 166, 102, 102, //
                1, 128, 2, //
                1, 192, 32, 0, 0, //
                0, 0, 0, //
                0, 0, 0, 0, 0, //
                1, 127, 251, //
                1, 192, 128, 0, 0, //
                1, 128, 2, //
                1, 189, 204, 204, 205, //
                1, 128, 2, //
                1, 63, 127, 255, 255, //
                1, 128, 0, //
                1, 127, 255, 255, 255 //
            ]
        );

        assert!(rows.row(3) < rows.row(6));
        assert!(rows.row(0) < rows.row(1));
        assert!(rows.row(3) < rows.row(0));
        assert!(rows.row(4) < rows.row(1));
        assert!(rows.row(5) < rows.row(4));

        let back = converter.convert_rows(&rows).unwrap();
        for (expected, actual) in cols.iter().zip(&back) {
            assert_eq!(expected, actual);
        }
    }

    fn test_roundtrip(sort_option: SortOptions, col: ArrayRef) {
        let converter = RowConverter::new(vec![SortField::new_with_options(
            col.data_type().clone(),
            sort_option,
        )])
        .unwrap();
        let rows = converter.convert_columns(&[Arc::clone(&col)]).unwrap();
        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(back.len(), 1);
        assert_eq!(&back[0], &col);
        back[0].to_data().validate_full().unwrap();
    }

    #[test]
    fn test_zero_width_fixed_size_binary_roundtrip() {
        for sort_option in all_sort_options() {
            // With a zero byte width and no nulls, the decoded length cannot be
            // inferred from the value or null buffers and must come from the row count
            for with_null in [true, false] {
                let nulls = if with_null {
                    Some(NullBuffer::from(vec![true, false, true, false, true]))
                } else {
                    None
                };
                let col: ArrayRef = Arc::new(
                    FixedSizeBinaryArray::try_new_with_len(0, Buffer::default(), nulls, 5).unwrap(),
                );

                test_roundtrip(sort_option, col);
            }
        }
    }

    #[test]
    fn test_zero_width_fixed_size_list_roundtrip() {
        for sort_option in all_sort_options() {
            // With a zero byte width and no nulls, the decoded length cannot be
            // inferred from the value or null buffers and must come from the row count
            for with_null in [true, false] {
                let nulls = if with_null {
                    Some(NullBuffer::from(vec![true, false, true, false, true]))
                } else {
                    None
                };
                let col: ArrayRef = Arc::new(
                    FixedSizeListArray::try_new_with_length(
                        Arc::new(Field::new("item", DataType::Boolean, false)),
                        0,
                        new_empty_array(&DataType::Boolean),
                        nulls,
                        5,
                    )
                    .unwrap(),
                );

                test_roundtrip(sort_option, col);
            }
        }
    }

    #[test]
    fn test_decimal32() {
        let converter = RowConverter::new(vec![SortField::new(DataType::Decimal32(
            DECIMAL32_MAX_PRECISION,
            7,
        ))])
        .unwrap();
        let col = Arc::new(
            Decimal32Array::from_iter([
                None,
                Some(i32::MIN),
                Some(-13),
                Some(46_i32),
                Some(5456_i32),
                Some(i32::MAX),
            ])
            .with_precision_and_scale(9, 7)
            .unwrap(),
        ) as ArrayRef;

        let rows = converter.convert_columns(&[Arc::clone(&col)]).unwrap();
        for i in 0..rows.num_rows() - 1 {
            assert!(rows.row(i) < rows.row(i + 1));
        }

        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(back.len(), 1);
        assert_eq!(col.as_ref(), back[0].as_ref())
    }

    #[test]
    fn test_decimal64() {
        let converter = RowConverter::new(vec![SortField::new(DataType::Decimal64(
            DECIMAL64_MAX_PRECISION,
            7,
        ))])
        .unwrap();
        let col = Arc::new(
            Decimal64Array::from_iter([
                None,
                Some(i64::MIN),
                Some(-13),
                Some(46_i64),
                Some(5456_i64),
                Some(i64::MAX),
            ])
            .with_precision_and_scale(18, 7)
            .unwrap(),
        ) as ArrayRef;

        let rows = converter.convert_columns(&[Arc::clone(&col)]).unwrap();
        for i in 0..rows.num_rows() - 1 {
            assert!(rows.row(i) < rows.row(i + 1));
        }

        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(back.len(), 1);
        assert_eq!(col.as_ref(), back[0].as_ref())
    }

    #[test]
    fn test_decimal128() {
        let converter = RowConverter::new(vec![SortField::new(DataType::Decimal128(
            DECIMAL128_MAX_PRECISION,
            7,
        ))])
        .unwrap();
        let col = Arc::new(
            Decimal128Array::from_iter([
                None,
                Some(i128::MIN),
                Some(-13),
                Some(46_i128),
                Some(5456_i128),
                Some(i128::MAX),
            ])
            .with_precision_and_scale(38, 7)
            .unwrap(),
        ) as ArrayRef;

        let rows = converter.convert_columns(&[Arc::clone(&col)]).unwrap();
        for i in 0..rows.num_rows() - 1 {
            assert!(rows.row(i) < rows.row(i + 1));
        }

        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(back.len(), 1);
        assert_eq!(col.as_ref(), back[0].as_ref())
    }

    #[test]
    fn test_decimal256() {
        let converter = RowConverter::new(vec![SortField::new(DataType::Decimal256(
            DECIMAL256_MAX_PRECISION,
            7,
        ))])
        .unwrap();
        let col = Arc::new(
            Decimal256Array::from_iter([
                None,
                Some(i256::MIN),
                Some(i256::from_parts(0, -1)),
                Some(i256::from_parts(u128::MAX, -1)),
                Some(i256::from_parts(u128::MAX, 0)),
                Some(i256::from_parts(0, 46_i128)),
                Some(i256::from_parts(5, 46_i128)),
                Some(i256::MAX),
            ])
            .with_precision_and_scale(DECIMAL256_MAX_PRECISION, 7)
            .unwrap(),
        ) as ArrayRef;

        let rows = converter.convert_columns(&[Arc::clone(&col)]).unwrap();
        for i in 0..rows.num_rows() - 1 {
            assert!(rows.row(i) < rows.row(i + 1));
        }

        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(back.len(), 1);
        assert_eq!(col.as_ref(), back[0].as_ref())
    }

    #[test]
    fn test_bool() {
        let converter = RowConverter::new(vec![SortField::new(DataType::Boolean)]).unwrap();

        let col = Arc::new(BooleanArray::from_iter([None, Some(false), Some(true)])) as ArrayRef;

        let rows = converter.convert_columns(&[Arc::clone(&col)]).unwrap();
        assert!(rows.row(2) > rows.row(1));
        assert!(rows.row(2) > rows.row(0));
        assert!(rows.row(1) > rows.row(0));

        let cols = converter.convert_rows(&rows).unwrap();
        assert_eq!(&cols[0], &col);

        let converter = RowConverter::new(vec![SortField::new_with_options(
            DataType::Boolean,
            SortOptions::default().desc().with_nulls_first(false),
        )])
        .unwrap();

        let rows = converter.convert_columns(&[Arc::clone(&col)]).unwrap();
        assert!(rows.row(2) < rows.row(1));
        assert!(rows.row(2) < rows.row(0));
        assert!(rows.row(1) < rows.row(0));
        let cols = converter.convert_rows(&rows).unwrap();
        assert_eq!(&cols[0], &col);
    }

    #[test]
    fn test_timezone() {
        let a =
            TimestampNanosecondArray::from(vec![1, 2, 3, 4, 5]).with_timezone("+01:00".to_string());
        let d = a.data_type().clone();

        let converter = RowConverter::new(vec![SortField::new(a.data_type().clone())]).unwrap();
        let rows = converter.convert_columns(&[Arc::new(a) as _]).unwrap();
        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(back.len(), 1);
        assert_eq!(back[0].data_type(), &d);

        // Test dictionary
        let mut a = PrimitiveDictionaryBuilder::<Int32Type, TimestampNanosecondType>::new();
        a.append(34).unwrap();
        a.append_null();
        a.append(345).unwrap();

        // Construct dictionary with a timezone
        let dict = a.finish();
        let values = TimestampNanosecondArray::from(dict.values().to_data());
        let dict_with_tz = dict.with_values(Arc::new(values.with_timezone("+02:00")));
        let v = DataType::Timestamp(TimeUnit::Nanosecond, Some("+02:00".into()));
        let d = DataType::Dictionary(Box::new(DataType::Int32), Box::new(v.clone()));

        assert_eq!(dict_with_tz.data_type(), &d);
        let converter = RowConverter::new(vec![SortField::new(d.clone())]).unwrap();
        let rows = converter
            .convert_columns(&[Arc::new(dict_with_tz) as _])
            .unwrap();
        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(back.len(), 1);
        assert_eq!(back[0].data_type(), &v);
    }

    #[test]
    fn test_null_encoding() {
        let col = Arc::new(NullArray::new(10));
        let converter = RowConverter::new(vec![SortField::new(DataType::Null)]).unwrap();
        let rows = converter.convert_columns(&[col]).unwrap();
        assert_eq!(rows.num_rows(), 10);
        // NULL element encoded as 2 bytes data
        assert_eq!(rows.row(1).data.len(), 2);
    }

    #[test]
    fn test_variable_width() {
        let col = Arc::new(StringArray::from_iter([
            Some("hello"),
            Some("he"),
            None,
            Some("foo"),
            Some(""),
        ])) as ArrayRef;

        let converter = RowConverter::new(vec![SortField::new(DataType::Utf8)]).unwrap();
        let rows = converter.convert_columns(&[Arc::clone(&col)]).unwrap();

        assert!(rows.row(1) < rows.row(0));
        assert!(rows.row(2) < rows.row(4));
        assert!(rows.row(3) < rows.row(0));
        assert!(rows.row(3) < rows.row(1));

        let cols = converter.convert_rows(&rows).unwrap();
        assert_eq!(&cols[0], &col);

        let col = Arc::new(BinaryArray::from_iter([
            None,
            Some(vec![0_u8; 0]),
            Some(vec![0_u8; 6]),
            Some(vec![0_u8; variable::MINI_BLOCK_SIZE]),
            Some(vec![0_u8; variable::MINI_BLOCK_SIZE + 1]),
            Some(vec![0_u8; variable::BLOCK_SIZE]),
            Some(vec![0_u8; variable::BLOCK_SIZE + 1]),
            Some(vec![1_u8; 6]),
            Some(vec![1_u8; variable::MINI_BLOCK_SIZE]),
            Some(vec![1_u8; variable::MINI_BLOCK_SIZE + 1]),
            Some(vec![1_u8; variable::BLOCK_SIZE]),
            Some(vec![1_u8; variable::BLOCK_SIZE + 1]),
            Some(vec![0xFF_u8; 6]),
            Some(vec![0xFF_u8; variable::MINI_BLOCK_SIZE]),
            Some(vec![0xFF_u8; variable::MINI_BLOCK_SIZE + 1]),
            Some(vec![0xFF_u8; variable::BLOCK_SIZE]),
            Some(vec![0xFF_u8; variable::BLOCK_SIZE + 1]),
        ])) as ArrayRef;

        let converter = RowConverter::new(vec![SortField::new(DataType::Binary)]).unwrap();
        let rows = converter.convert_columns(&[Arc::clone(&col)]).unwrap();

        for i in 0..rows.num_rows() {
            for j in i + 1..rows.num_rows() {
                assert!(
                    rows.row(i) < rows.row(j),
                    "{} < {} - {:?} < {:?}",
                    i,
                    j,
                    rows.row(i),
                    rows.row(j)
                );
            }
        }

        let cols = converter.convert_rows(&rows).unwrap();
        assert_eq!(&cols[0], &col);

        let converter = RowConverter::new(vec![SortField::new_with_options(
            DataType::Binary,
            SortOptions::default().desc().with_nulls_first(false),
        )])
        .unwrap();
        let rows = converter.convert_columns(&[Arc::clone(&col)]).unwrap();

        for i in 0..rows.num_rows() {
            for j in i + 1..rows.num_rows() {
                assert!(
                    rows.row(i) > rows.row(j),
                    "{} > {} - {:?} > {:?}",
                    i,
                    j,
                    rows.row(i),
                    rows.row(j)
                );
            }
        }

        let cols = converter.convert_rows(&rows).unwrap();
        assert_eq!(&cols[0], &col);
    }

    /// If `exact` is false performs a logical comparison between a and dictionary-encoded b
    fn dictionary_eq(a: &dyn Array, b: &dyn Array) {
        match b.data_type() {
            DataType::Dictionary(_, v) => {
                assert_eq!(a.data_type(), v.as_ref());
                let b = arrow_cast::cast(b, v).unwrap();
                assert_eq!(a, b.as_ref())
            }
            _ => assert_eq!(a, b),
        }
    }

    #[test]
    fn test_string_dictionary() {
        let a = Arc::new(DictionaryArray::<Int32Type>::from_iter([
            Some("foo"),
            Some("hello"),
            Some("he"),
            None,
            Some("hello"),
            Some(""),
            Some("hello"),
            Some("hello"),
        ])) as ArrayRef;

        let field = SortField::new(a.data_type().clone());
        let converter = RowConverter::new(vec![field]).unwrap();
        let rows_a = converter.convert_columns(&[Arc::clone(&a)]).unwrap();

        assert!(rows_a.row(3) < rows_a.row(5));
        assert!(rows_a.row(2) < rows_a.row(1));
        assert!(rows_a.row(0) < rows_a.row(1));
        assert!(rows_a.row(3) < rows_a.row(0));

        assert_eq!(rows_a.row(1), rows_a.row(4));
        assert_eq!(rows_a.row(1), rows_a.row(6));
        assert_eq!(rows_a.row(1), rows_a.row(7));

        let cols = converter.convert_rows(&rows_a).unwrap();
        dictionary_eq(&cols[0], &a);

        let b = Arc::new(DictionaryArray::<Int32Type>::from_iter([
            Some("hello"),
            None,
            Some("cupcakes"),
        ])) as ArrayRef;

        let rows_b = converter.convert_columns(&[Arc::clone(&b)]).unwrap();
        assert_eq!(rows_a.row(1), rows_b.row(0));
        assert_eq!(rows_a.row(3), rows_b.row(1));
        assert!(rows_b.row(2) < rows_a.row(0));

        let cols = converter.convert_rows(&rows_b).unwrap();
        dictionary_eq(&cols[0], &b);

        let converter = RowConverter::new(vec![SortField::new_with_options(
            a.data_type().clone(),
            SortOptions::default().desc().with_nulls_first(false),
        )])
        .unwrap();

        let rows_c = converter.convert_columns(&[Arc::clone(&a)]).unwrap();
        assert!(rows_c.row(3) > rows_c.row(5));
        assert!(rows_c.row(2) > rows_c.row(1));
        assert!(rows_c.row(0) > rows_c.row(1));
        assert!(rows_c.row(3) > rows_c.row(0));

        let cols = converter.convert_rows(&rows_c).unwrap();
        dictionary_eq(&cols[0], &a);

        let converter = RowConverter::new(vec![SortField::new_with_options(
            a.data_type().clone(),
            SortOptions::default().desc().with_nulls_first(true),
        )])
        .unwrap();

        let rows_c = converter.convert_columns(&[Arc::clone(&a)]).unwrap();
        assert!(rows_c.row(3) < rows_c.row(5));
        assert!(rows_c.row(2) > rows_c.row(1));
        assert!(rows_c.row(0) > rows_c.row(1));
        assert!(rows_c.row(3) < rows_c.row(0));

        let cols = converter.convert_rows(&rows_c).unwrap();
        dictionary_eq(&cols[0], &a);
    }

    #[test]
    fn test_struct() {
        // Test basic
        let a = Arc::new(Int32Array::from(vec![1, 1, 2, 2])) as ArrayRef;
        let a_f = Arc::new(Field::new("int", DataType::Int32, false));
        let u = Arc::new(StringArray::from(vec!["a", "b", "c", "d"])) as ArrayRef;
        let u_f = Arc::new(Field::new("s", DataType::Utf8, false));
        let s1 = Arc::new(StructArray::from(vec![(a_f, a), (u_f, u)])) as ArrayRef;

        let sort_fields = vec![SortField::new(s1.data_type().clone())];
        let converter = RowConverter::new(sort_fields).unwrap();
        let r1 = converter.convert_columns(&[Arc::clone(&s1)]).unwrap();

        for (a, b) in r1.iter().zip(r1.iter().skip(1)) {
            assert!(a < b);
        }

        let back = converter.convert_rows(&r1).unwrap();
        assert_eq!(back.len(), 1);
        assert_eq!(&back[0], &s1);

        // Test struct nullability
        let data = s1
            .to_data()
            .into_builder()
            .null_bit_buffer(Some(Buffer::from_slice_ref([0b00001010])))
            .null_count(2)
            .build()
            .unwrap();

        let s2 = Arc::new(StructArray::from(data)) as ArrayRef;
        let r2 = converter.convert_columns(&[Arc::clone(&s2)]).unwrap();
        assert_eq!(r2.row(0), r2.row(2)); // Nulls equal
        assert!(r2.row(0) < r2.row(1)); // Nulls first
        assert_ne!(r1.row(0), r2.row(0)); // Value does not equal null
        assert_eq!(r1.row(1), r2.row(1)); // Values equal

        let back = converter.convert_rows(&r2).unwrap();
        assert_eq!(back.len(), 1);
        assert_eq!(&back[0], &s2);

        back[0].to_data().validate_full().unwrap();
    }

    #[test]
    fn test_dictionary_in_struct() {
        let builder = StringDictionaryBuilder::<Int32Type>::new();
        let mut struct_builder = StructBuilder::new(
            vec![Field::new_dictionary(
                "foo",
                DataType::Int32,
                DataType::Utf8,
                true,
            )],
            vec![Box::new(builder)],
        );

        let dict_builder = struct_builder
            .field_builder::<StringDictionaryBuilder<Int32Type>>(0)
            .unwrap();

        // Flattened: ["a", null, "a", "b"]
        dict_builder.append_value("a");
        dict_builder.append_null();
        dict_builder.append_value("a");
        dict_builder.append_value("b");

        for _ in 0..4 {
            struct_builder.append(true);
        }

        let s = Arc::new(struct_builder.finish()) as ArrayRef;
        let sort_fields = vec![SortField::new(s.data_type().clone())];
        let converter = RowConverter::new(sort_fields).unwrap();
        let r = converter.convert_columns(&[Arc::clone(&s)]).unwrap();

        let back = converter.convert_rows(&r).unwrap();
        let [s2] = back.try_into().unwrap();

        // RowConverter flattens Dictionary
        // s.ty = Struct("foo": Dictionary(Int32, Utf8)), s2.ty = Struct("foo": Utf8)
        assert_ne!(&s.data_type(), &s2.data_type());
        s2.to_data().validate_full().unwrap();

        // Check if the logical data remains the same
        // Keys: [0, null, 0, 1]
        // Values: ["a", "b"]
        let s1_struct = s.as_struct();
        let s1_0 = s1_struct.column(0);
        let s1_idx_0 = s1_0.as_dictionary::<Int32Type>();
        let keys = s1_idx_0.keys();
        let values = s1_idx_0.values().as_string::<i32>();
        // Flattened: ["a", null, "a", "b"]
        let s2_struct = s2.as_struct();
        let s2_0 = s2_struct.column(0);
        let s2_idx_0 = s2_0.as_string::<i32>();

        for i in 0..keys.len() {
            if keys.is_null(i) {
                assert!(s2_idx_0.is_null(i));
            } else {
                let dict_index = keys.value(i) as usize;
                assert_eq!(values.value(dict_index), s2_idx_0.value(i));
            }
        }
    }

    #[test]
    fn test_dictionary_in_struct_empty() {
        let ty = DataType::Struct(
            vec![Field::new_dictionary(
                "foo",
                DataType::Int32,
                DataType::Int32,
                false,
            )]
            .into(),
        );
        let s = arrow_array::new_empty_array(&ty);

        let sort_fields = vec![SortField::new(s.data_type().clone())];
        let converter = RowConverter::new(sort_fields).unwrap();
        let r = converter.convert_columns(&[Arc::clone(&s)]).unwrap();

        let back = converter.convert_rows(&r).unwrap();
        let [s2] = back.try_into().unwrap();

        // RowConverter flattens Dictionary
        // s.ty = Struct("foo": Dictionary(Int32, Int32)), s2.ty = Struct("foo": Int32)
        assert_ne!(&s.data_type(), &s2.data_type());
        s2.to_data().validate_full().unwrap();
        assert_eq!(s.len(), 0);
        assert_eq!(s2.len(), 0);
    }

    #[test]
    fn test_list_of_string_dictionary() {
        let mut builder = ListBuilder::<StringDictionaryBuilder<Int32Type>>::default();
        // List[0] = ["a", "b", "zero", null, "c", "b", "d" (dict)]
        builder.values().append("a").unwrap();
        builder.values().append("b").unwrap();
        builder.values().append("zero").unwrap();
        builder.values().append_null();
        builder.values().append("c").unwrap();
        builder.values().append("b").unwrap();
        builder.values().append("d").unwrap();
        builder.append(true);
        // List[1] = null
        builder.append(false);
        // List[2] = ["e", "zero", "a" (dict)]
        builder.values().append("e").unwrap();
        builder.values().append("zero").unwrap();
        builder.values().append("a").unwrap();
        builder.append(true);

        let a = Arc::new(builder.finish()) as ArrayRef;
        let data_type = a.data_type().clone();

        let field = SortField::new(data_type.clone());
        let converter = RowConverter::new(vec![field]).unwrap();
        let rows = converter.convert_columns(&[Arc::clone(&a)]).unwrap();

        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(back.len(), 1);
        let [a2] = back.try_into().unwrap();

        // RowConverter flattens Dictionary
        // a.ty: List(Dictionary(Int32, Utf8)), a2.ty: List(Utf8)
        assert_ne!(&a.data_type(), &a2.data_type());

        a2.to_data().validate_full().unwrap();

        let a2_list = a2.as_list::<i32>();
        let a1_list = a.as_list::<i32>();

        // Check if the logical data remains the same
        // List[0] = ["a", "b", "zero", null, "c", "b", "d" (dict)]
        let a1_0 = a1_list.value(0);
        let a1_idx_0 = a1_0.as_dictionary::<Int32Type>();
        let keys = a1_idx_0.keys();
        let values = a1_idx_0.values().as_string::<i32>();
        let a2_0 = a2_list.value(0);
        let a2_idx_0 = a2_0.as_string::<i32>();

        for i in 0..keys.len() {
            if keys.is_null(i) {
                assert!(a2_idx_0.is_null(i));
            } else {
                let dict_index = keys.value(i) as usize;
                assert_eq!(values.value(dict_index), a2_idx_0.value(i));
            }
        }

        // List[1] = null
        assert!(a1_list.is_null(1));
        assert!(a2_list.is_null(1));

        // List[2] = ["e", "zero", "a" (dict)]
        let a1_2 = a1_list.value(2);
        let a1_idx_2 = a1_2.as_dictionary::<Int32Type>();
        let keys = a1_idx_2.keys();
        let values = a1_idx_2.values().as_string::<i32>();
        let a2_2 = a2_list.value(2);
        let a2_idx_2 = a2_2.as_string::<i32>();

        for i in 0..keys.len() {
            if keys.is_null(i) {
                assert!(a2_idx_2.is_null(i));
            } else {
                let dict_index = keys.value(i) as usize;
                assert_eq!(values.value(dict_index), a2_idx_2.value(i));
            }
        }
    }

    #[test]
    fn test_primitive_dictionary() {
        let mut builder = PrimitiveDictionaryBuilder::<Int32Type, Int32Type>::new();
        builder.append(2).unwrap();
        builder.append(3).unwrap();
        builder.append(0).unwrap();
        builder.append_null();
        builder.append(5).unwrap();
        builder.append(3).unwrap();
        builder.append(-1).unwrap();

        let a = builder.finish();
        let data_type = a.data_type().clone();
        let columns = [Arc::new(a) as ArrayRef];

        let field = SortField::new(data_type.clone());
        let converter = RowConverter::new(vec![field]).unwrap();
        let rows = converter.convert_columns(&columns).unwrap();
        assert!(rows.row(0) < rows.row(1));
        assert!(rows.row(2) < rows.row(0));
        assert!(rows.row(3) < rows.row(2));
        assert!(rows.row(6) < rows.row(2));
        assert!(rows.row(3) < rows.row(6));

        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(back.len(), 1);
        back[0].to_data().validate_full().unwrap();
    }

    #[test]
    fn test_dictionary_nulls() {
        let values = Int32Array::from_iter([Some(1), Some(-1), None, Some(4), None]).into_data();
        let keys =
            Int32Array::from_iter([Some(0), Some(0), Some(1), Some(2), Some(4), None]).into_data();

        let data_type = DataType::Dictionary(Box::new(DataType::Int32), Box::new(DataType::Int32));
        let data = keys
            .into_builder()
            .data_type(data_type.clone())
            .child_data(vec![values])
            .build()
            .unwrap();

        let columns = [Arc::new(DictionaryArray::<Int32Type>::from(data)) as ArrayRef];
        let field = SortField::new(data_type.clone());
        let converter = RowConverter::new(vec![field]).unwrap();
        let rows = converter.convert_columns(&columns).unwrap();

        assert_eq!(rows.row(0), rows.row(1));
        assert_eq!(rows.row(3), rows.row(4));
        assert_eq!(rows.row(4), rows.row(5));
        assert!(rows.row(3) < rows.row(0));
    }

    #[test]
    fn test_from_binary_shared_buffer() {
        let converter = RowConverter::new(vec![SortField::new(DataType::Binary)]).unwrap();
        let array = Arc::new(BinaryArray::from_iter_values([&[0xFF]])) as _;
        let rows = converter.convert_columns(&[array]).unwrap();
        let binary_rows = rows.try_into_binary().expect("known-small rows");
        let _binary_rows_shared_buffer = binary_rows.clone();

        let parsed = converter.from_binary(binary_rows);

        converter.convert_rows(parsed.iter()).unwrap();
    }

    #[test]
    #[should_panic(expected = "Encountered non UTF-8 data")]
    fn test_invalid_utf8() {
        let converter = RowConverter::new(vec![SortField::new(DataType::Binary)]).unwrap();
        let array = Arc::new(BinaryArray::from_iter_values([&[0xFF]])) as _;
        let rows = converter.convert_columns(&[array]).unwrap();
        let binary_row = rows.row(0);

        let converter = RowConverter::new(vec![SortField::new(DataType::Utf8)]).unwrap();
        let parser = converter.parser();
        let utf8_row = parser.parse(binary_row.as_ref());

        converter.convert_rows(std::iter::once(utf8_row)).unwrap();
    }

    #[test]
    #[should_panic(expected = "Encountered non UTF-8 data")]
    fn test_invalid_utf8_array() {
        let converter = RowConverter::new(vec![SortField::new(DataType::Binary)]).unwrap();
        let array = Arc::new(BinaryArray::from_iter_values([&[0xFF]])) as _;
        let rows = converter.convert_columns(&[array]).unwrap();
        let binary_rows = rows.try_into_binary().expect("known-small rows");

        let converter = RowConverter::new(vec![SortField::new(DataType::Utf8)]).unwrap();
        let parsed = converter.from_binary(binary_rows);

        converter.convert_rows(parsed.iter()).unwrap();
    }

    #[test]
    #[should_panic(expected = "index out of bounds")]
    fn test_invalid_empty() {
        let binary_row: &[u8] = &[];

        let converter = RowConverter::new(vec![SortField::new(DataType::Utf8)]).unwrap();
        let parser = converter.parser();
        let utf8_row = parser.parse(binary_row.as_ref());

        converter.convert_rows(std::iter::once(utf8_row)).unwrap();
    }

    #[test]
    #[should_panic(expected = "index out of bounds")]
    fn test_invalid_empty_array() {
        let row: &[u8] = &[];
        let binary_rows = BinaryArray::from(vec![row]);

        let converter = RowConverter::new(vec![SortField::new(DataType::Utf8)]).unwrap();
        let parsed = converter.from_binary(binary_rows);

        converter.convert_rows(parsed.iter()).unwrap();
    }

    #[test]
    #[should_panic(expected = "index out of bounds")]
    fn test_invalid_truncated() {
        let binary_row: &[u8] = &[0x02];

        let converter = RowConverter::new(vec![SortField::new(DataType::Utf8)]).unwrap();
        let parser = converter.parser();
        let utf8_row = parser.parse(binary_row.as_ref());

        converter.convert_rows(std::iter::once(utf8_row)).unwrap();
    }

    #[test]
    #[should_panic(expected = "index out of bounds")]
    fn test_invalid_truncated_array() {
        let row: &[u8] = &[0x02];
        let binary_rows = BinaryArray::from(vec![row]);

        let converter = RowConverter::new(vec![SortField::new(DataType::Utf8)]).unwrap();
        let parsed = converter.from_binary(binary_rows);

        converter.convert_rows(parsed.iter()).unwrap();
    }

    #[test]
    #[should_panic(expected = "rows were not produced by this RowConverter")]
    fn test_different_converter() {
        let values = Arc::new(Int32Array::from_iter([Some(1), Some(-1)]));
        let converter = RowConverter::new(vec![SortField::new(DataType::Int32)]).unwrap();
        let rows = converter.convert_columns(&[values]).unwrap();

        let converter = RowConverter::new(vec![SortField::new(DataType::Int32)]).unwrap();
        let _ = converter.convert_rows(&rows);
    }

    fn test_single_list<O: OffsetSizeTrait>() {
        let mut builder = GenericListBuilder::<O, _>::new(Int32Builder::new());
        builder.values().append_value(32);
        builder.values().append_value(52);
        builder.values().append_value(32);
        builder.append(true);
        builder.values().append_value(32);
        builder.values().append_value(52);
        builder.values().append_value(12);
        builder.append(true);
        builder.values().append_value(32);
        builder.values().append_value(52);
        builder.append(true);
        builder.values().append_value(32); // MASKED
        builder.values().append_value(52); // MASKED
        builder.append(false);
        builder.values().append_value(32);
        builder.values().append_null();
        builder.append(true);
        builder.append(true);
        builder.values().append_value(17); // MASKED
        builder.values().append_null(); // MASKED
        builder.append(false);

        let list = Arc::new(builder.finish()) as ArrayRef;
        let d = list.data_type().clone();

        let converter = RowConverter::new(vec![SortField::new(d.clone())]).unwrap();

        let rows = converter.convert_columns(&[Arc::clone(&list)]).unwrap();
        assert!(rows.row(0) > rows.row(1)); // [32, 52, 32] > [32, 52, 12]
        assert!(rows.row(2) < rows.row(1)); // [32, 52] < [32, 52, 12]
        assert!(rows.row(3) < rows.row(2)); // null < [32, 52]
        assert!(rows.row(4) < rows.row(2)); // [32, null] < [32, 52]
        assert!(rows.row(5) < rows.row(2)); // [] < [32, 52]
        assert!(rows.row(3) < rows.row(5)); // null < []
        assert_eq!(rows.row(3), rows.row(6)); // null = null (different masked values)

        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(back.len(), 1);
        back[0].to_data().validate_full().unwrap();
        assert_eq!(&back[0], &list);

        let options = SortOptions::default().asc().with_nulls_first(false);
        let field = SortField::new_with_options(d.clone(), options);
        let converter = RowConverter::new(vec![field]).unwrap();
        let rows = converter.convert_columns(&[Arc::clone(&list)]).unwrap();

        assert!(rows.row(0) > rows.row(1)); // [32, 52, 32] > [32, 52, 12]
        assert!(rows.row(2) < rows.row(1)); // [32, 52] < [32, 52, 12]
        assert!(rows.row(3) > rows.row(2)); // null > [32, 52]
        assert!(rows.row(4) > rows.row(2)); // [32, null] > [32, 52]
        assert!(rows.row(5) < rows.row(2)); // [] < [32, 52]
        assert!(rows.row(3) > rows.row(5)); // null > []
        assert_eq!(rows.row(3), rows.row(6)); // null = null (different masked values)

        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(back.len(), 1);
        back[0].to_data().validate_full().unwrap();
        assert_eq!(&back[0], &list);

        let options = SortOptions::default().desc().with_nulls_first(false);
        let field = SortField::new_with_options(d.clone(), options);
        let converter = RowConverter::new(vec![field]).unwrap();
        let rows = converter.convert_columns(&[Arc::clone(&list)]).unwrap();

        assert!(rows.row(0) < rows.row(1)); // [32, 52, 32] < [32, 52, 12]
        assert!(rows.row(2) > rows.row(1)); // [32, 52] > [32, 52, 12]
        assert!(rows.row(3) > rows.row(2)); // null > [32, 52]
        assert!(rows.row(4) > rows.row(2)); // [32, null] > [32, 52]
        assert!(rows.row(5) > rows.row(2)); // [] > [32, 52]
        assert!(rows.row(3) > rows.row(5)); // null > []
        assert_eq!(rows.row(3), rows.row(6)); // null = null (different masked values)

        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(back.len(), 1);
        back[0].to_data().validate_full().unwrap();
        assert_eq!(&back[0], &list);

        let options = SortOptions::default().desc().with_nulls_first(true);
        let field = SortField::new_with_options(d, options);
        let converter = RowConverter::new(vec![field]).unwrap();
        let rows = converter.convert_columns(&[Arc::clone(&list)]).unwrap();

        assert!(rows.row(0) < rows.row(1)); // [32, 52, 32] < [32, 52, 12]
        assert!(rows.row(2) > rows.row(1)); // [32, 52] > [32, 52, 12]
        assert!(rows.row(3) < rows.row(2)); // null < [32, 52]
        assert!(rows.row(4) < rows.row(2)); // [32, null] < [32, 52]
        assert!(rows.row(5) > rows.row(2)); // [] > [32, 52]
        assert!(rows.row(3) < rows.row(5)); // null < []
        assert_eq!(rows.row(3), rows.row(6)); // null = null (different masked values)

        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(back.len(), 1);
        back[0].to_data().validate_full().unwrap();
        assert_eq!(&back[0], &list);

        let sliced_list = list.slice(1, 5);
        let rows_on_sliced_list = converter
            .convert_columns(&[Arc::clone(&sliced_list)])
            .unwrap();

        assert!(rows_on_sliced_list.row(1) > rows_on_sliced_list.row(0)); // [32, 52] > [32, 52, 12]
        assert!(rows_on_sliced_list.row(2) < rows_on_sliced_list.row(1)); // null < [32, 52]
        assert!(rows_on_sliced_list.row(3) < rows_on_sliced_list.row(1)); // [32, null] < [32, 52]
        assert!(rows_on_sliced_list.row(4) > rows_on_sliced_list.row(1)); // [] > [32, 52]
        assert!(rows_on_sliced_list.row(2) < rows_on_sliced_list.row(4)); // null < []

        let back = converter.convert_rows(&rows_on_sliced_list).unwrap();
        assert_eq!(back.len(), 1);
        back[0].to_data().validate_full().unwrap();
        assert_eq!(&back[0], &sliced_list);
    }

    fn test_nested_list<O: OffsetSizeTrait>() {
        let mut builder =
            GenericListBuilder::<O, _>::new(GenericListBuilder::<O, _>::new(Int32Builder::new()));

        builder.values().values().append_value(1);
        builder.values().values().append_value(2);
        builder.values().append(true);
        builder.values().values().append_value(1);
        builder.values().values().append_null();
        builder.values().append(true);
        builder.append(true);

        builder.values().values().append_value(1);
        builder.values().values().append_null();
        builder.values().append(true);
        builder.values().values().append_value(1);
        builder.values().values().append_null();
        builder.values().append(true);
        builder.append(true);

        builder.values().values().append_value(1);
        builder.values().values().append_null();
        builder.values().append(true);
        builder.values().append(false);
        builder.append(true);
        builder.append(false);

        builder.values().values().append_value(1);
        builder.values().values().append_value(2);
        builder.values().append(true);
        builder.append(true);

        let list = Arc::new(builder.finish()) as ArrayRef;
        let d = list.data_type().clone();

        // [
        //   [[1, 2], [1, null]],
        //   [[1, null], [1, null]],
        //   [[1, null], null]
        //   null
        //   [[1, 2]]
        // ]
        let options = SortOptions::default().asc().with_nulls_first(true);
        let field = SortField::new_with_options(d.clone(), options);
        let converter = RowConverter::new(vec![field]).unwrap();
        let rows = converter.convert_columns(&[Arc::clone(&list)]).unwrap();

        assert!(rows.row(0) > rows.row(1));
        assert!(rows.row(1) > rows.row(2));
        assert!(rows.row(2) > rows.row(3));
        assert!(rows.row(4) < rows.row(0));
        assert!(rows.row(4) > rows.row(1));

        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(back.len(), 1);
        back[0].to_data().validate_full().unwrap();
        assert_eq!(&back[0], &list);

        let options = SortOptions::default().desc().with_nulls_first(true);
        let field = SortField::new_with_options(d.clone(), options);
        let converter = RowConverter::new(vec![field]).unwrap();
        let rows = converter.convert_columns(&[Arc::clone(&list)]).unwrap();

        assert!(rows.row(0) > rows.row(1));
        assert!(rows.row(1) > rows.row(2));
        assert!(rows.row(2) > rows.row(3));
        assert!(rows.row(4) > rows.row(0));
        assert!(rows.row(4) > rows.row(1));

        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(back.len(), 1);
        back[0].to_data().validate_full().unwrap();
        assert_eq!(&back[0], &list);

        let options = SortOptions::default().desc().with_nulls_first(false);
        let field = SortField::new_with_options(d, options);
        let converter = RowConverter::new(vec![field]).unwrap();
        let rows = converter.convert_columns(&[Arc::clone(&list)]).unwrap();

        assert!(rows.row(0) < rows.row(1));
        assert!(rows.row(1) < rows.row(2));
        assert!(rows.row(2) < rows.row(3));
        assert!(rows.row(4) > rows.row(0));
        assert!(rows.row(4) < rows.row(1));

        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(back.len(), 1);
        back[0].to_data().validate_full().unwrap();
        assert_eq!(&back[0], &list);

        let sliced_list = list.slice(1, 3);
        let rows = converter
            .convert_columns(&[Arc::clone(&sliced_list)])
            .unwrap();

        assert!(rows.row(0) < rows.row(1));
        assert!(rows.row(1) < rows.row(2));

        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(back.len(), 1);
        back[0].to_data().validate_full().unwrap();
        assert_eq!(&back[0], &sliced_list);
    }

    #[test]
    fn test_list() {
        test_single_list::<i32>();
        test_nested_list::<i32>();
    }

    #[test]
    fn test_large_list() {
        test_single_list::<i64>();
        test_nested_list::<i64>();
    }

    fn test_single_list_view<O: OffsetSizeTrait>() {
        let mut builder = GenericListViewBuilder::<O, _>::new(Int32Builder::new());
        builder.values().append_value(32);
        builder.values().append_value(52);
        builder.values().append_value(32);
        builder.append(true);
        builder.values().append_value(32);
        builder.values().append_value(52);
        builder.values().append_value(12);
        builder.append(true);
        builder.values().append_value(32);
        builder.values().append_value(52);
        builder.append(true);
        builder.values().append_value(32); // MASKED
        builder.values().append_value(52); // MASKED
        builder.append(false);
        builder.values().append_value(32);
        builder.values().append_null();
        builder.append(true);
        builder.append(true);
        builder.values().append_value(17); // MASKED
        builder.values().append_null(); // MASKED
        builder.append(false);

        let list = Arc::new(builder.finish()) as ArrayRef;
        let d = list.data_type().clone();

        let converter = RowConverter::new(vec![SortField::new(d.clone())]).unwrap();

        let rows = converter.convert_columns(&[Arc::clone(&list)]).unwrap();
        assert!(rows.row(0) > rows.row(1)); // [32, 52, 32] > [32, 52, 12]
        assert!(rows.row(2) < rows.row(1)); // [32, 52] < [32, 52, 12]
        assert!(rows.row(3) < rows.row(2)); // null < [32, 52]
        assert!(rows.row(4) < rows.row(2)); // [32, null] < [32, 52]
        assert!(rows.row(5) < rows.row(2)); // [] < [32, 52]
        assert!(rows.row(3) < rows.row(5)); // null < []
        assert_eq!(rows.row(3), rows.row(6)); // null = null (different masked values)

        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(back.len(), 1);
        back[0].to_data().validate_full().unwrap();

        // Verify the content matches (ListView may have different physical layout but same logical content)
        let back_list_view = back[0]
            .as_any()
            .downcast_ref::<GenericListViewArray<O>>()
            .unwrap();
        let orig_list_view = list
            .as_any()
            .downcast_ref::<GenericListViewArray<O>>()
            .unwrap();

        assert_eq!(back_list_view.len(), orig_list_view.len());
        for i in 0..back_list_view.len() {
            assert_eq!(back_list_view.is_valid(i), orig_list_view.is_valid(i));
            if back_list_view.is_valid(i) {
                assert_eq!(&back_list_view.value(i), &orig_list_view.value(i));
            }
        }

        let options = SortOptions::default().asc().with_nulls_first(false);
        let field = SortField::new_with_options(d.clone(), options);
        let converter = RowConverter::new(vec![field]).unwrap();
        let rows = converter.convert_columns(&[Arc::clone(&list)]).unwrap();

        assert!(rows.row(0) > rows.row(1)); // [32, 52, 32] > [32, 52, 12]
        assert!(rows.row(2) < rows.row(1)); // [32, 52] < [32, 52, 12]
        assert!(rows.row(3) > rows.row(2)); // null > [32, 52]
        assert!(rows.row(4) > rows.row(2)); // [32, null] > [32, 52]
        assert!(rows.row(5) < rows.row(2)); // [] < [32, 52]
        assert!(rows.row(3) > rows.row(5)); // null > []
        assert_eq!(rows.row(3), rows.row(6)); // null = null (different masked values)

        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(back.len(), 1);
        back[0].to_data().validate_full().unwrap();

        let options = SortOptions::default().desc().with_nulls_first(false);
        let field = SortField::new_with_options(d.clone(), options);
        let converter = RowConverter::new(vec![field]).unwrap();
        let rows = converter.convert_columns(&[Arc::clone(&list)]).unwrap();

        assert!(rows.row(0) < rows.row(1)); // [32, 52, 32] < [32, 52, 12]
        assert!(rows.row(2) > rows.row(1)); // [32, 52] > [32, 52, 12]
        assert!(rows.row(3) > rows.row(2)); // null > [32, 52]
        assert!(rows.row(4) > rows.row(2)); // [32, null] > [32, 52]
        assert!(rows.row(5) > rows.row(2)); // [] > [32, 52]
        assert!(rows.row(3) > rows.row(5)); // null > []
        assert_eq!(rows.row(3), rows.row(6)); // null = null (different masked values)

        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(back.len(), 1);
        back[0].to_data().validate_full().unwrap();

        let options = SortOptions::default().desc().with_nulls_first(true);
        let field = SortField::new_with_options(d, options);
        let converter = RowConverter::new(vec![field]).unwrap();
        let rows = converter.convert_columns(&[Arc::clone(&list)]).unwrap();

        assert!(rows.row(0) < rows.row(1)); // [32, 52, 32] < [32, 52, 12]
        assert!(rows.row(2) > rows.row(1)); // [32, 52] > [32, 52, 12]
        assert!(rows.row(3) < rows.row(2)); // null < [32, 52]
        assert!(rows.row(4) < rows.row(2)); // [32, null] < [32, 52]
        assert!(rows.row(5) > rows.row(2)); // [] > [32, 52]
        assert!(rows.row(3) < rows.row(5)); // null < []
        assert_eq!(rows.row(3), rows.row(6)); // null = null (different masked values)

        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(back.len(), 1);
        back[0].to_data().validate_full().unwrap();

        let sliced_list = list.slice(1, 5);
        let rows_on_sliced_list = converter
            .convert_columns(&[Arc::clone(&sliced_list)])
            .unwrap();

        assert!(rows_on_sliced_list.row(1) > rows_on_sliced_list.row(0)); // [32, 52] > [32, 52, 12]
        assert!(rows_on_sliced_list.row(2) < rows_on_sliced_list.row(1)); // null < [32, 52]
        assert!(rows_on_sliced_list.row(3) < rows_on_sliced_list.row(1)); // [32, null] < [32, 52]
        assert!(rows_on_sliced_list.row(4) > rows_on_sliced_list.row(1)); // [] > [32, 52]
        assert!(rows_on_sliced_list.row(2) < rows_on_sliced_list.row(4)); // null < []

        let back = converter.convert_rows(&rows_on_sliced_list).unwrap();
        assert_eq!(back.len(), 1);
        back[0].to_data().validate_full().unwrap();
    }

    fn test_nested_list_view<O: OffsetSizeTrait>() {
        let mut builder = GenericListViewBuilder::<O, _>::new(GenericListViewBuilder::<O, _>::new(
            Int32Builder::new(),
        ));

        // Row 0: [[1, 2], [1, null]]
        builder.values().values().append_value(1);
        builder.values().values().append_value(2);
        builder.values().append(true);
        builder.values().values().append_value(1);
        builder.values().values().append_null();
        builder.values().append(true);
        builder.append(true);

        // Row 1: [[1, null], [1, null]]
        builder.values().values().append_value(1);
        builder.values().values().append_null();
        builder.values().append(true);
        builder.values().values().append_value(1);
        builder.values().values().append_null();
        builder.values().append(true);
        builder.append(true);

        // Row 2: [[1, null], null]
        builder.values().values().append_value(1);
        builder.values().values().append_null();
        builder.values().append(true);
        builder.values().append(false);
        builder.append(true);

        // Row 3: null
        builder.append(false);

        // Row 4: [[1, 2]]
        builder.values().values().append_value(1);
        builder.values().values().append_value(2);
        builder.values().append(true);
        builder.append(true);

        let list = Arc::new(builder.finish()) as ArrayRef;
        let d = list.data_type().clone();

        // [
        //   [[1, 2], [1, null]],
        //   [[1, null], [1, null]],
        //   [[1, null], null]
        //   null
        //   [[1, 2]]
        // ]
        let options = SortOptions::default().asc().with_nulls_first(true);
        let field = SortField::new_with_options(d.clone(), options);
        let converter = RowConverter::new(vec![field]).unwrap();
        let rows = converter.convert_columns(&[Arc::clone(&list)]).unwrap();

        assert!(rows.row(0) > rows.row(1));
        assert!(rows.row(1) > rows.row(2));
        assert!(rows.row(2) > rows.row(3));
        assert!(rows.row(4) < rows.row(0));
        assert!(rows.row(4) > rows.row(1));

        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(back.len(), 1);
        back[0].to_data().validate_full().unwrap();

        // Verify the content matches (ListView may have different physical layout but same logical content)
        let back_list_view = back[0]
            .as_any()
            .downcast_ref::<GenericListViewArray<O>>()
            .unwrap();
        let orig_list_view = list
            .as_any()
            .downcast_ref::<GenericListViewArray<O>>()
            .unwrap();

        assert_eq!(back_list_view.len(), orig_list_view.len());
        for i in 0..back_list_view.len() {
            assert_eq!(back_list_view.is_valid(i), orig_list_view.is_valid(i));
            if back_list_view.is_valid(i) {
                assert_eq!(&back_list_view.value(i), &orig_list_view.value(i));
            }
        }

        let options = SortOptions::default().desc().with_nulls_first(true);
        let field = SortField::new_with_options(d.clone(), options);
        let converter = RowConverter::new(vec![field]).unwrap();
        let rows = converter.convert_columns(&[Arc::clone(&list)]).unwrap();

        assert!(rows.row(0) > rows.row(1));
        assert!(rows.row(1) > rows.row(2));
        assert!(rows.row(2) > rows.row(3));
        assert!(rows.row(4) > rows.row(0));
        assert!(rows.row(4) > rows.row(1));

        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(back.len(), 1);
        back[0].to_data().validate_full().unwrap();

        // Verify the content matches
        let back_list_view = back[0]
            .as_any()
            .downcast_ref::<GenericListViewArray<O>>()
            .unwrap();

        assert_eq!(back_list_view.len(), orig_list_view.len());
        for i in 0..back_list_view.len() {
            assert_eq!(back_list_view.is_valid(i), orig_list_view.is_valid(i));
            if back_list_view.is_valid(i) {
                assert_eq!(&back_list_view.value(i), &orig_list_view.value(i));
            }
        }

        let options = SortOptions::default().desc().with_nulls_first(false);
        let field = SortField::new_with_options(d.clone(), options);
        let converter = RowConverter::new(vec![field]).unwrap();
        let rows = converter.convert_columns(&[Arc::clone(&list)]).unwrap();

        assert!(rows.row(0) < rows.row(1));
        assert!(rows.row(1) < rows.row(2));
        assert!(rows.row(2) < rows.row(3));
        assert!(rows.row(4) > rows.row(0));
        assert!(rows.row(4) < rows.row(1));

        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(back.len(), 1);
        back[0].to_data().validate_full().unwrap();

        // Verify the content matches
        let back_list_view = back[0]
            .as_any()
            .downcast_ref::<GenericListViewArray<O>>()
            .unwrap();

        assert_eq!(back_list_view.len(), orig_list_view.len());
        for i in 0..back_list_view.len() {
            assert_eq!(back_list_view.is_valid(i), orig_list_view.is_valid(i));
            if back_list_view.is_valid(i) {
                assert_eq!(&back_list_view.value(i), &orig_list_view.value(i));
            }
        }

        let sliced_list = list.slice(1, 3);
        let rows = converter
            .convert_columns(&[Arc::clone(&sliced_list)])
            .unwrap();

        assert!(rows.row(0) < rows.row(1));
        assert!(rows.row(1) < rows.row(2));

        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(back.len(), 1);
        back[0].to_data().validate_full().unwrap();
    }

    #[test]
    fn test_list_view() {
        test_single_list_view::<i32>();
        test_nested_list_view::<i32>();
    }

    #[test]
    fn test_large_list_view() {
        test_single_list_view::<i64>();
        test_nested_list_view::<i64>();
    }

    fn test_list_view_with_shared_values<O: OffsetSizeTrait>() {
        // Create a values array: [1, 2, 3, 4, 5, 6, 7, 8]
        let values = Int32Array::from(vec![1, 2, 3, 4, 5, 6, 7, 8]);
        let field = Arc::new(Field::new_list_field(DataType::Int32, true));

        // Create a ListView where:
        // - Row 0: offset=0, size=3 -> [1, 2, 3]
        // - Row 1: offset=0, size=3 -> [1, 2, 3] (same offset+size as row 0)
        // - Row 2: offset=5, size=2 -> [6, 7] (non-monotonic offset)
        // - Row 3: offset=2, size=2 -> [3, 4] (offset goes back)
        // - Row 4: offset=1, size=4 -> [2, 3, 4, 5] (subset of values that contains row 3's range)
        // - Row 5: offset=2, size=1 -> [3] (subset of row 3 and row 4)
        let offsets = ScalarBuffer::<O>::from(vec![
            O::from_usize(0).unwrap(),
            O::from_usize(0).unwrap(),
            O::from_usize(5).unwrap(),
            O::from_usize(2).unwrap(),
            O::from_usize(1).unwrap(),
            O::from_usize(2).unwrap(),
        ]);
        let sizes = ScalarBuffer::<O>::from(vec![
            O::from_usize(3).unwrap(),
            O::from_usize(3).unwrap(),
            O::from_usize(2).unwrap(),
            O::from_usize(2).unwrap(),
            O::from_usize(4).unwrap(),
            O::from_usize(1).unwrap(),
        ]);

        let list_view: GenericListViewArray<O> =
            GenericListViewArray::try_new(field, offsets, sizes, Arc::new(values), None).unwrap();

        let d = list_view.data_type().clone();
        let list = Arc::new(list_view) as ArrayRef;

        let converter = RowConverter::new(vec![SortField::new(d.clone())]).unwrap();
        let rows = converter.convert_columns(&[Arc::clone(&list)]).unwrap();

        // Row 0 and Row 1 have the same content [1, 2, 3], so they should be equal
        assert_eq!(rows.row(0), rows.row(1));

        // [1, 2, 3] < [6, 7] (comparing first elements: 1 < 6)
        assert!(rows.row(0) < rows.row(2));

        // [3, 4] > [1, 2, 3] (comparing first elements: 3 > 1)
        assert!(rows.row(3) > rows.row(0));

        // [2, 3, 4, 5] > [1, 2, 3] (comparing first elements: 2 > 1)
        assert!(rows.row(4) > rows.row(0));

        // [3] < [3, 4] (same prefix but shorter)
        assert!(rows.row(5) < rows.row(3));

        // [3] < [2, 3, 4, 5] (comparing first elements: 3 > 2)
        assert!(rows.row(5) > rows.row(4));

        // Round-trip conversion
        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(back.len(), 1);
        back[0].to_data().validate_full().unwrap();

        // Verify logical content matches
        let back_list_view = back[0]
            .as_any()
            .downcast_ref::<GenericListViewArray<O>>()
            .unwrap();
        let orig_list_view = list
            .as_any()
            .downcast_ref::<GenericListViewArray<O>>()
            .unwrap();

        assert_eq!(back_list_view.len(), orig_list_view.len());
        for i in 0..back_list_view.len() {
            assert_eq!(back_list_view.is_valid(i), orig_list_view.is_valid(i));
            if back_list_view.is_valid(i) {
                assert_eq!(&back_list_view.value(i), &orig_list_view.value(i));
            }
        }

        // Test with descending order
        let options = SortOptions::default().desc();
        let field = SortField::new_with_options(d, options);
        let converter = RowConverter::new(vec![field]).unwrap();
        let rows = converter.convert_columns(&[Arc::clone(&list)]).unwrap();

        // In descending order, comparisons are reversed
        assert_eq!(rows.row(0), rows.row(1)); // Equal rows stay equal
        assert!(rows.row(0) > rows.row(2)); // [1, 2, 3] > [6, 7] in desc
        assert!(rows.row(3) < rows.row(0)); // [3, 4] < [1, 2, 3] in desc

        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(back.len(), 1);
        back[0].to_data().validate_full().unwrap();
    }

    #[test]
    fn test_list_view_shared_values() {
        test_list_view_with_shared_values::<i32>();
    }

    #[test]
    fn test_large_list_view_shared_values() {
        test_list_view_with_shared_values::<i64>();
    }

    #[test]
    fn test_fixed_size_list() {
        let mut builder = FixedSizeListBuilder::new(Int32Builder::new(), 3);
        builder.values().append_value(32);
        builder.values().append_value(52);
        builder.values().append_value(32);
        builder.append(true);
        builder.values().append_value(32);
        builder.values().append_value(52);
        builder.values().append_value(12);
        builder.append(true);
        builder.values().append_value(32);
        builder.values().append_value(52);
        builder.values().append_null();
        builder.append(true);
        builder.values().append_value(32); // MASKED
        builder.values().append_value(52); // MASKED
        builder.values().append_value(13); // MASKED
        builder.append(false);
        builder.values().append_value(32);
        builder.values().append_null();
        builder.values().append_null();
        builder.append(true);
        builder.values().append_null();
        builder.values().append_null();
        builder.values().append_null();
        builder.append(true);
        builder.values().append_value(17); // MASKED
        builder.values().append_null(); // MASKED
        builder.values().append_value(77); // MASKED
        builder.append(false);

        let list = Arc::new(builder.finish()) as ArrayRef;
        let d = list.data_type().clone();

        // Default sorting (ascending, nulls first)
        let converter = RowConverter::new(vec![SortField::new(d.clone())]).unwrap();

        let rows = converter.convert_columns(&[Arc::clone(&list)]).unwrap();
        assert!(rows.row(0) > rows.row(1)); // [32, 52, 32] > [32, 52, 12]
        assert!(rows.row(2) < rows.row(1)); // [32, 52, null] < [32, 52, 12]
        assert!(rows.row(3) < rows.row(2)); // null < [32, 52, null]
        assert!(rows.row(4) < rows.row(2)); // [32, null, null] < [32, 52, null]
        assert!(rows.row(5) < rows.row(2)); // [null, null, null] < [32, 52, null]
        assert!(rows.row(3) < rows.row(5)); // null < [null, null, null]
        assert_eq!(rows.row(3), rows.row(6)); // null = null (different masked values)

        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(back.len(), 1);
        back[0].to_data().validate_full().unwrap();
        assert_eq!(&back[0], &list);

        // Ascending, null last
        let options = SortOptions::default().asc().with_nulls_first(false);
        let field = SortField::new_with_options(d.clone(), options);
        let converter = RowConverter::new(vec![field]).unwrap();
        let rows = converter.convert_columns(&[Arc::clone(&list)]).unwrap();
        assert!(rows.row(0) > rows.row(1)); // [32, 52, 32] > [32, 52, 12]
        assert!(rows.row(2) > rows.row(1)); // [32, 52, null] > [32, 52, 12]
        assert!(rows.row(3) > rows.row(2)); // null > [32, 52, null]
        assert!(rows.row(4) > rows.row(2)); // [32, null, null] > [32, 52, null]
        assert!(rows.row(5) > rows.row(2)); // [null, null, null] > [32, 52, null]
        assert!(rows.row(3) > rows.row(5)); // null > [null, null, null]
        assert_eq!(rows.row(3), rows.row(6)); // null = null (different masked values)

        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(back.len(), 1);
        back[0].to_data().validate_full().unwrap();
        assert_eq!(&back[0], &list);

        // Descending, nulls last
        let options = SortOptions::default().desc().with_nulls_first(false);
        let field = SortField::new_with_options(d.clone(), options);
        let converter = RowConverter::new(vec![field]).unwrap();
        let rows = converter.convert_columns(&[Arc::clone(&list)]).unwrap();
        assert!(rows.row(0) < rows.row(1)); // [32, 52, 32] < [32, 52, 12]
        assert!(rows.row(2) > rows.row(1)); // [32, 52, null] > [32, 52, 12]
        assert!(rows.row(3) > rows.row(2)); // null > [32, 52, null]
        assert!(rows.row(4) > rows.row(2)); // [32, null, null] > [32, 52, null]
        assert!(rows.row(5) > rows.row(2)); // [null, null, null] > [32, 52, null]
        assert!(rows.row(3) > rows.row(5)); // null > [null, null, null]
        assert_eq!(rows.row(3), rows.row(6)); // null = null (different masked values)

        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(back.len(), 1);
        back[0].to_data().validate_full().unwrap();
        assert_eq!(&back[0], &list);

        // Descending, nulls first
        let options = SortOptions::default().desc().with_nulls_first(true);
        let field = SortField::new_with_options(d, options);
        let converter = RowConverter::new(vec![field]).unwrap();
        let rows = converter.convert_columns(&[Arc::clone(&list)]).unwrap();

        assert!(rows.row(0) < rows.row(1)); // [32, 52, 32] < [32, 52, 12]
        assert!(rows.row(2) < rows.row(1)); // [32, 52, null] > [32, 52, 12]
        assert!(rows.row(3) < rows.row(2)); // null < [32, 52, null]
        assert!(rows.row(4) < rows.row(2)); // [32, null, null] < [32, 52, null]
        assert!(rows.row(5) < rows.row(2)); // [null, null, null] > [32, 52, null]
        assert!(rows.row(3) < rows.row(5)); // null < [null, null, null]
        assert_eq!(rows.row(3), rows.row(6)); // null = null (different masked values)

        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(back.len(), 1);
        back[0].to_data().validate_full().unwrap();
        assert_eq!(&back[0], &list);

        let sliced_list = list.slice(1, 5);
        let rows_on_sliced_list = converter
            .convert_columns(&[Arc::clone(&sliced_list)])
            .unwrap();

        assert!(rows_on_sliced_list.row(2) < rows_on_sliced_list.row(1)); // null < [32, 52, null]
        assert!(rows_on_sliced_list.row(3) < rows_on_sliced_list.row(1)); // [32, null, null] < [32, 52, null]
        assert!(rows_on_sliced_list.row(4) < rows_on_sliced_list.row(1)); // [null, null, null] > [32, 52, null]
        assert!(rows_on_sliced_list.row(2) < rows_on_sliced_list.row(4)); // null < [null, null, null]

        let back = converter.convert_rows(&rows_on_sliced_list).unwrap();
        assert_eq!(back.len(), 1);
        back[0].to_data().validate_full().unwrap();
        assert_eq!(&back[0], &sliced_list);
    }

    #[test]
    fn test_two_fixed_size_lists() {
        let mut first = FixedSizeListBuilder::new(UInt8Builder::new(), 1);
        // 0: [100]
        first.values().append_value(100);
        first.append(true);
        // 1: [101]
        first.values().append_value(101);
        first.append(true);
        // 2: [102]
        first.values().append_value(102);
        first.append(true);
        // 3: [null]
        first.values().append_null();
        first.append(true);
        // 4: null
        first.values().append_null(); // MASKED
        first.append(false);
        let first = Arc::new(first.finish()) as ArrayRef;
        let first_type = first.data_type().clone();

        let mut second = FixedSizeListBuilder::new(UInt8Builder::new(), 1);
        // 0: [200]
        second.values().append_value(200);
        second.append(true);
        // 1: [201]
        second.values().append_value(201);
        second.append(true);
        // 2: [202]
        second.values().append_value(202);
        second.append(true);
        // 3: [null]
        second.values().append_null();
        second.append(true);
        // 4: null
        second.values().append_null(); // MASKED
        second.append(false);
        let second = Arc::new(second.finish()) as ArrayRef;
        let second_type = second.data_type().clone();

        let converter = RowConverter::new(vec![
            SortField::new(first_type.clone()),
            SortField::new(second_type.clone()),
        ])
        .unwrap();

        let rows = converter
            .convert_columns(&[Arc::clone(&first), Arc::clone(&second)])
            .unwrap();

        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(back.len(), 2);
        back[0].to_data().validate_full().unwrap();
        assert_eq!(&back[0], &first);
        back[1].to_data().validate_full().unwrap();
        assert_eq!(&back[1], &second);
    }

    #[test]
    fn test_fixed_size_list_with_variable_width_content() {
        let mut first = FixedSizeListBuilder::new(
            StructBuilder::from_fields(
                vec![
                    Field::new(
                        "timestamp",
                        DataType::Timestamp(TimeUnit::Microsecond, Some(Arc::from("UTC"))),
                        false,
                    ),
                    Field::new("offset_minutes", DataType::Int16, false),
                    Field::new("time_zone", DataType::Utf8, false),
                ],
                1,
            ),
            1,
        );
        // 0: null
        first
            .values()
            .field_builder::<TimestampMicrosecondBuilder>(0)
            .unwrap()
            .append_null();
        first
            .values()
            .field_builder::<Int16Builder>(1)
            .unwrap()
            .append_null();
        first
            .values()
            .field_builder::<StringBuilder>(2)
            .unwrap()
            .append_null();
        first.values().append(false);
        first.append(false);
        // 1: [null]
        first
            .values()
            .field_builder::<TimestampMicrosecondBuilder>(0)
            .unwrap()
            .append_null();
        first
            .values()
            .field_builder::<Int16Builder>(1)
            .unwrap()
            .append_null();
        first
            .values()
            .field_builder::<StringBuilder>(2)
            .unwrap()
            .append_null();
        first.values().append(false);
        first.append(true);
        // 2: [1970-01-01 00:00:00.000000 UTC]
        first
            .values()
            .field_builder::<TimestampMicrosecondBuilder>(0)
            .unwrap()
            .append_value(0);
        first
            .values()
            .field_builder::<Int16Builder>(1)
            .unwrap()
            .append_value(0);
        first
            .values()
            .field_builder::<StringBuilder>(2)
            .unwrap()
            .append_value("UTC");
        first.values().append(true);
        first.append(true);
        // 3: [2005-09-10 13:30:00.123456 Europe/Warsaw]
        first
            .values()
            .field_builder::<TimestampMicrosecondBuilder>(0)
            .unwrap()
            .append_value(1126351800123456);
        first
            .values()
            .field_builder::<Int16Builder>(1)
            .unwrap()
            .append_value(120);
        first
            .values()
            .field_builder::<StringBuilder>(2)
            .unwrap()
            .append_value("Europe/Warsaw");
        first.values().append(true);
        first.append(true);
        let first = Arc::new(first.finish()) as ArrayRef;
        let first_type = first.data_type().clone();

        let mut second = StringBuilder::new();
        second.append_value("somewhere near");
        second.append_null();
        second.append_value("Greenwich");
        second.append_value("Warsaw");
        let second = Arc::new(second.finish()) as ArrayRef;
        let second_type = second.data_type().clone();

        let converter = RowConverter::new(vec![
            SortField::new(first_type.clone()),
            SortField::new(second_type.clone()),
        ])
        .unwrap();

        let rows = converter
            .convert_columns(&[Arc::clone(&first), Arc::clone(&second)])
            .unwrap();

        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(back.len(), 2);
        back[0].to_data().validate_full().unwrap();
        assert_eq!(&back[0], &first);
        back[1].to_data().validate_full().unwrap();
        assert_eq!(&back[1], &second);
    }

    #[test]
    fn test_single_map() {
        let mut builder = MapBuilder::new(None, StringBuilder::new(), Int32Builder::new());
        // Entry 0: {"hello": 1, "world": 2}
        builder.keys().append_value("hello");
        builder.values().append_value(1);
        builder.keys().append_value("world");
        builder.values().append_value(2);
        builder.append(true).unwrap();

        // Entry 1: {"foo": 3}
        builder.keys().append_value("foo");
        builder.values().append_value(3);
        builder.append(true).unwrap();

        // Entry 2: {} (empty map)
        builder.append(true).unwrap();

        // Entry 3: null (with masked data)
        builder.keys().append_value("masked_key");
        builder.values().append_value(999);
        builder.append(false).unwrap();

        // Entry 4: null (with unmasked data)
        builder.append(false).unwrap();

        // Entry 5: {"bar": null}
        builder.keys().append_value("bar");
        builder.values().append_null();
        builder.append(true).unwrap();

        // Entry 6: null (with different masked data)
        builder.keys().append_value("other_masked");
        builder.values().append_value(0);
        builder.append(false).unwrap();

        // Entry 7: {"a": 10, "b": 20, "c": 30}
        builder.keys().append_value("a");
        builder.values().append_value(10);
        builder.keys().append_value("b");
        builder.values().append_value(20);
        builder.keys().append_value("c");
        builder.values().append_value(30);
        builder.append(true).unwrap();

        let map = Arc::new(builder.finish()) as ArrayRef;
        let d = map.data_type().clone();

        let converter = RowConverter::new(vec![SortField::new(d.clone())]).unwrap();

        let rows = converter.convert_columns(&[Arc::clone(&map)]).unwrap();

        // null = null (different masked values or masking vs non-masking)
        assert_eq!(rows.row(3), rows.row(4));
        assert_eq!(rows.row(4), rows.row(6));

        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(back.len(), 1);
        back[0].to_data().validate_full().unwrap();
        assert_eq!(&back[0], &map);

        let sliced_map = map.slice(1, map.len() - 2);
        let rows_on_sliced = converter
            .convert_columns(&[Arc::clone(&sliced_map)])
            .unwrap();

        let back = converter.convert_rows(&rows_on_sliced).unwrap();
        assert_eq!(back.len(), 1);
        back[0].to_data().validate_full().unwrap();
        assert_eq!(&back[0], &sliced_map);
    }

    #[test]
    fn two_maps_with_different_keys_order_should_sort_by_entry_order() {
        let map_1: ArrayRef =
            Arc::new(MapArray::from_vec_of_maps::<StringArray, Int32Array, _, _>(
                vec![Some(vec![("hello", Some(1)), ("world", Some(2))])],
                false,
            ));
        // { "world": 2, "hello": 1 }
        let map_2: ArrayRef =
            Arc::new(MapArray::from_vec_of_maps::<StringArray, Int32Array, _, _>(
                vec![Some(vec![("world", Some(2)), ("hello", Some(1))])],
                false,
            ));

        let converter = RowConverter::new(vec![SortField::new(map_1.data_type().clone())]).unwrap();

        let map_1_rows = converter.convert_columns(&[Arc::clone(&map_1)]).unwrap();
        let map_2_rows = converter.convert_columns(&[Arc::clone(&map_2)]).unwrap();

        assert_ne!(map_1_rows.row(0), map_2_rows.row(0));
        assert!(map_1_rows.row(0) < map_2_rows.row(0));

        let back_1 = converter.convert_rows(&map_1_rows).unwrap();
        let back_2 = converter.convert_rows(&map_2_rows).unwrap();
        assert_eq!(&back_1[0], &map_1);
        assert_eq!(&back_2[0], &map_2);
    }

    #[test]
    fn test_nested_map() {
        // Map<Utf8, Map<Utf8, Int32>>
        let mut builder = MapBuilder::new(
            None,
            StringBuilder::new(),
            MapBuilder::new(None, StringBuilder::new(), Int32Builder::new()),
        );

        // Entry 0: {"outer1": {"inner_a": 1, "inner_b": 2}, "outer2": {"inner_c": 3}}
        builder.keys().append_value("outer1");
        builder.values().keys().append_value("inner_a");
        builder.values().values().append_value(1);
        builder.values().keys().append_value("inner_b");
        builder.values().values().append_value(2);
        builder.values().append(true).unwrap();
        builder.keys().append_value("outer2");
        builder.values().keys().append_value("inner_c");
        builder.values().values().append_value(3);
        builder.values().append(true).unwrap();
        builder.append(true).unwrap();

        // Entry 1: {"x": {}} (inner map is empty)
        builder.keys().append_value("x");
        builder.values().append(true).unwrap();
        builder.append(true).unwrap();

        // Entry 2: {"y": null} (inner map is null)
        builder.keys().append_value("y");
        builder.values().keys().append_value("masked"); // MASKED
        builder.values().values().append_value(0); // MASKED
        builder.values().append(false).unwrap();
        builder.append(true).unwrap();

        // Entry 3: {"y": null} (inner map is non masked null)
        builder.keys().append_value("y");
        builder.values().append(false).unwrap(); // not masked
        builder.append(true).unwrap();

        // Entry 4: null (outer map is null)
        builder.keys().append_value("masked_outer"); // MASKED
        builder.values().keys().append_value("masked_inner"); // MASKED
        builder.values().values().append_value(0); // MASKED
        builder.values().append(true).unwrap(); // MASKED
        builder.append(false).unwrap();

        // Entry 5: null (outer map is null with inner map non masked null)
        builder.keys().append_value("masked_outer"); // MASKED
        builder.values().append(false).unwrap(); // not masked
        builder.append(false).unwrap();

        // Entry 6: non masked null (outer map is null)
        builder.append(false).unwrap(); // not masked

        // Entry 7: {} (outer map is empty)
        builder.append(true).unwrap();

        let map = Arc::new(builder.finish()) as ArrayRef;
        let d = map.data_type().clone();

        let converter = RowConverter::new(vec![SortField::new(d.clone())]).unwrap();

        let rows = converter.convert_columns(&[Arc::clone(&map)]).unwrap();

        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(back.len(), 1);
        back[0].to_data().validate_full().unwrap();
        assert_eq!(&back[0], &map);

        let sliced_map = map.slice(1, 3);
        let rows_on_sliced = converter
            .convert_columns(&[Arc::clone(&sliced_map)])
            .unwrap();

        let back = converter.convert_rows(&rows_on_sliced).unwrap();
        assert_eq!(back.len(), 1);
        back[0].to_data().validate_full().unwrap();
        assert_eq!(&back[0], &sliced_map);
    }

    #[test]
    fn test_single_map_with_non_nullable_values() {
        // Use `with_values_field` on `MapBuilder` to set the values are not nullable
        let value_field = Arc::new(Field::new(
            Field::MAP_VALUE_FIELD_DEFAULT_NAME,
            DataType::Int32,
            false,
        ));
        let mut builder = MapBuilder::new(None, StringBuilder::new(), Int32Builder::new())
            .with_values_field(value_field);
        // Entry 0: {"a": 1, "b": 2}
        builder.keys().append_value("a");
        builder.values().append_value(1);
        builder.keys().append_value("b");
        builder.values().append_value(2);
        builder.append(true).unwrap();
        // Entry 1: null
        builder.append(false).unwrap();
        // Entry 2: {"c": 3}
        builder.keys().append_value("c");
        builder.values().append_value(3);
        builder.append(true).unwrap();
        // Entry 3: {}
        builder.append(true).unwrap();
        // Entry 4: null
        builder.keys().append_value("masked"); // MASKED
        builder.values().append_value(0); // MASKED
        builder.append(false).unwrap();

        let map = Arc::new(builder.finish()) as ArrayRef;
        let d = map.data_type().clone();

        let converter = RowConverter::new(vec![SortField::new(d.clone())]).unwrap();

        let rows = converter.convert_columns(&[Arc::clone(&map)]).unwrap();

        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(back.len(), 1);
        back[0].to_data().validate_full().unwrap();
        assert_eq!(&back[0], &map);
    }

    #[test]
    fn test_single_map_with_non_nullable_map_but_with_nullable_values() {
        // Map column is non-nullable, but values are nullable
        let value_field = Arc::new(Field::new(
            Field::MAP_VALUE_FIELD_DEFAULT_NAME,
            DataType::Int32,
            true,
        ));
        let mut builder = MapBuilder::new(None, StringBuilder::new(), Int32Builder::new())
            .with_values_field(value_field);

        // Entry 0: {"a": 1, "b": null}
        builder.keys().append_value("a");
        builder.values().append_value(1);
        builder.keys().append_value("b");
        builder.values().append_null();
        builder.append(true).unwrap();
        // Entry 1: {"c": null, "d": null}
        builder.keys().append_value("c");
        builder.values().append_null();
        builder.keys().append_value("d");
        builder.values().append_null();
        builder.append(true).unwrap();
        // Entry 2: {}
        builder.append(true).unwrap();
        // Entry 3: {"e": 5}
        builder.keys().append_value("e");
        builder.values().append_value(5);
        builder.append(true).unwrap();

        let map = Arc::new(builder.finish()) as ArrayRef;
        let d = map.data_type().clone();

        let converter = RowConverter::new(vec![SortField::new(d.clone())]).unwrap();

        let rows = converter.convert_columns(&[Arc::clone(&map)]).unwrap();

        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(back.len(), 1);
        back[0].to_data().validate_full().unwrap();
        assert_eq!(&back[0], &map);
    }

    #[test]
    fn test_map_all_nulls() {
        let mut builder = MapBuilder::new(None, StringBuilder::new(), Int32Builder::new());
        // All entries are null
        builder.keys().append_value("m1"); // MASKED
        builder.values().append_value(1); // MASKED
        builder.append(false).unwrap();
        builder.keys().append_value("m2"); // MASKED
        builder.values().append_value(2); // MASKED
        builder.append(false).unwrap();

        builder.append(false).unwrap(); // non masked

        builder.keys().append_value("m3"); // MASKED
        builder.values().append_value(3); // MASKED
        builder.append(false).unwrap();

        let map = Arc::new(builder.finish()) as ArrayRef;
        let d = map.data_type().clone();

        let converter = RowConverter::new(vec![SortField::new(d.clone())]).unwrap();

        let rows = converter.convert_columns(&[Arc::clone(&map)]).unwrap();

        // All null rows should be equal
        rows.iter().for_each(|row| assert_eq!(row, rows.row(0)));

        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(back.len(), 1);
        back[0].to_data().validate_full().unwrap();
        assert_eq!(&back[0], &map);
    }

    #[test]
    fn test_map_all_empty() {
        let mut builder = MapBuilder::new(None, StringBuilder::new(), Int32Builder::new());
        // All entries are empty maps
        builder.append(true).unwrap();
        builder.append(true).unwrap();
        builder.append(true).unwrap();

        let map = Arc::new(builder.finish()) as ArrayRef;
        let d = map.data_type().clone();

        let converter = RowConverter::new(vec![SortField::new(d.clone())]).unwrap();

        let rows = converter.convert_columns(&[Arc::clone(&map)]).unwrap();

        // All empty maps should be equal
        rows.iter().for_each(|row| assert_eq!(row, rows.row(0)));

        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(back.len(), 1);
        back[0].to_data().validate_full().unwrap();
        assert_eq!(&back[0], &map);
    }

    #[test]
    fn test_map_empty_array() {
        // Zero-length map array
        let builder = MapBuilder::new(None, StringBuilder::new(), Int32Builder::new());
        let map = Arc::new(builder.finish_cloned()) as ArrayRef;
        let d = map.data_type().clone();

        let converter = RowConverter::new(vec![SortField::new(d.clone())]).unwrap();

        let rows = converter.convert_columns(&[Arc::clone(&map)]).unwrap();

        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(back.len(), 1);
        back[0].to_data().validate_full().unwrap();
        assert_eq!(&back[0], &map);
    }

    fn generate_primitive_array<K>(
        rng: &mut StdRng,
        len: usize,
        valid_percent: f64,
    ) -> PrimitiveArray<K>
    where
        K: ArrowPrimitiveType,
        StandardUniform: Distribution<K::Native>,
    {
        (0..len)
            .map(|_| rng.random_bool(valid_percent).then(|| rng.random()))
            .collect()
    }

    fn generate_all_unique_primitive_array<K>(rng: &mut StdRng, len: usize) -> PrimitiveArray<K>
    where
        K: ArrowPrimitiveType,
        K::Native: Hash + Eq,
        StandardUniform: Distribution<K::Native>,
    {
        let possible_number_of_values = 2i32.saturating_pow(size_of::<K::Native>() as u32 * 8);
        assert!(
            len <= possible_number_of_values as usize,
            "len {len} is larger than the number of possible values {possible_number_of_values}"
        );

        let mut seen = std::collections::HashSet::new();
        (0..len)
            .map(|_| {
                let mut value;
                loop {
                    value = rng.random();

                    if seen.insert(value) {
                        break;
                    }
                }

                Some(value)
            })
            .collect()
    }

    fn generate_boolean_array(rng: &mut StdRng, len: usize, valid_percent: f64) -> BooleanArray {
        (0..len)
            .map(|_| rng.random_bool(valid_percent).then(|| rng.random_bool(0.5)))
            .collect()
    }

    fn generate_strings<O: OffsetSizeTrait>(
        rng: &mut StdRng,
        len: usize,
        valid_percent: f64,
    ) -> GenericStringArray<O> {
        (0..len)
            .map(|_| {
                rng.random_bool(valid_percent).then(|| {
                    let len = rng.random_range(0..100);
                    let bytes = (0..len).map(|_| rng.random_range(0..128)).collect();
                    String::from_utf8(bytes).unwrap()
                })
            })
            .collect()
    }

    fn generate_string_view(rng: &mut StdRng, len: usize, valid_percent: f64) -> StringViewArray {
        (0..len)
            .map(|_| {
                rng.random_bool(valid_percent).then(|| {
                    let len = rng.random_range(0..100);
                    let bytes = (0..len).map(|_| rng.random_range(0..128)).collect();
                    String::from_utf8(bytes).unwrap()
                })
            })
            .collect()
    }

    fn generate_byte_view(rng: &mut StdRng, len: usize, valid_percent: f64) -> BinaryViewArray {
        (0..len)
            .map(|_| {
                rng.random_bool(valid_percent).then(|| {
                    let len = rng.random_range(0..100);
                    let bytes: Vec<_> = (0..len).map(|_| rng.random_range(0..128)).collect();
                    bytes
                })
            })
            .collect()
    }

    fn generate_fixed_stringview_column(len: usize) -> StringViewArray {
        let edge_cases = vec![
            Some("bar".to_string()),
            Some("bar\0".to_string()),
            Some("LongerThan12Bytes".to_string()),
            Some("LongerThan12Bytez".to_string()),
            Some("LongerThan12Bytes\0".to_string()),
            Some("LongerThan12Byt".to_string()),
            Some("backend one".to_string()),
            Some("backend two".to_string()),
            Some("a".repeat(257)),
            Some("a".repeat(300)),
        ];

        // Fill up to `len` by repeating edge cases and trimming
        let mut values = Vec::with_capacity(len);
        for i in 0..len {
            values.push(
                edge_cases
                    .get(i % edge_cases.len())
                    .cloned()
                    .unwrap_or(None),
            );
        }

        StringViewArray::from(values)
    }

    fn generate_dictionary<K>(
        rng: &mut StdRng,
        values: ArrayRef,
        len: usize,
        valid_percent: f64,
    ) -> DictionaryArray<K>
    where
        K: ArrowDictionaryKeyType,
        K::Native: SampleUniform,
    {
        let min_key = K::Native::from_usize(0).unwrap();
        let max_key = K::Native::from_usize(values.len()).unwrap();
        let keys: PrimitiveArray<K> = (0..len)
            .map(|_| {
                rng.random_bool(valid_percent)
                    .then(|| rng.random_range(min_key..max_key))
            })
            .collect();

        let data_type =
            DataType::Dictionary(Box::new(K::DATA_TYPE), Box::new(values.data_type().clone()));

        let data = keys
            .into_data()
            .into_builder()
            .data_type(data_type)
            .add_child_data(values.to_data())
            .build()
            .unwrap();

        DictionaryArray::from(data)
    }

    fn generate_fixed_size_binary(
        rng: &mut StdRng,
        len: usize,
        valid_percent: f64,
    ) -> FixedSizeBinaryArray {
        let width = rng.random_range(0..20);
        let mut builder = FixedSizeBinaryBuilder::new(width);

        let mut b = vec![0; width as usize];
        for _ in 0..len {
            match rng.random_bool(valid_percent) {
                true => {
                    b.iter_mut().for_each(|x| *x = rng.random());
                    builder.append_value(&b).unwrap();
                }
                false => builder.append_null(),
            }
        }

        builder.finish()
    }

    fn generate_struct(rng: &mut StdRng, len: usize, valid_percent: f64) -> StructArray {
        let nulls = NullBuffer::from_iter((0..len).map(|_| rng.random_bool(valid_percent)));
        let a = generate_primitive_array::<Int32Type>(rng, len, valid_percent);
        let b = generate_strings::<i32>(rng, len, valid_percent);
        let fields = Fields::from(vec![
            Field::new("a", DataType::Int32, true),
            Field::new("b", DataType::Utf8, true),
        ]);
        let values = vec![Arc::new(a) as _, Arc::new(b) as _];
        StructArray::new(fields, values, Some(nulls))
    }

    fn generate_list<F>(rng: &mut StdRng, len: usize, valid_percent: f64, values: F) -> ListArray
    where
        F: FnOnce(&mut StdRng, usize) -> ArrayRef,
    {
        let offsets = OffsetBuffer::<i32>::from_lengths((0..len).map(|_| rng.random_range(0..10)));
        let values_len = offsets.last().as_usize();
        let values = values(rng, values_len);
        let nulls = NullBuffer::from_iter((0..len).map(|_| rng.random_bool(valid_percent)));
        let field = Arc::new(Field::new_list_field(values.data_type().clone(), true));
        ListArray::new(field, offsets, values, Some(nulls))
    }

    fn generate_list_view<F>(
        rng: &mut StdRng,
        len: usize,
        valid_percent: f64,
        values: F,
    ) -> ListViewArray
    where
        F: FnOnce(&mut StdRng, usize) -> ArrayRef,
    {
        // Generate sizes first, then create a values array large enough
        let sizes: Vec<i32> = (0..len).map(|_| rng.random_range(0..10)).collect();
        let values_len: usize = sizes.iter().map(|s| *s as usize).sum::<usize>().max(1);
        let values = values(rng, values_len);

        // Generate offsets that can overlap, be non-monotonic, or share ranges
        let offsets: Vec<i32> = sizes
            .iter()
            .map(|&size| {
                if size == 0 {
                    0
                } else {
                    rng.random_range(0..=(values_len as i32 - size))
                }
            })
            .collect();

        let nulls = NullBuffer::from_iter((0..len).map(|_| rng.random_bool(valid_percent)));
        let field = Arc::new(Field::new_list_field(values.data_type().clone(), true));
        ListViewArray::new(
            field,
            ScalarBuffer::from(offsets),
            ScalarBuffer::from(sizes),
            values,
            Some(nulls),
        )
    }

    fn generate_map<KeysFn, ValuesFn>(
        rng: &mut StdRng,
        len: usize,
        valid_percent: f64,
        gen_keys: KeysFn,
        gen_values: ValuesFn,
    ) -> MapArray
    where
        KeysFn: FnOnce(&mut StdRng, usize) -> ArrayRef,
        ValuesFn: FnOnce(&mut StdRng, usize) -> ArrayRef,
    {
        let offsets = OffsetBuffer::<i32>::from_lengths((0..len).map(|_| rng.random_range(0..10)));
        let entries_len = offsets.last().as_usize();
        let keys = gen_keys(rng, entries_len);
        let values = gen_values(rng, entries_len);
        let nulls = NullBuffer::from_iter((0..len).map(|_| rng.random_bool(valid_percent)));
        let field = Arc::new(Field::new_map(
            "",
            Field::MAP_ENTRIES_FIELD_DEFAULT_NAME,
            Field::new(
                Field::MAP_KEY_FIELD_DEFAULT_NAME,
                keys.data_type().clone(),
                false,
            ),
            Field::new(
                Field::MAP_VALUE_FIELD_DEFAULT_NAME,
                values.data_type().clone(),
                true,
            ),
            false,
            true,
        ));
        let DataType::Map(struct_field, _) = field.data_type() else {
            unreachable!();
        };

        let DataType::Struct(fields) = struct_field.data_type() else {
            unreachable!();
        };

        let entries = StructArray::new(fields.clone(), vec![keys, values], None);

        let map_array = MapArray::new(struct_field.clone(), offsets, entries, Some(nulls), false);

        assert_valid_map(&map_array);

        map_array
    }

    /// Assert that the map is valid; this includes unique map keys
    ///
    /// Can be removed once https://github.com/apache/arrow-rs/issues/9475 is resolved and `MapArray`
    /// will not allow creation of invalid map arrays with duplicate keys.
    ///
    /// # Unique map keys
    /// According to [arrow specification][1], the keys of a map must be unique, so this asserts that
    ///
    /// [1]: https://github.com/apache/arrow/blob/cbe2618431e413f12aa16aeba88b3a98914f194b/format/Schema.fbs#L124
    fn assert_valid_map(array: &MapArray) {
        let keys_arrow_row_converter =
            RowConverter::new(vec![SortField::new(array.key_type().clone())]).unwrap();

        array.iter().enumerate().filter_map(|(index, entry)| entry.map(|entry| (index, Arc::clone(entry.column(0))))).for_each(|(entry_index, keys)| {
            let keys_as_rows = keys_arrow_row_converter.convert_columns(&[Arc::clone(&keys)]).expect("should be able to convert keys");

            for i in 0..keys_as_rows.num_rows() {
                for j in (i + 1)..keys_as_rows.num_rows() {
                    if keys_as_rows.row(i) == keys_as_rows.row(j) {
                        let key_i = keys.slice(i, 1);
                        let key_j = keys.slice(j, 1);

                        assert_ne!(keys_as_rows.row(i), keys_as_rows.row(j), "map keys should be unique, but key {i} and key {j} are equal in entry {entry_index}. key {i} value is {key_i:?} and key {j} value is {key_j:?}");
                    }
                }
            }
        })
    }

    fn generate_nulls(rng: &mut StdRng, len: usize) -> Option<NullBuffer> {
        Some(NullBuffer::from_iter(
            (0..len).map(|_| rng.random_bool(0.8)),
        ))
    }

    fn change_underlying_null_values_for_primitive<T: ArrowPrimitiveType>(
        array: &PrimitiveArray<T>,
    ) -> PrimitiveArray<T> {
        let (dt, values, nulls) = array.clone().into_parts();

        let new_values = ScalarBuffer::<T::Native>::from_iter(
            values
                .iter()
                .zip(nulls.as_ref().unwrap().iter())
                .map(|(val, is_valid)| {
                    if is_valid {
                        *val
                    } else {
                        val.add_wrapping(T::Native::usize_as(1))
                    }
                }),
        );

        PrimitiveArray::new(new_values, nulls).with_data_type(dt)
    }

    fn change_underline_null_values_for_byte_array<T: ByteArrayType>(
        array: &GenericByteArray<T>,
    ) -> GenericByteArray<T> {
        let (offsets, values, nulls) = array.clone().into_parts();

        let new_offsets = OffsetBuffer::<T::Offset>::from_lengths(
            offsets
                .lengths()
                .zip(nulls.as_ref().unwrap().iter())
                .map(|(len, is_valid)| if is_valid { len } else { len + 1 }),
        );

        let mut new_bytes = Vec::<u8>::with_capacity(new_offsets[new_offsets.len() - 1].as_usize());

        offsets
            .windows(2)
            .zip(nulls.as_ref().unwrap().iter())
            .for_each(|(start_and_end, is_valid)| {
                let start = start_and_end[0].as_usize();
                let end = start_and_end[1].as_usize();
                new_bytes.extend_from_slice(&values.as_slice()[start..end]);

                // add an extra byte
                if !is_valid {
                    new_bytes.push(b'c');
                }
            });

        GenericByteArray::<T>::new(new_offsets, Buffer::from_vec(new_bytes), nulls)
    }

    fn change_underline_null_values_for_list_array<O: OffsetSizeTrait>(
        array: &GenericListArray<O>,
    ) -> GenericListArray<O> {
        let (field, offsets, values, nulls) = array.clone().into_parts();

        let (new_values, new_offsets) = {
            let concat_values = offsets
                .windows(2)
                .zip(nulls.as_ref().unwrap().iter())
                .map(|(start_and_end, is_valid)| {
                    let start = start_and_end[0].as_usize();
                    let end = start_and_end[1].as_usize();
                    if is_valid {
                        return (start, end - start);
                    }

                    // If reached end, we take one less
                    if end == values.len() {
                        (start, (end - start).saturating_sub(1))
                    } else {
                        (start, end - start + 1)
                    }
                })
                .map(|(start, length)| values.slice(start, length))
                .collect::<Vec<_>>();

            let new_offsets =
                OffsetBuffer::<O>::from_lengths(concat_values.iter().map(|s| s.len()));

            let new_values = {
                let values = concat_values.iter().map(|a| a.as_ref()).collect::<Vec<_>>();
                arrow_select::concat::concat(&values).expect("should be able to concat")
            };

            (new_values, new_offsets)
        };

        GenericListArray::<O>::new(field, new_offsets, new_values, nulls)
    }

    fn change_underline_null_values_for_map_array(array: &MapArray) -> MapArray {
        let (field, offsets, entries, nulls, ordered) = array.clone().into_parts();
        assert!(
            !ordered,
            "can't replace underlying null values for ordered map array as this can violate the ordering"
        );

        let (new_entries, new_offsets) = {
            let concat_values = offsets
                .windows(2)
                .zip(nulls.as_ref().unwrap().iter())
                .map(|(start_and_end, is_valid)| {
                    let start = start_and_end[0].as_usize();
                    let end = start_and_end[1].as_usize();
                    if is_valid {
                        return (start, end - start);
                    }

                    // If reached end, we take one less
                    if end == entries.len() {
                        (start, (end - start).saturating_sub(1))
                    } else {
                        // The keys may no longer be unique
                        (start, end - start + 1)
                    }
                })
                .map(|(start, length)| entries.slice(start, length))
                .collect::<Vec<_>>();

            let new_offsets = OffsetBuffer::from_lengths(concat_values.iter().map(|s| s.len()));

            let new_values = {
                let values = concat_values
                    .iter()
                    .map(|a| a as &dyn Array)
                    .collect::<Vec<_>>();
                arrow_select::concat::concat(&values).expect("should be able to concat")
            };

            (new_values.as_struct().clone(), new_offsets)
        };

        let new_map = MapArray::new(field, new_offsets, new_entries, nulls, ordered);

        assert_valid_map(&new_map);

        new_map
    }

    fn change_underline_null_values(array: &ArrayRef) -> ArrayRef {
        if array.null_count() == 0 {
            return Arc::clone(array);
        }

        downcast_primitive_array!(
            array => {
                let output = change_underlying_null_values_for_primitive(array);

                Arc::new(output)
            }

            DataType::Utf8 => {
                Arc::new(change_underline_null_values_for_byte_array(array.as_string::<i32>()))
            }
            DataType::LargeUtf8 => {
                Arc::new(change_underline_null_values_for_byte_array(array.as_string::<i64>()))
            }
            DataType::Binary => {
                Arc::new(change_underline_null_values_for_byte_array(array.as_binary::<i32>()))
            }
            DataType::LargeBinary => {
                Arc::new(change_underline_null_values_for_byte_array(array.as_binary::<i64>()))
            }
            DataType::List(_) => {
                Arc::new(change_underline_null_values_for_list_array(array.as_list::<i32>()))
            }
            DataType::LargeList(_) => {
                Arc::new(change_underline_null_values_for_list_array(array.as_list::<i64>()))
            }
            DataType::Map(_, _) => {
                Arc::new(change_underline_null_values_for_map_array(array.as_map()))
            }
            _ => {
                Arc::clone(array)
            }
        )
    }

    fn generate_column(rng: &mut StdRng, len: usize) -> ArrayRef {
        match rng.random_range(0..24) {
            0 => Arc::new(generate_primitive_array::<Int32Type>(rng, len, 0.8)),
            1 => Arc::new(generate_primitive_array::<UInt32Type>(rng, len, 0.8)),
            2 => Arc::new(generate_primitive_array::<Int64Type>(rng, len, 0.8)),
            3 => Arc::new(generate_primitive_array::<UInt64Type>(rng, len, 0.8)),
            4 => Arc::new(generate_primitive_array::<Float32Type>(rng, len, 0.8)),
            5 => Arc::new(generate_primitive_array::<Float64Type>(rng, len, 0.8)),
            6 => Arc::new(generate_strings::<i32>(rng, len, 0.8)),
            7 => {
                let dict_values_len = rng.random_range(1..len);
                // Cannot test dictionaries containing null values because of #2687
                let strings = Arc::new(generate_strings::<i32>(rng, dict_values_len, 1.0));
                Arc::new(generate_dictionary::<Int64Type>(rng, strings, len, 0.8))
            }
            8 => {
                let dict_values_len = rng.random_range(1..len);
                // Cannot test dictionaries containing null values because of #2687
                let values = Arc::new(generate_primitive_array::<Int64Type>(
                    rng,
                    dict_values_len,
                    1.0,
                ));
                Arc::new(generate_dictionary::<Int64Type>(rng, values, len, 0.8))
            }
            9 => Arc::new(generate_fixed_size_binary(rng, len, 0.8)),
            10 => Arc::new(generate_struct(rng, len, 0.8)),
            11 => Arc::new(generate_list(rng, len, 0.8, |rng, values_len| {
                Arc::new(generate_primitive_array::<Int64Type>(rng, values_len, 0.8))
            })),
            12 => Arc::new(generate_list(rng, len, 0.8, |rng, values_len| {
                Arc::new(generate_strings::<i32>(rng, values_len, 0.8))
            })),
            13 => Arc::new(generate_list(rng, len, 0.8, |rng, values_len| {
                Arc::new(generate_struct(rng, values_len, 0.8))
            })),
            14 => Arc::new(generate_string_view(rng, len, 0.8)),
            15 => Arc::new(generate_byte_view(rng, len, 0.8)),
            16 => Arc::new(generate_fixed_stringview_column(len)),
            17 => Arc::new(
                generate_list(rng, len + 1000, 0.8, |rng, values_len| {
                    Arc::new(generate_primitive_array::<Int64Type>(rng, values_len, 0.8))
                })
                .slice(500, len),
            ),
            18 => Arc::new(generate_boolean_array(rng, len, 0.8)),
            19 => Arc::new(generate_list_view(rng, len, 0.8, |rng, values_len| {
                Arc::new(generate_primitive_array::<Int64Type>(rng, values_len, 0.8))
            })),
            20 => Arc::new(generate_list_view(rng, len, 0.8, |rng, values_len| {
                Arc::new(generate_strings::<i32>(rng, values_len, 0.8))
            })),
            21 => Arc::new(generate_list_view(rng, len, 0.8, |rng, values_len| {
                Arc::new(generate_struct(rng, values_len, 0.8))
            })),
            22 => Arc::new(
                generate_list_view(rng, len + 1000, 0.8, |rng, values_len| {
                    Arc::new(generate_primitive_array::<Int64Type>(rng, values_len, 0.8))
                })
                .slice(500, len),
            ),
            23 => Arc::new(generate_map(
                rng,
                len,
                0.9,
                // Need to generate all unique keys or make sure between each map every key is unique,
                |rng, keys_len| {
                    Arc::new(generate_all_unique_primitive_array::<Int64Type>(
                        rng, keys_len,
                    ))
                },
                |rng, values_len| Arc::new(generate_strings::<i32>(rng, values_len, 0.7)),
            )),
            _ => unreachable!(),
        }
    }

    fn print_row(cols: &[SortColumn], row: usize) -> String {
        let t: Vec<_> = cols
            .iter()
            .map(|x| match x.values.is_valid(row) {
                true => {
                    let opts = FormatOptions::default().with_null("NULL");
                    let formatter = ArrayFormatter::try_new(x.values.as_ref(), &opts).unwrap();
                    formatter.value(row).to_string()
                }
                false => "NULL".to_string(),
            })
            .collect();
        t.join(",")
    }

    fn print_col_types(cols: &[SortColumn]) -> String {
        let t: Vec<_> = cols
            .iter()
            .map(|x| x.values.data_type().to_string())
            .collect();
        t.join(",")
    }

    #[derive(Debug, PartialEq)]
    enum Nulls {
        /// Keep the generated array as is
        AsIs,

        /// Replace the null buffer with different null buffer to point to different positions as null
        Different,

        /// Remove all nulls
        None,
    }

    #[test]
    #[cfg_attr(miri, ignore)] // Takes too long
    fn fuzz_test() {
        let mut rng = StdRng::seed_from_u64(42);
        for _ in 0..100 {
            for null_behavior in [Nulls::AsIs, Nulls::Different, Nulls::None] {
                let num_columns = rng.random_range(1..5);
                let len = rng.random_range(5..100);
                let mut arrays: Vec<_> = (0..num_columns)
                    .map(|_| generate_column(&mut rng, len))
                    .collect();

                match null_behavior {
                    Nulls::AsIs => {
                        // Keep as is
                    }
                    Nulls::Different => {
                        // Replace nulls with different nulls to allow for testing different underlying null values
                        arrays = arrays
                            .into_iter()
                            .map(|a| replace_array_nulls(a, generate_nulls(&mut rng, len)))
                            .collect()
                    }
                    Nulls::None => {
                        // Remove nulls
                        arrays = arrays
                            .into_iter()
                            .map(|a| replace_array_nulls(a, None))
                            .collect()
                    }
                }

                let options: Vec<_> = (0..num_columns)
                    .map(|_| SortOptions {
                        descending: rng.random_bool(0.5),
                        nulls_first: rng.random_bool(0.5),
                    })
                    .collect();

                let sort_columns: Vec<_> = options
                    .iter()
                    .zip(&arrays)
                    .map(|(o, c)| SortColumn {
                        values: Arc::clone(c),
                        options: Some(*o),
                    })
                    .collect();

                let comparator = LexicographicalComparator::try_new(&sort_columns).unwrap();

                let columns: Vec<SortField> = options
                    .into_iter()
                    .zip(&arrays)
                    .map(|(o, a)| SortField::new_with_options(a.data_type().clone(), o))
                    .collect();

                let converter = RowConverter::new(columns).unwrap();
                let rows = converter.convert_columns(&arrays).unwrap();

                // Assert that the underlying null values are not taken into account when converting
                // even for different inputs
                if !matches!(null_behavior, Nulls::None) {
                    assert_same_rows_when_changing_input_underlying_null_values(
                        &arrays, &converter, &rows,
                    );
                }

                for i in 0..len {
                    for j in 0..len {
                        let row_i = rows.row(i);
                        let row_j = rows.row(j);
                        let row_cmp = row_i.cmp(&row_j);
                        let lex_cmp = comparator.compare(i, j);
                        assert_eq!(
                            row_cmp,
                            lex_cmp,
                            "({:?} vs {:?}) vs ({:?} vs {:?}) for types {}",
                            print_row(&sort_columns, i),
                            print_row(&sort_columns, j),
                            row_i,
                            row_j,
                            print_col_types(&sort_columns)
                        );
                    }
                }

                // Validate rows length iterator
                {
                    let mut rows_iter = rows.iter();
                    let mut rows_lengths_iter = rows.lengths();
                    for (index, row) in rows_iter.by_ref().enumerate() {
                        let len = rows_lengths_iter
                            .next()
                            .expect("Reached end of length iterator while still have rows");
                        assert_eq!(
                            row.data.len(),
                            len,
                            "Row length mismatch: {} vs {}",
                            row.data.len(),
                            len
                        );
                        assert_eq!(
                            len,
                            rows.row_len(index),
                            "Row length mismatch at index {}: {} vs {}",
                            index,
                            len,
                            rows.row_len(index)
                        );
                    }

                    assert_eq!(
                        rows_lengths_iter.next(),
                        None,
                        "Length iterator did not reach end"
                    );
                }

                // Convert rows produced from convert_columns().
                // Note: validate_utf8 is set to false since Row is initialized through empty_rows()
                let back = converter.convert_rows(&rows).unwrap();
                for (actual, expected) in back.iter().zip(&arrays) {
                    actual.to_data().validate_full().unwrap();
                    dictionary_eq(actual, expected)
                }

                // Check that we can convert rows into ByteArray and then parse, convert it back to array
                // Note: validate_utf8 is set to true since Row is initialized through RowParser
                let rows = rows.try_into_binary().expect("reasonable size");
                let parser = converter.parser();
                let back = converter
                    .convert_rows(rows.iter().map(|b| parser.parse(b.expect("valid bytes"))))
                    .unwrap();
                for (actual, expected) in back.iter().zip(&arrays) {
                    actual.to_data().validate_full().unwrap();
                    dictionary_eq(actual, expected)
                }

                let rows = converter.from_binary(rows);
                let back = converter.convert_rows(&rows).unwrap();
                for (actual, expected) in back.iter().zip(&arrays) {
                    actual.to_data().validate_full().unwrap();
                    dictionary_eq(actual, expected)
                }
            }
        }
    }

    fn replace_array_nulls(array: ArrayRef, new_nulls: Option<NullBuffer>) -> ArrayRef {
        make_array(
            array
                .into_data()
                .into_builder()
                // Replace the nulls
                .nulls(new_nulls)
                .build()
                .unwrap(),
        )
    }

    fn assert_same_rows_when_changing_input_underlying_null_values(
        arrays: &[ArrayRef],
        converter: &RowConverter,
        rows: &Rows,
    ) {
        let arrays_with_different_data_behind_nulls = arrays
            .iter()
            .map(|arr| change_underline_null_values(arr))
            .collect::<Vec<_>>();

        // Skip assertion if we did not change anything
        if arrays
            .iter()
            .zip(arrays_with_different_data_behind_nulls.iter())
            .all(|(a, b)| Arc::ptr_eq(a, b))
        {
            return;
        }

        let rows_with_different_nulls = converter
            .convert_columns(&arrays_with_different_data_behind_nulls)
            .unwrap();

        assert_eq!(
            rows.iter().collect::<Vec<_>>(),
            rows_with_different_nulls.iter().collect::<Vec<_>>(),
            "Different underlying nulls should not output different rows"
        )
    }

    #[test]
    fn test_clear() {
        let converter = RowConverter::new(vec![SortField::new(DataType::Int32)]).unwrap();
        let mut rows = converter.empty_rows(3, 128);

        let first = Int32Array::from(vec![None, Some(2), Some(4)]);
        let second = Int32Array::from(vec![Some(2), None, Some(4)]);
        let arrays = [Arc::new(first) as ArrayRef, Arc::new(second) as ArrayRef];

        for array in &arrays {
            rows.clear();
            converter
                .append(&mut rows, std::slice::from_ref(array))
                .unwrap();
            let back = converter.convert_rows(&rows).unwrap();
            assert_eq!(&back[0], array);
        }

        let mut rows_expected = converter.empty_rows(3, 128);
        converter.append(&mut rows_expected, &arrays[1..]).unwrap();

        for (i, (actual, expected)) in rows.iter().zip(rows_expected.iter()).enumerate() {
            assert_eq!(
                actual, expected,
                "For row {i}: expected {expected:?}, actual: {actual:?}",
            );
        }
    }

    #[test]
    fn test_append_codec_dictionary_binary() {
        use DataType::*;
        // Dictionary RowConverter
        let converter = RowConverter::new(vec![SortField::new(Dictionary(
            Box::new(Int32),
            Box::new(Binary),
        ))])
        .unwrap();
        let mut rows = converter.empty_rows(4, 128);

        let keys = Int32Array::from_iter_values([0, 1, 2, 3]);
        let values = BinaryArray::from(vec![
            Some(b"a".as_slice()),
            Some(b"b"),
            Some(b"c"),
            Some(b"d"),
        ]);
        let dict_array = DictionaryArray::new(keys, Arc::new(values));

        rows.clear();
        let array = Arc::new(dict_array) as ArrayRef;
        converter
            .append(&mut rows, std::slice::from_ref(&array))
            .unwrap();
        let back = converter.convert_rows(&rows).unwrap();

        dictionary_eq(&back[0], &array);
    }

    #[test]
    fn test_list_prefix() {
        let mut a = ListBuilder::new(Int8Builder::new());
        a.append_value([None]);
        a.append_value([None, None]);
        let a = a.finish();

        let converter = RowConverter::new(vec![SortField::new(a.data_type().clone())]).unwrap();
        let rows = converter.convert_columns(&[Arc::new(a) as _]).unwrap();
        assert_eq!(rows.row(0).cmp(&rows.row(1)), Ordering::Less);
    }

    #[test]
    fn test_utf8_validation_doesnt_affect_values_buffer_size() {
        fn assert_values_buffer_lens(col: ArrayRef) -> usize {
            // 1. Convert cols into rows
            let converter = RowConverter::new(vec![SortField::new(DataType::Utf8View)]).unwrap();

            // 2a. Convert rows into cols (validate_utf8 = false)
            let rows = converter.convert_columns(&[col]).unwrap();
            let converted = converter.convert_rows(&rows).unwrap();
            let unchecked_values_len = converted[0].as_string_view().data_buffers()[0].len();

            // 2b. Convert rows into cols (validate_utf8 = true since Row is initialized through RowParser)
            let rows = rows.try_into_binary().expect("reasonable size");
            let parser = converter.parser();
            let converted = converter
                .convert_rows(rows.iter().map(|b| parser.parse(b.expect("valid bytes"))))
                .unwrap();
            let checked_values_len = converted[0].as_string_view().data_buffers()[0].len();
            // Regardless of utf8 validation flag, we should always have minimal data in buffers
            assert_eq!(unchecked_values_len, checked_values_len);
            checked_values_len
        }

        // Case1. StringViewArray with inline strings
        let col = Arc::new(StringViewArray::from_iter([
            Some("hello"), // short(5)
            None,          // null
            Some("short"), // short(5)
            Some("tiny"),  // short(4)
        ])) as ArrayRef;

        let values_len = assert_values_buffer_lens(col);
        // Since there are no long (>12) strings, len of values buffer is 0
        assert_eq!(values_len, 0);

        // Case2. StringViewArray with long(>12) strings
        let col = Arc::new(StringViewArray::from_iter([
            Some("1234567890123"),  // 13
            Some("12345678901234"), // 14
        ])) as ArrayRef;

        let values_len = assert_values_buffer_lens(col);
        assert_eq!(values_len, 13 + 14);

        // Case3. StringViewArray with both short and long strings
        let col = Arc::new(StringViewArray::from_iter([
            Some("tiny"),          // 4 (short)
            Some("thisisexact13"), // 13 (long)
            None,
            Some("short"), // 5 (short)
        ])) as ArrayRef;

        let values_len = assert_values_buffer_lens(col);
        // Since there is single long string, len of values buffer is 13
        assert_eq!(values_len, 13);
    }

    #[test]
    fn test_sparse_union() {
        // create a sparse union with Int32 (type_id = 0) and Utf8 (type_id = 1)
        let int_array = Int32Array::from(vec![Some(1), None, Some(3), None, Some(5)]);
        let str_array = StringArray::from(vec![None, Some("b"), None, Some("d"), None]);

        // [1, "b", 3, "d", 5]
        let type_ids = vec![0, 1, 0, 1, 0].into();

        let union_fields = [
            (0, Arc::new(Field::new("int", DataType::Int32, false))),
            (1, Arc::new(Field::new("str", DataType::Utf8, false))),
        ]
        .into_iter()
        .collect();

        let union_array = UnionArray::try_new(
            union_fields,
            type_ids,
            None,
            vec![Arc::new(int_array) as ArrayRef, Arc::new(str_array)],
        )
        .unwrap();

        let union_type = union_array.data_type().clone();
        let converter = RowConverter::new(vec![SortField::new(union_type)]).unwrap();

        let rows = converter
            .convert_columns(&[Arc::new(union_array.clone())])
            .unwrap();

        // round trip
        let back = converter.convert_rows(&rows).unwrap();
        let back_union = back[0].as_any().downcast_ref::<UnionArray>().unwrap();

        assert_eq!(union_array.len(), back_union.len());
        for i in 0..union_array.len() {
            assert_eq!(union_array.type_id(i), back_union.type_id(i));
        }
    }

    #[test]
    fn test_sparse_union_with_nulls() {
        // create a sparse union with Int32 (type_id = 0) and Utf8 (type_id = 1)
        let int_array = Int32Array::from(vec![Some(1), None, Some(3), None, Some(5)]);
        let str_array = StringArray::from(vec![None::<&str>; 5]);

        // [1, null (both children null), 3, null (both children null), 5]
        let type_ids = vec![0, 1, 0, 1, 0].into();

        let union_fields = [
            (0, Arc::new(Field::new("int", DataType::Int32, true))),
            (1, Arc::new(Field::new("str", DataType::Utf8, true))),
        ]
        .into_iter()
        .collect();

        let union_array = UnionArray::try_new(
            union_fields,
            type_ids,
            None,
            vec![Arc::new(int_array) as ArrayRef, Arc::new(str_array)],
        )
        .unwrap();

        let union_type = union_array.data_type().clone();
        let converter = RowConverter::new(vec![SortField::new(union_type)]).unwrap();

        let rows = converter
            .convert_columns(&[Arc::new(union_array.clone())])
            .unwrap();

        // round trip
        let back = converter.convert_rows(&rows).unwrap();
        let back_union = back[0].as_any().downcast_ref::<UnionArray>().unwrap();

        assert_eq!(union_array.len(), back_union.len());
        for i in 0..union_array.len() {
            let expected_null = union_array.is_null(i);
            let actual_null = back_union.is_null(i);
            assert_eq!(expected_null, actual_null, "Null mismatch at index {i}");
            if !expected_null {
                assert_eq!(union_array.type_id(i), back_union.type_id(i));
            }
        }
    }

    #[test]
    fn test_dense_union() {
        // create a dense union with Int32 (type_id = 0) and use Utf8 (type_id = 1)
        let int_array = Int32Array::from(vec![1, 3, 5]);
        let str_array = StringArray::from(vec!["a", "b"]);

        let type_ids = vec![0, 1, 0, 1, 0].into();

        // [1, "a", 3, "b", 5]
        let offsets = vec![0, 0, 1, 1, 2].into();

        let union_fields = [
            (0, Arc::new(Field::new("int", DataType::Int32, false))),
            (1, Arc::new(Field::new("str", DataType::Utf8, false))),
        ]
        .into_iter()
        .collect();

        let union_array = UnionArray::try_new(
            union_fields,
            type_ids,
            Some(offsets), // Dense mode
            vec![Arc::new(int_array) as ArrayRef, Arc::new(str_array)],
        )
        .unwrap();

        let union_type = union_array.data_type().clone();
        let converter = RowConverter::new(vec![SortField::new(union_type)]).unwrap();

        let rows = converter
            .convert_columns(&[Arc::new(union_array.clone())])
            .unwrap();

        // round trip
        let back = converter.convert_rows(&rows).unwrap();
        let back_union = back[0].as_any().downcast_ref::<UnionArray>().unwrap();

        assert_eq!(union_array.len(), back_union.len());
        for i in 0..union_array.len() {
            assert_eq!(union_array.type_id(i), back_union.type_id(i));
        }
    }

    #[test]
    fn test_dense_union_with_nulls() {
        // create a dense union with Int32 (type_id = 0) and Utf8 (type_id = 1)
        let int_array = Int32Array::from(vec![Some(1), None, Some(5)]);
        let str_array = StringArray::from(vec![Some("a"), None]);

        // [1, "a", 5, null (str null), null (int null)]
        let type_ids = vec![0, 1, 0, 1, 0].into();
        let offsets = vec![0, 0, 1, 1, 2].into();

        let union_fields = [
            (0, Arc::new(Field::new("int", DataType::Int32, true))),
            (1, Arc::new(Field::new("str", DataType::Utf8, true))),
        ]
        .into_iter()
        .collect();

        let union_array = UnionArray::try_new(
            union_fields,
            type_ids,
            Some(offsets),
            vec![Arc::new(int_array) as ArrayRef, Arc::new(str_array)],
        )
        .unwrap();

        let union_type = union_array.data_type().clone();
        let converter = RowConverter::new(vec![SortField::new(union_type)]).unwrap();

        let rows = converter
            .convert_columns(&[Arc::new(union_array.clone())])
            .unwrap();

        // round trip
        let back = converter.convert_rows(&rows).unwrap();
        let back_union = back[0].as_any().downcast_ref::<UnionArray>().unwrap();

        assert_eq!(union_array.len(), back_union.len());
        for i in 0..union_array.len() {
            let expected_null = union_array.is_null(i);
            let actual_null = back_union.is_null(i);
            assert_eq!(expected_null, actual_null, "Null mismatch at index {i}");
            if !expected_null {
                assert_eq!(union_array.type_id(i), back_union.type_id(i));
            }
        }
    }

    #[test]
    fn test_union_ordering() {
        let int_array = Int32Array::from(vec![100, 5, 20]);
        let str_array = StringArray::from(vec!["z", "a"]);

        // [100, "z", 5, "a", 20]
        let type_ids = vec![0, 1, 0, 1, 0].into();
        let offsets = vec![0, 0, 1, 1, 2].into();

        let union_fields = [
            (0, Arc::new(Field::new("int", DataType::Int32, false))),
            (1, Arc::new(Field::new("str", DataType::Utf8, false))),
        ]
        .into_iter()
        .collect();

        let union_array = UnionArray::try_new(
            union_fields,
            type_ids,
            Some(offsets),
            vec![Arc::new(int_array) as ArrayRef, Arc::new(str_array)],
        )
        .unwrap();

        let union_type = union_array.data_type().clone();
        let converter = RowConverter::new(vec![SortField::new(union_type)]).unwrap();

        let rows = converter.convert_columns(&[Arc::new(union_array)]).unwrap();

        /*
        expected ordering

        row 2: 5    - type_id 0
        row 4: 20   - type_id 0
        row 0: 100  - type id 0
        row 3: "a"  - type id 1
        row 1: "z"  - type id 1
        */

        // 5 < "z"
        assert!(rows.row(2) < rows.row(1));

        // 100 < "a"
        assert!(rows.row(0) < rows.row(3));

        // among ints
        // 5 < 20
        assert!(rows.row(2) < rows.row(4));
        // 20 < 100
        assert!(rows.row(4) < rows.row(0));

        // among strigns
        // "a" < "z"
        assert!(rows.row(3) < rows.row(1));
    }

    #[test]
    fn test_row_converter_roundtrip_with_many_union_columns() {
        // col 1: Union(Int32, Utf8) [67, "hello"]
        let fields1 = UnionFields::try_new(
            vec![0, 1],
            vec![
                Field::new("int", DataType::Int32, true),
                Field::new("string", DataType::Utf8, true),
            ],
        )
        .unwrap();

        let int_array1 = Int32Array::from(vec![Some(67), None]);
        let string_array1 = StringArray::from(vec![None::<&str>, Some("hello")]);
        let type_ids1 = vec![0i8, 1].into();

        let union_array1 = UnionArray::try_new(
            fields1.clone(),
            type_ids1,
            None,
            vec![
                Arc::new(int_array1) as ArrayRef,
                Arc::new(string_array1) as ArrayRef,
            ],
        )
        .unwrap();

        // col 2: Union(Int32, Utf8) [100, "world"]
        let fields2 = UnionFields::try_new(
            vec![0, 1],
            vec![
                Field::new("int", DataType::Int32, true),
                Field::new("string", DataType::Utf8, true),
            ],
        )
        .unwrap();

        let int_array2 = Int32Array::from(vec![Some(100), None]);
        let string_array2 = StringArray::from(vec![None::<&str>, Some("world")]);
        let type_ids2 = vec![0i8, 1].into();

        let union_array2 = UnionArray::try_new(
            fields2.clone(),
            type_ids2,
            None,
            vec![
                Arc::new(int_array2) as ArrayRef,
                Arc::new(string_array2) as ArrayRef,
            ],
        )
        .unwrap();

        // create a row converter with 2 union columns
        let field1 = Field::new("col1", DataType::Union(fields1, UnionMode::Sparse), true);
        let field2 = Field::new("col2", DataType::Union(fields2, UnionMode::Sparse), true);

        let sort_field1 = SortField::new(field1.data_type().clone());
        let sort_field2 = SortField::new(field2.data_type().clone());

        let converter = RowConverter::new(vec![sort_field1, sort_field2]).unwrap();

        let rows = converter
            .convert_columns(&[
                Arc::new(union_array1.clone()) as ArrayRef,
                Arc::new(union_array2.clone()) as ArrayRef,
            ])
            .unwrap();

        // roundtrip
        let out = converter.convert_rows(&rows).unwrap();

        let [col1, col2] = out.as_slice() else {
            panic!("expected 2 columns")
        };

        let col1 = col1.as_any().downcast_ref::<UnionArray>().unwrap();
        let col2 = col2.as_any().downcast_ref::<UnionArray>().unwrap();

        for (expected, got) in [union_array1, union_array2].iter().zip([col1, col2]) {
            assert_eq!(expected.len(), got.len());
            assert_eq!(expected.type_ids(), got.type_ids());

            for i in 0..expected.len() {
                assert_eq!(expected.value(i).as_ref(), got.value(i).as_ref());
            }
        }
    }

    #[test]
    fn test_row_converter_roundtrip_with_one_union_column() {
        let fields = UnionFields::try_new(
            vec![0, 1],
            vec![
                Field::new("int", DataType::Int32, true),
                Field::new("string", DataType::Utf8, true),
            ],
        )
        .unwrap();

        let int_array = Int32Array::from(vec![Some(67), None]);
        let string_array = StringArray::from(vec![None::<&str>, Some("hello")]);
        let type_ids = vec![0i8, 1].into();

        let union_array = UnionArray::try_new(
            fields.clone(),
            type_ids,
            None,
            vec![
                Arc::new(int_array) as ArrayRef,
                Arc::new(string_array) as ArrayRef,
            ],
        )
        .unwrap();

        let field = Field::new("col", DataType::Union(fields, UnionMode::Sparse), true);
        let sort_field = SortField::new(field.data_type().clone());
        let converter = RowConverter::new(vec![sort_field]).unwrap();

        let rows = converter
            .convert_columns(&[Arc::new(union_array.clone()) as ArrayRef])
            .unwrap();

        // roundtrip
        let out = converter.convert_rows(&rows).unwrap();

        let [col1] = out.as_slice() else {
            panic!("expected 1 column")
        };

        let col = col1.as_any().downcast_ref::<UnionArray>().unwrap();
        assert_eq!(col.len(), union_array.len());
        assert_eq!(col.type_ids(), union_array.type_ids());

        for i in 0..col.len() {
            assert_eq!(col.value(i).as_ref(), union_array.value(i).as_ref());
        }
    }

    #[test]
    fn test_row_converter_roundtrip_with_non_default_union_type_ids() {
        // test with non-sequential type IDs (70, 85) instead of (0, 1)
        let fields = UnionFields::try_new(
            vec![70, 85],
            vec![
                Field::new("int", DataType::Int32, true),
                Field::new("string", DataType::Utf8, true),
            ],
        )
        .unwrap();

        let int_array = Int32Array::from(vec![Some(67), None]);
        let string_array = StringArray::from(vec![None::<&str>, Some("hello")]);
        let type_ids = vec![70i8, 85].into();

        let union_array = UnionArray::try_new(
            fields.clone(),
            type_ids,
            None,
            vec![
                Arc::new(int_array) as ArrayRef,
                Arc::new(string_array) as ArrayRef,
            ],
        )
        .unwrap();

        let field = Field::new("col", DataType::Union(fields, UnionMode::Sparse), true);
        let sort_field = SortField::new(field.data_type().clone());
        let converter = RowConverter::new(vec![sort_field]).unwrap();

        let rows = converter
            .convert_columns(&[Arc::new(union_array.clone()) as ArrayRef])
            .unwrap();

        // roundtrip
        let out = converter.convert_rows(&rows).unwrap();

        let [col1] = out.as_slice() else {
            panic!("expected 1 column")
        };

        let col = col1.as_any().downcast_ref::<UnionArray>().unwrap();
        assert_eq!(col.len(), union_array.len());
        assert_eq!(col.type_ids(), union_array.type_ids());

        for i in 0..col.len() {
            assert_eq!(col.value(i).as_ref(), union_array.value(i).as_ref());
        }
    }

    #[test]
    fn rows_size_should_count_for_capacity() {
        let row_converter = RowConverter::new(vec![SortField::new(DataType::UInt8)]).unwrap();

        let empty_rows_size_with_preallocate_rows_and_data = {
            let rows = row_converter.empty_rows(1000, 1000);

            rows.size()
        };
        let empty_rows_size_with_preallocate_rows = {
            let rows = row_converter.empty_rows(1000, 0);

            rows.size()
        };
        let empty_rows_size_with_preallocate_data = {
            let rows = row_converter.empty_rows(0, 1000);

            rows.size()
        };
        let empty_rows_size_without_preallocate = {
            let rows = row_converter.empty_rows(0, 0);

            rows.size()
        };

        assert!(
            empty_rows_size_with_preallocate_rows_and_data > empty_rows_size_with_preallocate_rows,
            "{empty_rows_size_with_preallocate_rows_and_data} should be larger than {empty_rows_size_with_preallocate_rows}"
        );
        assert!(
            empty_rows_size_with_preallocate_rows_and_data > empty_rows_size_with_preallocate_data,
            "{empty_rows_size_with_preallocate_rows_and_data} should be larger than {empty_rows_size_with_preallocate_data}"
        );
        assert!(
            empty_rows_size_with_preallocate_rows > empty_rows_size_without_preallocate,
            "{empty_rows_size_with_preallocate_rows} should be larger than {empty_rows_size_without_preallocate}"
        );
        assert!(
            empty_rows_size_with_preallocate_data > empty_rows_size_without_preallocate,
            "{empty_rows_size_with_preallocate_data} should be larger than {empty_rows_size_without_preallocate}"
        );
    }

    #[test]
    fn test_struct_no_child_fields() {
        fn run_test(array: ArrayRef) {
            let sort_fields = vec![SortField::new(array.data_type().clone())];
            let converter = RowConverter::new(sort_fields).unwrap();
            let r = converter.convert_columns(&[Arc::clone(&array)]).unwrap();

            let back = converter.convert_rows(&r).unwrap();
            assert_eq!(back.len(), 1);
            assert_eq!(&back[0], &array);
        }

        let s = Arc::new(StructArray::new_empty_fields(5, None)) as ArrayRef;
        run_test(s);

        let s = Arc::new(StructArray::new_empty_fields(
            5,
            Some(vec![true, false, true, false, false].into()),
        )) as ArrayRef;
        run_test(s);
    }

    #[test]
    fn reserve_should_increase_capacity_to_the_requested_size() {
        let row_converter = RowConverter::new(vec![SortField::new(DataType::UInt8)]).unwrap();
        let mut empty_rows = row_converter.empty_rows(0, 0);
        empty_rows.reserve(50, 50);
        let before_size = empty_rows.size();
        empty_rows.reserve(50, 50);
        assert_eq!(
            empty_rows.size(),
            before_size,
            "Size should not change when reserving already reserved space"
        );
        empty_rows.reserve(10, 20);
        assert_eq!(
            empty_rows.size(),
            before_size,
            "Size should not change when already have space for the expected reserved data"
        );

        empty_rows.reserve(100, 20);
        assert!(
            empty_rows.size() > before_size,
            "Size should increase when reserving more space than previously reserved"
        );

        let before_size = empty_rows.size();

        empty_rows.reserve(20, 100);
        assert!(
            empty_rows.size() > before_size,
            "Size should increase when reserving more space than previously reserved"
        );
    }

    #[test]
    fn empty_rows_should_return_empty_lengths_iterator() {
        let rows = RowConverter::new(vec![SortField::new(DataType::UInt8)])
            .unwrap()
            .empty_rows(0, 0);
        let mut lengths_iter = rows.lengths();
        assert_eq!(lengths_iter.next(), None);
    }

    #[test]
    #[should_panic(expected = "row index out of bounds")]
    fn row_should_panic_on_overflowing_index() {
        let rows = RowConverter::new(vec![SortField::new(DataType::Int32)])
            .unwrap()
            .empty_rows(0, 0);
        rows.row(usize::MAX);
    }

    #[test]
    #[should_panic(expected = "row index out of bounds")]
    fn row_len_should_panic_on_overflowing_index() {
        let rows = RowConverter::new(vec![SortField::new(DataType::Int32)])
            .unwrap()
            .empty_rows(0, 0);
        rows.row_len(usize::MAX);
    }

    #[test]
    fn test_nested_null_list() {
        let null_array = Arc::new(NullArray::new(3));
        // [[NULL], [], [NULL, NULL]]
        let list: ArrayRef = Arc::new(ListArray::new(
            Field::new_list_field(DataType::Null, true).into(),
            OffsetBuffer::from_lengths(vec![1, 0, 2]),
            null_array,
            None,
        ));

        let converter = RowConverter::new(vec![SortField::new(list.data_type().clone())]).unwrap();
        let rows = converter.convert_columns(&[Arc::clone(&list)]).unwrap();
        let back = converter.convert_rows(&rows).unwrap();

        assert_eq!(&list, &back[0]);
    }

    // Test case from original issue #9227: [[NULL]] - double nested List with Null
    #[test]
    fn test_double_nested_null_list() {
        let null_array = Arc::new(NullArray::new(1));
        // [NULL]
        let nested_field = Arc::new(Field::new_list_field(DataType::Null, true));
        let nested_list = Arc::new(ListArray::new(
            nested_field.clone(),
            OffsetBuffer::from_lengths(vec![1]),
            null_array,
            None,
        ));
        // [[NULL]]
        let list = Arc::new(ListArray::new(
            Field::new_list_field(DataType::List(nested_field), true).into(),
            OffsetBuffer::from_lengths(vec![1]),
            nested_list,
            None,
        )) as ArrayRef;

        let converter = RowConverter::new(vec![SortField::new(list.data_type().clone())]).unwrap();
        let rows = converter.convert_columns(&[Arc::clone(&list)]).unwrap();
        let back = converter.convert_rows(&rows).unwrap();

        assert_eq!(&list, &back[0]);
    }

    // Test LargeList<Null>
    #[test]
    fn test_large_list_null() {
        let null_array = Arc::new(NullArray::new(3));
        // [[NULL], [], [NULL, NULL]] as LargeList
        let list: ArrayRef = Arc::new(LargeListArray::new(
            Field::new_list_field(DataType::Null, true).into(),
            OffsetBuffer::from_lengths(vec![1, 0, 2]),
            null_array,
            None,
        ));

        let converter = RowConverter::new(vec![SortField::new(list.data_type().clone())]).unwrap();
        let rows = converter.convert_columns(&[Arc::clone(&list)]).unwrap();
        let back = converter.convert_rows(&rows).unwrap();

        assert_eq!(&list, &back[0]);
    }

    // Test FixedSizeList<Null>
    #[test]
    fn test_fixed_size_list_null() {
        let null_array = Arc::new(NullArray::new(6));
        // [[NULL, NULL], [NULL, NULL], [NULL, NULL]] as FixedSizeList
        let list: ArrayRef = Arc::new(FixedSizeListArray::new(
            Arc::new(Field::new_list_field(DataType::Null, true)),
            2,
            null_array,
            None,
        ));

        let converter = RowConverter::new(vec![SortField::new(list.data_type().clone())]).unwrap();
        let rows = converter.convert_columns(&[Arc::clone(&list)]).unwrap();
        let back = converter.convert_rows(&rows).unwrap();

        assert_eq!(&list, &back[0]);
    }

    /// Ensure dictionaries nested within FixedSizeLists are not flattened
    #[test]
    fn test_fixed_size_list_of_dictionaries_round_trips() {
        // Build one row = ["a", "b"] as
        // `FixedSizeList<Dictionary<Int32, Utf8>, 2>`.
        let dict_dt = DataType::Dictionary(Box::new(DataType::Int32), Box::new(DataType::Utf8));
        let element_field = Arc::new(Field::new("item", dict_dt.clone(), true));
        let fsl_dt = DataType::FixedSizeList(Arc::clone(&element_field), 2);

        let values = Arc::new(StringArray::from(vec!["a", "b"]));
        let keys = Int32Array::from(vec![0, 1]);
        let dict = DictionaryArray::<Int32Type>::try_new(keys, values).unwrap();
        let fsl: ArrayRef = Arc::new(FixedSizeListArray::new(
            Arc::clone(&element_field),
            2,
            Arc::new(dict),
            None,
        ));

        assert!(RowConverter::supports_fields(&[SortField::new(
            fsl_dt.clone()
        )]));

        let converter = RowConverter::new(vec![SortField::new(fsl_dt.clone())]).unwrap();
        let rows = converter.convert_columns(&[Arc::clone(&fsl)]).unwrap();

        // Before the fix this panicked at the `.unwrap()` because
        // `convert_rows` returned `Err(InvalidArgumentError(...))`.
        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(back.len(), 1);

        // The returned array is a `FixedSizeList` with a decoded
        // (flattened) child — same self-consistent shape the other
        // list-like decoders produce for dictionary children.
        let out = back[0]
            .as_any()
            .downcast_ref::<FixedSizeListArray>()
            .expect("decoded array must be a FixedSizeListArray");
        assert_eq!(out.len(), 1);
        assert_eq!(out.value_length(), 2);
        // Child data type is the flattened `Utf8`, not the declared
        // `Dictionary`. Callers that want the dictionary back need to
        // re-encode (see the module docs).
        assert_eq!(out.values().data_type(), &DataType::Utf8);

        // Sanity: values survived the round trip.
        let values = out
            .values()
            .as_any()
            .downcast_ref::<StringArray>()
            .expect("child must be a StringArray after flattening");
        assert_eq!(values.value(0), "a");
        assert_eq!(values.value(1), "b");
    }

    // Test List<Null> with various combinations of nulls and empty lists
    #[test]
    fn test_list_null_variations() {
        // Test case: [[NULL], [], [NULL, NULL]]
        let null_array = Arc::new(NullArray::new(3));
        let list: ArrayRef = Arc::new(ListArray::new(
            Field::new_list_field(DataType::Null, true).into(),
            OffsetBuffer::from_lengths(vec![1, 0, 2]),
            null_array,
            None,
        ));

        let converter = RowConverter::new(vec![SortField::new(list.data_type().clone())]).unwrap();
        let rows = converter.convert_columns(&[Arc::clone(&list)]).unwrap();
        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(&list, &back[0]);

        // Test case: [[NULL], null, [NULL, NULL]] - middle element is null
        let null_array = Arc::new(NullArray::new(3));
        let list: ArrayRef = Arc::new(ListArray::new(
            Field::new_list_field(DataType::Null, true).into(),
            OffsetBuffer::from_lengths(vec![1, 0, 2]),
            null_array,
            Some(vec![true, false, true].into()),
        ));

        let rows = converter.convert_columns(&[Arc::clone(&list)]).unwrap();
        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(&list, &back[0]);

        // Test case: [] - empty list
        let null_array = Arc::new(NullArray::new(0));
        let list: ArrayRef = Arc::new(ListArray::new(
            Field::new_list_field(DataType::Null, true).into(),
            OffsetBuffer::from_lengths(vec![]),
            null_array,
            None,
        ));

        let rows = converter.convert_columns(&[Arc::clone(&list)]).unwrap();
        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(&list, &back[0]);

        // Test case: [[], [], []] - all empty sublists
        let null_array = Arc::new(NullArray::new(0));
        let list: ArrayRef = Arc::new(ListArray::new(
            Field::new_list_field(DataType::Null, true).into(),
            OffsetBuffer::from_lengths(vec![0, 0, 0]),
            null_array,
            None,
        ));

        let rows = converter.convert_columns(&[Arc::clone(&list)]).unwrap();
        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(&list, &back[0]);
    }

    // Test List<Null> with descending order
    #[test]
    fn test_list_null_descending() {
        let null_array = Arc::new(NullArray::new(3));
        // [[NULL], [], [NULL, NULL]]
        let list: ArrayRef = Arc::new(ListArray::new(
            Field::new_list_field(DataType::Null, true).into(),
            OffsetBuffer::from_lengths(vec![1, 0, 2]),
            null_array,
            None,
        ));

        let options = SortOptions::default().with_descending(true);
        let field = SortField::new_with_options(list.data_type().clone(), options);
        let converter = RowConverter::new(vec![field]).unwrap();
        let rows = converter.convert_columns(&[Arc::clone(&list)]).unwrap();
        let back = converter.convert_rows(&rows).unwrap();

        assert_eq!(&list, &back[0]);
    }

    // Test Struct with Null field
    #[test]
    fn test_struct_with_null_field() {
        // Struct { a: Null, b: Int32 }
        let null_array = Arc::new(NullArray::new(3));
        let int_array = Arc::new(Int32Array::from(vec![1, 2, 3]));

        let struct_array: ArrayRef = Arc::new(StructArray::new(
            vec![
                Arc::new(Field::new("a", DataType::Null, true)),
                Arc::new(Field::new("b", DataType::Int32, true)),
            ]
            .into(),
            vec![null_array, int_array],
            Some(vec![true, true, false].into()), // validity bitmap
        ));

        let converter =
            RowConverter::new(vec![SortField::new(struct_array.data_type().clone())]).unwrap();
        let rows = converter
            .convert_columns(&[Arc::clone(&struct_array)])
            .unwrap();
        let back = converter.convert_rows(&rows).unwrap();

        assert_eq!(&struct_array, &back[0]);
    }

    // Test nested Struct with Null
    #[test]
    fn test_nested_struct_with_null() {
        // Inner struct: { x: Null }
        let inner_null = Arc::new(NullArray::new(2));
        let inner_struct = Arc::new(StructArray::new(
            vec![Arc::new(Field::new("x", DataType::Null, true))].into(),
            vec![inner_null],
            None,
        ));

        // Outer struct: { inner: Struct { x: Null }, y: Int32 }
        let y_array = Arc::new(Int32Array::from(vec![10, 20]));
        let outer_struct: ArrayRef = Arc::new(StructArray::new(
            vec![
                Arc::new(Field::new("inner", inner_struct.data_type().clone(), true)),
                Arc::new(Field::new("y", DataType::Int32, true)),
            ]
            .into(),
            vec![inner_struct, y_array],
            None,
        ));

        let converter =
            RowConverter::new(vec![SortField::new(outer_struct.data_type().clone())]).unwrap();
        let rows = converter
            .convert_columns(&[Arc::clone(&outer_struct)])
            .unwrap();
        let back = converter.convert_rows(&rows).unwrap();

        assert_eq!(&outer_struct, &back[0]);
    }

    // Test Map<Utf8, Null> with various combinations of nulls and empty maps
    #[test]
    fn test_map_null_variations() {
        // Map with Null values: [{a: NULL}, {}, {b: NULL, c: NULL}]
        let keys = Arc::new(StringArray::from(vec!["a", "b", "c"])) as ArrayRef;
        let null_values = Arc::new(NullArray::new(3)) as ArrayRef;

        let offsets = OffsetBuffer::new(vec![0, 1, 1, 3].into());
        let entries_fields = vec![
            Arc::new(Field::new(
                Field::MAP_KEY_FIELD_DEFAULT_NAME,
                DataType::Utf8,
                false,
            )),
            Arc::new(Field::new(
                Field::MAP_VALUE_FIELD_DEFAULT_NAME,
                DataType::Null,
                true,
            )),
        ];
        let struct_field = Arc::new(Field::new(
            Field::MAP_ENTRIES_FIELD_DEFAULT_NAME,
            DataType::Struct(entries_fields.clone().into()),
            false,
        ));
        let entries = StructArray::new(entries_fields.into(), vec![keys, null_values], None);

        let map: ArrayRef = Arc::new(MapArray::new(
            struct_field.clone(),
            offsets,
            entries,
            None,
            false,
        ));

        let converter = RowConverter::new(vec![SortField::new(map.data_type().clone())]).unwrap();
        let rows = converter.convert_columns(&[Arc::clone(&map)]).unwrap();
        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(back.len(), 1);
        back[0].to_data().validate_full().unwrap();
        assert_eq!(&map, &back[0]);

        // Map with Null values and null map entries: [{a: NULL}, null, {b: NULL, c: NULL}]
        let keys = Arc::new(StringArray::from(vec!["a", "b", "c"])) as ArrayRef;
        let null_values = Arc::new(NullArray::new(3)) as ArrayRef;

        let offsets = OffsetBuffer::new(vec![0, 1, 1, 3].into());
        let entries_fields = vec![
            Arc::new(Field::new(
                Field::MAP_KEY_FIELD_DEFAULT_NAME,
                DataType::Utf8,
                false,
            )),
            Arc::new(Field::new(
                Field::MAP_VALUE_FIELD_DEFAULT_NAME,
                DataType::Null,
                true,
            )),
        ];
        let struct_field = Arc::new(Field::new(
            Field::MAP_ENTRIES_FIELD_DEFAULT_NAME,
            DataType::Struct(entries_fields.clone().into()),
            false,
        ));
        let entries = StructArray::new(entries_fields.into(), vec![keys, null_values], None);

        let map: ArrayRef = Arc::new(MapArray::new(
            struct_field.clone(),
            offsets,
            entries,
            Some(vec![true, false, true].into()),
            false,
        ));

        let rows = converter.convert_columns(&[Arc::clone(&map)]).unwrap();
        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(back.len(), 1);
        back[0].to_data().validate_full().unwrap();
        assert_eq!(&map, &back[0]);

        // Empty map array with Null value type
        let keys = Arc::new(StringArray::from(Vec::<&str>::new())) as ArrayRef;
        let null_values = Arc::new(NullArray::new(0)) as ArrayRef;

        let offsets = OffsetBuffer::new(vec![0i32].into());
        let entries_fields = vec![
            Arc::new(Field::new(
                Field::MAP_KEY_FIELD_DEFAULT_NAME,
                DataType::Utf8,
                false,
            )),
            Arc::new(Field::new(
                Field::MAP_VALUE_FIELD_DEFAULT_NAME,
                DataType::Null,
                true,
            )),
        ];
        let struct_field = Arc::new(Field::new(
            Field::MAP_ENTRIES_FIELD_DEFAULT_NAME,
            DataType::Struct(entries_fields.clone().into()),
            false,
        ));
        let entries = StructArray::new(entries_fields.into(), vec![keys, null_values], None);

        let map: ArrayRef = Arc::new(MapArray::new(struct_field, offsets, entries, None, false));

        let rows = converter.convert_columns(&[Arc::clone(&map)]).unwrap();
        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(back.len(), 1);
        back[0].to_data().validate_full().unwrap();
        assert_eq!(&map, &back[0]);
    }

    // Test Map<Utf8, Null> with descending order
    #[test]
    fn test_map_null_descending() {
        // [{a: NULL}, {}, {b: NULL, c: NULL}]
        let keys = Arc::new(StringArray::from(vec!["a", "b", "c"])) as ArrayRef;
        let null_values = Arc::new(NullArray::new(3)) as ArrayRef;

        let offsets = OffsetBuffer::new(vec![0, 1, 1, 3].into());
        let entries_fields = vec![
            Arc::new(Field::new(
                Field::MAP_KEY_FIELD_DEFAULT_NAME,
                DataType::Utf8,
                false,
            )),
            Arc::new(Field::new(
                Field::MAP_VALUE_FIELD_DEFAULT_NAME,
                DataType::Null,
                true,
            )),
        ];
        let struct_field = Arc::new(Field::new(
            Field::MAP_ENTRIES_FIELD_DEFAULT_NAME,
            DataType::Struct(entries_fields.clone().into()),
            false,
        ));
        let entries = StructArray::new(entries_fields.into(), vec![keys, null_values], None);

        let map: ArrayRef = Arc::new(MapArray::new(struct_field, offsets, entries, None, false));

        let options = SortOptions::default().with_descending(true);
        let field = SortField::new_with_options(map.data_type().clone(), options);
        let converter = RowConverter::new(vec![field]).unwrap();
        let rows = converter.convert_columns(&[Arc::clone(&map)]).unwrap();
        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(back.len(), 1);
        back[0].to_data().validate_full().unwrap();
        assert_eq!(&map, &back[0]);
    }

    // Test Map<Utf8, Null> all empty maps
    #[test]
    fn test_map_null_all_empty() {
        let keys = Arc::new(StringArray::from(Vec::<&str>::new())) as ArrayRef;
        let null_values = Arc::new(NullArray::new(0)) as ArrayRef;

        let offsets = OffsetBuffer::new(vec![0, 0, 0, 0].into());
        let entries_fields = vec![
            Arc::new(Field::new(
                Field::MAP_KEY_FIELD_DEFAULT_NAME,
                DataType::Utf8,
                false,
            )),
            Arc::new(Field::new(
                Field::MAP_VALUE_FIELD_DEFAULT_NAME,
                DataType::Null,
                true,
            )),
        ];
        let struct_field = Arc::new(Field::new(
            Field::MAP_ENTRIES_FIELD_DEFAULT_NAME,
            DataType::Struct(entries_fields.clone().into()),
            false,
        ));
        let entries = StructArray::new(entries_fields.into(), vec![keys, null_values], None);

        let map: ArrayRef = Arc::new(MapArray::new(struct_field, offsets, entries, None, false));

        let converter = RowConverter::new(vec![SortField::new(map.data_type().clone())]).unwrap();
        let rows = converter.convert_columns(&[Arc::clone(&map)]).unwrap();

        // All empty maps should be equal
        assert_eq!(rows.row(0), rows.row(1));
        assert_eq!(rows.row(1), rows.row(2));

        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(back.len(), 1);
        back[0].to_data().validate_full().unwrap();
        assert_eq!(&map, &back[0]);
    }

    // Test Map<Utf8, Map<Utf8, Null>> - nested map with Null leaf values
    #[test]
    fn test_nested_map_null() {
        // Inner map entries: {a: NULL, b: NULL, c: NULL}
        let inner_keys = Arc::new(StringArray::from(vec!["a", "b", "c"])) as ArrayRef;
        let inner_null_values = Arc::new(NullArray::new(3)) as ArrayRef;

        let inner_entries_fields = vec![
            Arc::new(Field::new(
                Field::MAP_KEY_FIELD_DEFAULT_NAME,
                DataType::Utf8,
                false,
            )),
            Arc::new(Field::new(
                Field::MAP_VALUE_FIELD_DEFAULT_NAME,
                DataType::Null,
                true,
            )),
        ];
        let inner_struct_field = Arc::new(Field::new(
            Field::MAP_ENTRIES_FIELD_DEFAULT_NAME,
            DataType::Struct(inner_entries_fields.clone().into()),
            false,
        ));
        let inner_entries = StructArray::new(
            inner_entries_fields.clone().into(),
            vec![inner_keys, inner_null_values],
            None,
        );

        // Inner maps: [{a: NULL}, {b: NULL, c: NULL}]
        let inner_map = Arc::new(MapArray::new(
            inner_struct_field.clone(),
            OffsetBuffer::new(vec![0, 1, 3].into()),
            inner_entries,
            None,
            false,
        )) as ArrayRef;

        // Outer map entries
        let outer_keys = Arc::new(StringArray::from(vec!["x", "y"])) as ArrayRef;

        let inner_map_type = DataType::Map(inner_struct_field.clone(), false);
        let outer_entries_fields = vec![
            Arc::new(Field::new(
                Field::MAP_KEY_FIELD_DEFAULT_NAME,
                DataType::Utf8,
                false,
            )),
            Arc::new(Field::new(
                Field::MAP_VALUE_FIELD_DEFAULT_NAME,
                inner_map_type,
                true,
            )),
        ];
        let outer_struct_field = Arc::new(Field::new(
            Field::MAP_ENTRIES_FIELD_DEFAULT_NAME,
            DataType::Struct(outer_entries_fields.clone().into()),
            false,
        ));
        let outer_entries = StructArray::new(
            outer_entries_fields.into(),
            vec![outer_keys, inner_map],
            None,
        );

        // Outer map: [{x: {a: NULL}}, {y: {b: NULL, c: NULL}}]
        let map: ArrayRef = Arc::new(MapArray::new(
            outer_struct_field,
            OffsetBuffer::new(vec![0, 1, 2].into()),
            outer_entries,
            None,
            false,
        ));

        let converter = RowConverter::new(vec![SortField::new(map.data_type().clone())]).unwrap();
        let rows = converter.convert_columns(&[Arc::clone(&map)]).unwrap();
        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(back.len(), 1);
        back[0].to_data().validate_full().unwrap();
        assert_eq!(&map, &back[0]);
    }

    // Test List<Map<Utf8, Null>> - list containing maps with Null values
    #[test]
    fn test_list_of_map_null() {
        // Map entries: {a: NULL, b: NULL, c: NULL}
        let keys = Arc::new(StringArray::from(vec!["a", "b", "c"])) as ArrayRef;
        let null_values = Arc::new(NullArray::new(3)) as ArrayRef;

        let entries_fields = vec![
            Arc::new(Field::new(
                Field::MAP_KEY_FIELD_DEFAULT_NAME,
                DataType::Utf8,
                false,
            )),
            Arc::new(Field::new(
                Field::MAP_VALUE_FIELD_DEFAULT_NAME,
                DataType::Null,
                true,
            )),
        ];
        let struct_field = Arc::new(Field::new(
            Field::MAP_ENTRIES_FIELD_DEFAULT_NAME,
            DataType::Struct(entries_fields.clone().into()),
            false,
        ));
        let entries = StructArray::new(entries_fields.into(), vec![keys, null_values], None);

        // Maps: [{a: NULL}, {}, {b: NULL, c: NULL}]
        let map_array = Arc::new(MapArray::new(
            struct_field.clone(),
            OffsetBuffer::new(vec![0, 1, 1, 3].into()),
            entries,
            None,
            false,
        )) as ArrayRef;

        let map_type = DataType::Map(struct_field, false);
        // List of maps: [[{a: NULL}], [{}, {b: NULL, c: NULL}]]
        let list: ArrayRef = Arc::new(ListArray::new(
            Arc::new(Field::new_list_field(map_type, true)),
            OffsetBuffer::new(vec![0, 1, 3].into()),
            map_array,
            None,
        ));

        let converter = RowConverter::new(vec![SortField::new(list.data_type().clone())]).unwrap();
        let rows = converter.convert_columns(&[Arc::clone(&list)]).unwrap();
        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(&list, &back[0]);
    }

    // Test Map<Utf8, List<Null>> - map with list values containing Null
    #[test]
    fn test_map_of_list_null() {
        // Inner list values: [NULL, NULL, NULL]
        let null_array = Arc::new(NullArray::new(3)) as ArrayRef;
        // Lists: [[NULL], [], [NULL, NULL]]
        let list_array = Arc::new(ListArray::new(
            Arc::new(Field::new_list_field(DataType::Null, true)),
            OffsetBuffer::from_lengths(vec![1, 0, 2]),
            null_array,
            None,
        )) as ArrayRef;

        let keys = Arc::new(StringArray::from(vec!["a", "b", "c"])) as ArrayRef;

        let list_type = list_array.data_type().clone();
        let entries_fields = vec![
            Arc::new(Field::new(
                Field::MAP_KEY_FIELD_DEFAULT_NAME,
                DataType::Utf8,
                false,
            )),
            Arc::new(Field::new(
                Field::MAP_VALUE_FIELD_DEFAULT_NAME,
                list_type,
                true,
            )),
        ];
        let struct_field = Arc::new(Field::new(
            Field::MAP_ENTRIES_FIELD_DEFAULT_NAME,
            DataType::Struct(entries_fields.clone().into()),
            false,
        ));
        let entries = StructArray::new(entries_fields.into(), vec![keys, list_array], None);

        // Map: [{a: [NULL], b: [], c: [NULL, NULL]}]
        let map: ArrayRef = Arc::new(MapArray::new(
            struct_field,
            OffsetBuffer::new(vec![0, 3].into()),
            entries,
            None,
            false,
        ));

        let converter = RowConverter::new(vec![SortField::new(map.data_type().clone())]).unwrap();
        let rows = converter.convert_columns(&[Arc::clone(&map)]).unwrap();
        let back = converter.convert_rows(&rows).unwrap();
        assert_eq!(back.len(), 1);
        back[0].to_data().validate_full().unwrap();
        assert_eq!(&map, &back[0]);
    }

    #[test]
    fn empty_row_iter_next_back() {
        let rows = RowConverter::new(vec![SortField::new(DataType::UInt8)])
            .unwrap()
            .empty_rows(0, 0);
        let mut rows_iter = rows.iter();
        assert_eq!(rows_iter.next_back(), None);
        assert_eq!(rows_iter.next_back(), None);
        assert_eq!(rows_iter.next_back(), None);
    }

    /// Round-trip through `with_skip_utf8_validate` confirms skipping validation preserves values.
    #[test]
    fn test_row_parser_skip_utf8_validation_roundtrip() {
        let converter = RowConverter::new(vec![SortField::new(DataType::Utf8)]).unwrap();
        let array = StringArray::from(vec!["arrow", "rust"]);
        let rows = converter.convert_columns(&[Arc::new(array) as _]).unwrap();
        let binary = rows.try_into_binary().expect("fits in i32 offsets");

        // SAFETY: bytes come from this RowConverter and are known-valid UTF-8.
        let parser = unsafe { RowParser::with_skip_utf8_validate(Arc::clone(&converter.fields)) };

        let decoded = converter
            .convert_rows(binary.iter().map(|b| parser.parse(b.unwrap())))
            .unwrap();
        let got: Vec<_> = decoded[0].as_string::<i32>().iter().flatten().collect();
        assert_eq!(got, vec!["arrow", "rust"]);
    }

    #[test]
    fn row_iter_next_back() {
        let row_converter = RowConverter::new(vec![SortField::new(DataType::UInt8)]).unwrap();
        let mut rng = StdRng::seed_from_u64(42);
        let array = generate_primitive_array::<UInt8Type>(&mut rng, 100, 0.8);
        let rows = row_converter.convert_columns(&[Arc::new(array)]).unwrap();

        let mut rows_iter = rows.iter();
        let mut bytes: Vec<u8> = vec![];

        while let Some(row) = rows_iter.next_back() {
            bytes.extend(row.data.iter().rev());
        }

        bytes.reverse();

        assert_eq!(
            bytes,
            &rows.buffer.as_slice()[..*rows.offsets.last().unwrap()]
        );

        assert_eq!(rows_iter.next_back(), None);
        assert_eq!(rows_iter.next(), None);
    }
}