arrow 28.0.0

// 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
//! [here](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::{as_primitive_array, 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 mut 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 = as_primitive_array::<Int32Type>(converted[0].as_ref());
//! assert_eq!(c1.values(), &[-1, 4, 0, 3]);
//!
//! let c2 = as_string_array(converted[1].as_ref());
//! let c2_values: Vec<_> = c2.iter().flatten().collect();
//! assert_eq!(&c2_values, &["a", "f", "c", "e"]);
//! ```
//!
//! # 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 mut 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))
//! }
//! ```
//!
//! [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]:[https://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.83.1080&rep=rep1&type=pdf]
//! [`memcmp`]:[https://www.man7.org/linux/man-pages/man3/memcmp.3.html]
//! [`lexsort`]: crate::compute::kernels::sort::lexsort
//! [compared]: PartialOrd
//! [compare]: PartialOrd

use std::cmp::Ordering;
use std::hash::{Hash, Hasher};
use std::sync::Arc;

use arrow_array::cast::*;
use arrow_array::*;

use crate::compute::SortOptions;
use crate::datatypes::*;
use crate::error::{ArrowError, Result};
use crate::row::dictionary::{
    compute_dictionary_mapping, decode_dictionary, encode_dictionary,
};
use crate::row::fixed::{decode_bool, decode_primitive};
use crate::row::interner::OrderPreservingInterner;
use crate::row::variable::{decode_binary, decode_string};
use crate::{downcast_dictionary_array, downcast_primitive_array};

mod dictionary;
mod fixed;
mod interner;
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.
///
/// ## 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 32-byte 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`.
///
/// Note the following example encodings use a block size of 4 bytes,
/// as opposed to 32 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
///
/// [`RowConverter`] needs to support converting dictionary encoded arrays with unsorted, and
/// potentially distinct dictionaries. One simple mechanism to avoid this would be to reverse
/// the dictionary encoding, and encode the array values directly, however, this would lose
/// the benefits of dictionary encoding to reduce memory and CPU consumption.
///
/// As such the [`RowConverter`] creates an order-preserving mapping
/// for each dictionary encoded column, which allows new dictionary
/// values to be added whilst preserving the sort order.
///
/// A null dictionary value is encoded as `0_u8`.
///
/// A non-null dictionary value is encoded as `1_u8` followed by a null-terminated byte array
/// key determined by the order-preserving dictionary encoding
///
/// ```text
/// ┌──────────┐                 ┌─────┐
/// │  "Bar"   │ ───────────────▶│ 01  │
/// └──────────┘                 └─────┘
/// ┌──────────┐                 ┌─────┬─────┐
/// │"Fabulous"│ ───────────────▶│ 01  │ 02  │
/// └──────────┘                 └─────┴─────┘
/// ┌──────────┐                 ┌─────┐
/// │  "Soup"  │ ───────────────▶│ 05  │
/// └──────────┘                 └─────┘
/// ┌──────────┐                 ┌─────┐
/// │   "ZZ"   │ ───────────────▶│ 07  │
/// └──────────┘                 └─────┘
///
/// Example Order Preserving Mapping
/// ```
/// Using the map above, the corresponding row format will be
///
/// ```text
///                           ┌─────┬─────┬─────┬─────┐
///    "Fabulous"             │ 01  │ 03  │ 05  │ 00  │
///                           └─────┴─────┴─────┴─────┘
///
///                           ┌─────┬─────┬─────┐
///    "ZZ"                   │ 01  │ 07  │ 00  │
///                           └─────┴─────┴─────┘
///
///                           ┌─────┐
///     NULL                  │ 00  │
///                           └─────┘
///
///      Input                  Row Format
/// ```
/// # 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`
///
/// ## 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]>,
    /// interning state for column `i`, if column`i` is a dictionary
    interners: Vec<Option<Box<OrderPreservingInterner>>>,
}

/// 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> {
        if !Self::supports_fields(&fields) {
            return Err(ArrowError::NotYetImplemented(format!(
                "not yet implemented: {:?}",
                fields
            )));
        }

        let interners = (0..fields.len()).map(|_| None).collect();
        Ok(Self {
            fields: fields.into(),
            interners,
        })
    }

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

    /// Convert [`ArrayRef`] columns into [`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`]
    pub fn convert_columns(&mut self, columns: &[ArrayRef]) -> Result<Rows> {
        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()
            )));
        }

        let dictionaries = columns
            .iter()
            .zip(&mut self.interners)
            .zip(self.fields.iter())
            .map(|((column, interner), 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()
                    )));
                }

                let values = downcast_dictionary_array! {
                    column => column.values(),
                    _ => return Ok(None)
                };

                let interner = interner.get_or_insert_with(Default::default);

                let mapping: Vec<_> = compute_dictionary_mapping(interner, values)
                    .into_iter()
                    .map(|maybe_interned| {
                        maybe_interned.map(|interned| interner.normalized_key(interned))
                    })
                    .collect();

                Ok(Some(mapping))
            })
            .collect::<Result<Vec<_>>>()?;

        let config = RowConfig {
            fields: Arc::clone(&self.fields),
            // Don't need to validate UTF-8 as came from arrow array
            validate_utf8: false,
        };
        let mut rows = new_empty_rows(columns, &dictionaries, config);

        for ((column, field), dictionary) in
            columns.iter().zip(self.fields.iter()).zip(dictionaries)
        {
            // We encode a column at a time to minimise dispatch overheads
            encode_column(&mut rows, column, field.options, dictionary.as_deref())
        }

        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(rows)
    }

    /// Convert [`Rows`] columns into [`ArrayRef`]
    ///
    /// # Panics
    ///
    /// Panics if the rows were not produced by this [`RowConverter`]
    pub fn convert_rows<'a, I>(&self, rows: I) -> Result<Vec<ArrayRef>>
    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();

        self.fields
            .iter()
            .zip(&self.interners)
            .map(|(field, interner)| {
                // SAFETY
                // We have validated that the rows came from this [`RowConverter`]
                // and therefore must be valid
                unsafe {
                    decode_column(field, &mut rows, interner.as_deref(), 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))
    }

    /// 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.interners.capacity()
                * std::mem::size_of::<Option<Box<OrderPreservingInterner>>>()
            + self
                .interners
                .iter()
                .filter_map(|x| x.as_ref().map(|x| x.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,
            },
        }
    }

    /// 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)]
pub struct Rows {
    /// Underlying row bytes
    buffer: Box<[u8]>,
    /// Row `i` has data `&buffer[offsets[i]..offsets[i+1]]`
    offsets: Box<[usize]>,
    /// The config for these rows
    config: RowConfig,
}

impl Rows {
    pub fn row(&self, row: usize) -> Row<'_> {
        let end = self.offsets[row + 1];
        let start = self.offsets[row];
        Row {
            data: &self.buffer[start..end],
            config: &self.config,
        }
    }

    pub fn num_rows(&self) -> usize {
        self.offsets.len() - 1
    }

    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.len()
            + self.offsets.len() * std::mem::size_of::<usize>()
    }
}

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;
        }
        let row = self.rows.row(self.start);
        self.start += 1;
        Some(row)
    }

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

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

impl<'a> DoubleEndedIterator for RowsIter<'a> {
    fn next_back(&mut self) -> Option<Self::Item> {
        if self.end == self.start {
            return None;
        }
        let row = self.rows.row(self.end);
        self.end -= 1;
        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.to_vec(),
            config: self.config.clone(),
        }
    }
}

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

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

impl<'a> Eq for Row<'a> {}

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

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

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

impl<'a> AsRef<[u8]> for Row<'a> {
    #[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: Vec<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> {
        self.row().partial_cmp(&other.row())
    }
}

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,
    }
}

/// Computes the length of each encoded [`Rows`] and returns an empty [`Rows`]
fn new_empty_rows(
    cols: &[ArrayRef],
    dictionaries: &[Option<Vec<Option<&[u8]>>>],
    config: RowConfig,
) -> Rows {
    use fixed::FixedLengthEncoding;

    let num_rows = cols.first().map(|x| x.len()).unwrap_or(0);
    let mut lengths = vec![0; num_rows];

    for (array, dict) in cols.iter().zip(dictionaries) {
        downcast_primitive_array! {
            array => lengths.iter_mut().for_each(|x| *x += fixed::encoded_len(array)),
            DataType::Null => {},
            DataType::Boolean => lengths.iter_mut().for_each(|x| *x += bool::ENCODED_LEN),
            DataType::Binary => as_generic_binary_array::<i32>(array)
                .iter()
                .zip(lengths.iter_mut())
                .for_each(|(slice, length)| *length += variable::encoded_len(slice)),
            DataType::LargeBinary => as_generic_binary_array::<i64>(array)
                .iter()
                .zip(lengths.iter_mut())
                .for_each(|(slice, length)| *length += variable::encoded_len(slice)),
            DataType::Utf8 => as_string_array(array)
                .iter()
                .zip(lengths.iter_mut())
                .for_each(|(slice, length)| {
                    *length += variable::encoded_len(slice.map(|x| x.as_bytes()))
                }),
            DataType::LargeUtf8 => as_largestring_array(array)
                .iter()
                .zip(lengths.iter_mut())
                .for_each(|(slice, length)| {
                    *length += variable::encoded_len(slice.map(|x| x.as_bytes()))
                }),
            DataType::Dictionary(_, _) => downcast_dictionary_array! {
                array => {
                    let dict = dict.as_ref().unwrap();
                    for (v, length) in array.keys().iter().zip(lengths.iter_mut()) {
                        match v.and_then(|v| dict[v as usize]) {
                            Some(k) => *length += k.len() + 1,
                            None => *length += 1,
                        }
                    }
                }
                _ => unreachable!(),
            }
            _ => unreachable!(),
        }
    }

    let mut offsets = Vec::with_capacity(num_rows + 1);
    offsets.push(0);

    // 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
    // as identifying the offsets of the written rows
    let mut cur_offset = 0_usize;
    for l in lengths {
        offsets.push(cur_offset);
        cur_offset = cur_offset.checked_add(l).expect("overflow");
    }

    let buffer = vec![0_u8; cur_offset];

    Rows {
        buffer: buffer.into(),
        offsets: offsets.into(),
        config,
    }
}

/// Encodes a column to the provided [`Rows`] incrementing the offsets as it progresses
fn encode_column(
    out: &mut Rows,
    column: &ArrayRef,
    opts: SortOptions,
    dictionary: Option<&[Option<&[u8]>]>,
) {
    downcast_primitive_array! {
        column => fixed::encode(out, column, opts),
        DataType::Null => {}
        DataType::Boolean => fixed::encode(out, as_boolean_array(column), opts),
        DataType::Binary => {
            variable::encode(out, as_generic_binary_array::<i32>(column).iter(), opts)
        }
        DataType::LargeBinary => {
            variable::encode(out, as_generic_binary_array::<i64>(column).iter(), opts)
        }
        DataType::Utf8 => variable::encode(
            out,
            as_string_array(column).iter().map(|x| x.map(|x| x.as_bytes())),
            opts,
        ),
        DataType::LargeUtf8 => variable::encode(
            out,
            as_largestring_array(column)
                .iter()
                .map(|x| x.map(|x| x.as_bytes())),
            opts,
        ),
        DataType::Dictionary(_, _) => downcast_dictionary_array! {
            column => encode_dictionary(out, column, dictionary.unwrap(), opts),
            _ => unreachable!()
        }
        _ => unreachable!(),
    }
}

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

macro_rules! decode_dictionary_helper {
    ($t:ty, $interner:ident, $v:ident, $options:ident, $rows:ident) => {
        Arc::new(decode_dictionary::<$t>(
            $interner.unwrap(),
            $v.as_ref(),
            $options,
            $rows,
        )?)
    };
}

/// 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]],
    interner: Option<&OrderPreservingInterner>,
    validate_utf8: bool,
) -> Result<ArrayRef> {
    let options = field.options;
    let data_type = field.data_type.clone();
    let array: ArrayRef = downcast_primitive! {
        data_type => (decode_primitive_helper, rows, data_type, options),
        DataType::Null => 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::Utf8 => Arc::new(decode_string::<i32>(rows, options, validate_utf8)),
        DataType::LargeUtf8 => Arc::new(decode_string::<i64>(rows, options, validate_utf8)),
        DataType::Dictionary(k, v) => downcast_integer! {
            k.as_ref() => (decode_dictionary_helper, interner, v, options, rows),
            _ => unreachable!()
        },
        _ => {
            return Err(ArrowError::NotYetImplemented(format!(
                "converting {} row is not supported",
                field.data_type
            )))
        }
    };
    Ok(array)
}

#[cfg(test)]
mod tests {
    use std::sync::Arc;

    use rand::distributions::uniform::SampleUniform;
    use rand::distributions::{Distribution, Standard};
    use rand::{thread_rng, Rng};

    use arrow_array::NullArray;

    use crate::array::{
        BinaryArray, BooleanArray, DictionaryArray, Float32Array, GenericStringArray,
        Int16Array, Int32Array, OffsetSizeTrait, PrimitiveArray,
        PrimitiveDictionaryBuilder, StringArray,
    };
    use crate::compute::{LexicographicalComparator, SortColumn};
    use crate::util::display::array_value_to_string;

    use super::*;

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

        assert_eq!(rows.offsets.as_ref(), &[0, 8, 16, 24, 32, 40, 48, 56]);
        assert_eq!(
            rows.buffer.as_ref(),
            &[
                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);
        }
    }

    #[test]
    fn test_decimal128() {
        let mut 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 mut 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 mut 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 mut converter = RowConverter::new(vec![SortField::new_with_options(
            DataType::Boolean,
            SortOptions {
                descending: true,
                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 mut 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().data().clone());
        let dict_with_tz = dict.with_values(&values.with_timezone("+02:00".to_string()));
        let d = DataType::Dictionary(
            Box::new(DataType::Int32),
            Box::new(DataType::Timestamp(
                TimeUnit::Nanosecond,
                Some("+02:00".to_string()),
            )),
        );

        assert_eq!(dict_with_tz.data_type(), &d);
        let mut 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(), &d);
    }

    #[test]
    fn test_null_encoding() {
        let col = Arc::new(NullArray::new(10));
        let mut converter =
            RowConverter::new(vec![SortField::new(DataType::Null)]).unwrap();
        let rows = converter.convert_columns(&[col]).unwrap();
        assert_eq!(rows.num_rows(), 10);
        assert_eq!(rows.row(1).data.len(), 0);
    }

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

        let mut 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::BLOCK_SIZE]),
            Some(vec![0_u8; variable::BLOCK_SIZE + 1]),
            Some(vec![1_u8; 6]),
            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::BLOCK_SIZE]),
            Some(vec![0xFF_u8; variable::BLOCK_SIZE + 1]),
        ])) as ArrayRef;

        let mut 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 mut converter = RowConverter::new(vec![SortField::new_with_options(
            DataType::Binary,
            SortOptions {
                descending: true,
                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);
    }

    #[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 mut converter =
            RowConverter::new(vec![SortField::new(a.data_type().clone())]).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();
        assert_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();
        assert_eq!(&cols[0], &b);

        let mut converter = RowConverter::new(vec![SortField::new_with_options(
            a.data_type().clone(),
            SortOptions {
                descending: true,
                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();
        assert_eq!(&cols[0], &a);
    }

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

    #[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 mut converter = RowConverter::new(vec![SortField::new(data_type)]).unwrap();
        let rows = converter
            .convert_columns(&[Arc::new(DictionaryArray::<Int32Type>::from(data))])
            .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]
    #[should_panic(expected = "Invalid UTF-8 sequence")]
    fn test_invalid_utf8() {
        let mut 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 = "rows were not produced by this RowConverter")]
    fn test_different_converter() {
        let values = Arc::new(Int32Array::from_iter([Some(1), Some(-1)]));
        let mut 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 generate_primitive_array<K>(len: usize, valid_percent: f64) -> PrimitiveArray<K>
    where
        K: ArrowPrimitiveType,
        Standard: Distribution<K::Native>,
    {
        let mut rng = thread_rng();
        (0..len)
            .map(|_| rng.gen_bool(valid_percent).then(|| rng.gen()))
            .collect()
    }

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

    fn generate_dictionary<K>(
        values: ArrayRef,
        len: usize,
        valid_percent: f64,
    ) -> DictionaryArray<K>
    where
        K: ArrowDictionaryKeyType,
        K::Native: SampleUniform,
    {
        let mut rng = thread_rng();
        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.gen_bool(valid_percent)
                    .then(|| rng.gen_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.data().clone())
            .build()
            .unwrap();

        DictionaryArray::from(data)
    }

    fn generate_column(len: usize) -> ArrayRef {
        let mut rng = thread_rng();
        match rng.gen_range(0..9) {
            0 => Arc::new(generate_primitive_array::<Int32Type>(len, 0.8)),
            1 => Arc::new(generate_primitive_array::<UInt32Type>(len, 0.8)),
            2 => Arc::new(generate_primitive_array::<Int64Type>(len, 0.8)),
            3 => Arc::new(generate_primitive_array::<UInt64Type>(len, 0.8)),
            4 => Arc::new(generate_primitive_array::<Float32Type>(len, 0.8)),
            5 => Arc::new(generate_primitive_array::<Float64Type>(len, 0.8)),
            6 => Arc::new(generate_strings::<i32>(len, 0.8)),
            7 => Arc::new(generate_dictionary::<Int64Type>(
                // Cannot test dictionaries containing null values because of #2687
                Arc::new(generate_strings::<i32>(rng.gen_range(1..len), 1.0)),
                len,
                0.8,
            )),
            8 => Arc::new(generate_dictionary::<Int64Type>(
                // Cannot test dictionaries containing null values because of #2687
                Arc::new(generate_primitive_array::<Int64Type>(
                    rng.gen_range(1..len),
                    1.0,
                )),
                len,
                0.8,
            )),
            _ => unreachable!(),
        }
    }

    fn print_row(cols: &[SortColumn], row: usize) -> String {
        let t: Vec<_> = cols
            .iter()
            .map(|x| array_value_to_string(&x.values, row).unwrap())
            .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(",")
    }

    #[test]
    #[cfg_attr(miri, ignore)]
    fn fuzz_test() {
        for _ in 0..100 {
            let mut rng = thread_rng();
            let num_columns = rng.gen_range(1..5);
            let len = rng.gen_range(5..100);
            let arrays: Vec<_> = (0..num_columns).map(|_| generate_column(len)).collect();

            let options: Vec<_> = (0..num_columns)
                .map(|_| SortOptions {
                    descending: rng.gen_bool(0.5),
                    nulls_first: rng.gen_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 = options
                .into_iter()
                .zip(&arrays)
                .map(|(o, a)| SortField::new_with_options(a.data_type().clone(), o))
                .collect();

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

            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)
                    );
                }
            }

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