arrow_ipc/

reader.rs

// SPDX-License-Identifier: Apache-2.0
// SPDX-FileCopyrightText: See the NOTICE file distributed with this work for additional information regarding copyright ownership
// SPDX-FileContributor: Licensed to the Apache Software Foundation (ASF) under one or more contributor license agreements

//! Arrow IPC File and Stream Readers
//!
//! # Notes
//!
//! The [`FileReader`] and [`StreamReader`] have similar interfaces,
//! however the [`FileReader`] expects a reader that supports [`Seek`]ing
//!
//! [`Seek`]: std::io::Seek

mod stream;
pub use stream::*;

use arrow_select::concat;

use flatbuffers::{VectorIter, VerifierOptions};
use std::collections::{HashMap, VecDeque};
use std::fmt;
use std::io::{BufReader, Read, Seek, SeekFrom};
use std::sync::Arc;

use arrow_array::*;
use arrow_buffer::{
    ArrowNativeType, BooleanBuffer, Buffer, MutableBuffer, NullBuffer, ScalarBuffer,
};
use arrow_data::{ArrayData, ArrayDataBuilder, UnsafeFlag};
use arrow_schema::*;

use crate::compression::{CompressionCodec, DecompressionContext};
use crate::r#gen::Message;
use crate::{Block, CONTINUATION_MARKER, FieldNode, MetadataVersion};
use DataType::*;

/// Read a buffer based on offset and length
/// From <https://github.com/apache/arrow/blob/6a936c4ff5007045e86f65f1a6b6c3c955ad5103/format/Message.fbs#L58>
/// Each constituent buffer is first compressed with the indicated
/// compressor, and then written with the uncompressed length in the first 8
/// bytes as a 64-bit little-endian signed integer followed by the compressed
/// buffer bytes (and then padding as required by the protocol). The
/// uncompressed length may be set to -1 to indicate that the data that
/// follows is not compressed, which can be useful for cases where
/// compression does not yield appreciable savings.
fn read_buffer(
    buf: &crate::Buffer,
    a_data: &Buffer,
    compression_codec: Option<CompressionCodec>,
    decompression_context: &mut DecompressionContext,
) -> Result<Buffer, ArrowError> {
    let start_offset = buf.offset() as usize;
    let buf_data = a_data
        .try_slice_with_length(start_offset, buf.length() as usize)
        .map_err(|e| ArrowError::IpcError(e.to_string()))?;
    // corner case: empty buffer
    match (buf_data.is_empty(), compression_codec) {
        (true, _) | (_, None) => Ok(buf_data),
        (false, Some(decompressor)) => {
            decompressor.decompress_to_buffer(&buf_data, decompression_context)
        }
    }
}

impl RecordBatchDecoder<'_> {
    /// Coordinates reading arrays based on data types.
    ///
    /// `variadic_counts` encodes the number of buffers to read for variadic types (e.g., Utf8View, BinaryView)
    /// When encounter such types, we pop from the front of the queue to get the number of buffers to read.
    ///
    /// Notes:
    /// * In the IPC format, null buffers are always set, but may be empty. We discard them if an array has 0 nulls
    /// * Numeric values inside list arrays are often stored as 64-bit values regardless of their data type size.
    ///   We thus:
    ///     - check if the bit width of non-64-bit numbers is 64, and
    ///     - read the buffer as 64-bit (signed integer or float), and
    ///     - cast the 64-bit array to the appropriate data type
    fn create_array(
        &mut self,
        field: &Field,
        variadic_counts: &mut VecDeque<i64>,
    ) -> Result<ArrayRef, ArrowError> {
        let data_type = field.data_type();
        match data_type {
            Utf8 | Binary | LargeBinary | LargeUtf8 => {
                let field_node = self.next_node(field)?;
                let buffers = [
                    self.next_buffer()?,
                    self.next_buffer()?,
                    self.next_buffer()?,
                ];
                self.create_primitive_array(field_node, data_type, &buffers)
            }
            BinaryView | Utf8View => {
                let count = self.next_variadic_buffer_count(variadic_counts, data_type)?;
                let buffers = (0..count)
                    .map(|_| self.next_buffer())
                    .collect::<Result<Vec<_>, _>>()?;
                let field_node = self.next_node(field)?;
                self.create_primitive_array(field_node, data_type, &buffers)
            }
            FixedSizeBinary(_) => {
                let field_node = self.next_node(field)?;
                let buffers = [self.next_buffer()?, self.next_buffer()?];
                self.create_primitive_array(field_node, data_type, &buffers)
            }
            List(list_field) | LargeList(list_field) | Map(list_field, _) => {
                let list_node = self.next_node(field)?;
                let list_buffers = [self.next_buffer()?, self.next_buffer()?];
                let values = self.create_array(list_field, variadic_counts)?;
                self.create_list_array(list_node, data_type, &list_buffers, values)
            }
            ListView(list_field) | LargeListView(list_field) => {
                let list_node = self.next_node(field)?;
                let list_buffers = [
                    self.next_buffer()?, // null buffer
                    self.next_buffer()?, // offsets
                    self.next_buffer()?, // sizes
                ];
                let values = self.create_array(list_field, variadic_counts)?;
                self.create_list_view_array(list_node, data_type, &list_buffers, values)
            }
            FixedSizeList(list_field, _) => {
                let list_node = self.next_node(field)?;
                let list_buffers = [self.next_buffer()?];
                let values = self.create_array(list_field, variadic_counts)?;
                self.create_list_array(list_node, data_type, &list_buffers, values)
            }
            Struct(struct_fields) => {
                let struct_node = self.next_node(field)?;
                let null_buffer = self.next_buffer()?;

                // read the arrays for each field
                let mut struct_arrays = Vec::with_capacity(struct_fields.len());
                // TODO investigate whether just knowing the number of buffers could
                // still work
                for struct_field in struct_fields {
                    let child = self.create_array(struct_field, variadic_counts)?;
                    struct_arrays.push(child);
                }
                self.create_struct_array(struct_node, null_buffer, struct_fields, struct_arrays)
            }
            RunEndEncoded(run_ends_field, values_field) => {
                let run_node = self.next_node(field)?;
                let run_ends = self.create_array(run_ends_field, variadic_counts)?;
                let values = self.create_array(values_field, variadic_counts)?;

                let run_array_length = run_node.length() as usize;
                let builder = ArrayData::builder(data_type.clone())
                    .len(run_array_length)
                    .offset(0)
                    .add_child_data(run_ends.into_data())
                    .add_child_data(values.into_data())
                    .null_count(run_node.null_count() as usize);

                self.create_array_from_builder(builder)
            }
            // Create dictionary array from RecordBatch
            Dictionary(_, _) => {
                let index_node = self.next_node(field)?;
                let index_buffers = [self.next_buffer()?, self.next_buffer()?];

                #[expect(deprecated)]
                let dict_id = field.dict_id().ok_or_else(|| {
                    ArrowError::ParseError(format!("Field {field} does not have dict id"))
                })?;

                let value_array = match self.dictionaries_by_id.get(&dict_id) {
                    Some(array) => array.clone(),
                    None => {
                        // Per the IPC spec, dictionary batches may be omitted when all
                        // values in the column are null. In that case we synthesize an
                        // empty values array so decoding can proceed.
                        if let Dictionary(_, value_type) = data_type {
                            arrow_array::new_empty_array(value_type.as_ref())
                        } else {
                            unreachable!()
                        }
                    }
                };

                self.create_dictionary_array(index_node, data_type, &index_buffers, value_array)
            }
            Union(fields, mode) => {
                let union_node = self.next_node(field)?;
                let len = union_node.length() as usize;

                // In V4, union types has validity bitmap
                // In V5 and later, union types have no validity bitmap
                if self.version < MetadataVersion::V5 {
                    self.next_buffer()?;
                }

                let type_ids: ScalarBuffer<i8> = self
                    .next_buffer()?
                    .try_slice_with_length(0, len)
                    .map_err(|e| ArrowError::IpcError(e.to_string()))?
                    .into();

                let value_offsets = match mode {
                    UnionMode::Dense => {
                        let offsets: ScalarBuffer<i32> = self
                            .next_buffer()?
                            .try_slice_with_length(0, len * 4)
                            .map_err(|e| ArrowError::IpcError(e.to_string()))?
                            .into();
                        Some(offsets)
                    }
                    UnionMode::Sparse => None,
                };

                let mut children = Vec::with_capacity(fields.len());

                for (_id, field) in fields.iter() {
                    let child = self.create_array(field, variadic_counts)?;
                    children.push(child);
                }

                let array = if self.skip_validation.get() {
                    // safety: flag can only be set via unsafe code
                    unsafe {
                        UnionArray::new_unchecked(fields.clone(), type_ids, value_offsets, children)
                    }
                } else {
                    UnionArray::try_new(fields.clone(), type_ids, value_offsets, children)?
                };
                Ok(Arc::new(array))
            }
            Null => {
                let node = self.next_node(field)?;
                let length = node.length();
                let null_count = node.null_count();

                if length != null_count {
                    return Err(ArrowError::SchemaError(format!(
                        "Field {field} of NullArray has unequal null_count {null_count} and len {length}"
                    )));
                }

                let builder = ArrayData::builder(data_type.clone())
                    .len(length as usize)
                    .offset(0);
                self.create_array_from_builder(builder)
            }
            _ => {
                let field_node = self.next_node(field)?;
                let buffers = [self.next_buffer()?, self.next_buffer()?];
                self.create_primitive_array(field_node, data_type, &buffers)
            }
        }
    }

    /// Reads the correct number of buffers based on data type and null_count, and creates a
    /// primitive array ref
    fn create_primitive_array(
        &self,
        field_node: &FieldNode,
        data_type: &DataType,
        buffers: &[Buffer],
    ) -> Result<ArrayRef, ArrowError> {
        let length = field_node.length() as usize;
        let null_buffer = (field_node.null_count() > 0).then_some(buffers[0].clone());
        let mut builder = match data_type {
            Utf8 | Binary | LargeBinary | LargeUtf8 => {
                // read 3 buffers: null buffer (optional), offsets buffer and data buffer
                ArrayData::builder(data_type.clone())
                    .len(length)
                    .buffers(buffers[1..3].to_vec())
                    .null_bit_buffer(null_buffer)
            }
            BinaryView | Utf8View => ArrayData::builder(data_type.clone())
                .len(length)
                .buffers(buffers[1..].to_vec())
                .null_bit_buffer(null_buffer),
            _ if data_type.is_primitive() || matches!(data_type, Boolean | FixedSizeBinary(_)) => {
                // read 2 buffers: null buffer (optional) and data buffer
                ArrayData::builder(data_type.clone())
                    .len(length)
                    .add_buffer(buffers[1].clone())
                    .null_bit_buffer(null_buffer)
            }
            t => unreachable!("Data type {:?} either unsupported or not primitive", t),
        };

        builder = builder.null_count(field_node.null_count() as usize);

        self.create_array_from_builder(builder)
    }

    /// Update the ArrayDataBuilder based on settings in this decoder
    fn create_array_from_builder(&self, builder: ArrayDataBuilder) -> Result<ArrayRef, ArrowError> {
        let mut builder = builder.align_buffers(!self.require_alignment);
        if self.skip_validation.get() {
            // SAFETY: flag can only be set via unsafe code
            unsafe { builder = builder.skip_validation(true) }
        }
        Ok(make_array(builder.build()?))
    }

    /// Reads the correct number of buffers based on list type and null_count, and creates a
    /// list array ref
    fn create_list_array(
        &self,
        field_node: &FieldNode,
        data_type: &DataType,
        buffers: &[Buffer],
        child_array: ArrayRef,
    ) -> Result<ArrayRef, ArrowError> {
        let null_buffer = (field_node.null_count() > 0).then_some(buffers[0].clone());
        let length = field_node.length() as usize;
        let child_data = child_array.into_data();
        let mut builder = match data_type {
            List(_) | LargeList(_) | Map(_, _) => ArrayData::builder(data_type.clone())
                .len(length)
                .add_buffer(buffers[1].clone())
                .add_child_data(child_data)
                .null_bit_buffer(null_buffer),

            FixedSizeList(_, _) => ArrayData::builder(data_type.clone())
                .len(length)
                .add_child_data(child_data)
                .null_bit_buffer(null_buffer),

            _ => unreachable!("Cannot create list or map array from {:?}", data_type),
        };

        builder = builder.null_count(field_node.null_count() as usize);

        self.create_array_from_builder(builder)
    }

    fn create_list_view_array(
        &self,
        field_node: &FieldNode,
        data_type: &DataType,
        buffers: &[Buffer],
        child_array: ArrayRef,
    ) -> Result<ArrayRef, ArrowError> {
        assert!(matches!(data_type, ListView(_) | LargeListView(_)));

        let null_buffer = (field_node.null_count() > 0).then_some(buffers[0].clone());
        let length = field_node.length() as usize;
        let child_data = child_array.into_data();

        self.create_array_from_builder(
            ArrayData::builder(data_type.clone())
                .len(length)
                .add_buffer(buffers[1].clone()) // offsets
                .add_buffer(buffers[2].clone()) // sizes
                .add_child_data(child_data)
                .null_bit_buffer(null_buffer)
                .null_count(field_node.null_count() as usize),
        )
    }

    fn create_struct_array(
        &self,
        struct_node: &FieldNode,
        null_buffer: Buffer,
        struct_fields: &Fields,
        struct_arrays: Vec<ArrayRef>,
    ) -> Result<ArrayRef, ArrowError> {
        let null_count = struct_node.null_count() as usize;
        let len = struct_node.length() as usize;
        let skip_validation = self.skip_validation.get();

        let nulls = if null_count > 0 {
            if !skip_validation {
                // Check before constructing the BooleanBuffer, which would otherwise panic.
                let needed_len = len.div_ceil(8);
                if null_buffer.len() < needed_len {
                    return Err(ArrowError::InvalidArgumentError(format!(
                        "null_bit_buffer size too small. got {} needed {}",
                        null_buffer.len(),
                        needed_len
                    )));
                }
            }
            let validity_buffer = BooleanBuffer::new(null_buffer, 0, len);
            let null_buffer = if skip_validation {
                // safety: flag can only be set via unsafe code
                unsafe { NullBuffer::new_unchecked(validity_buffer, null_count) }
            } else {
                let null_buffer = NullBuffer::new(validity_buffer);

                if null_buffer.null_count() != null_count {
                    return Err(ArrowError::InvalidArgumentError(format!(
                        "null_count value ({}) doesn't match actual number of nulls in array ({})",
                        null_count,
                        null_buffer.null_count()
                    )));
                }

                null_buffer
            };

            Some(null_buffer)
        } else {
            None
        };
        if struct_arrays.is_empty() {
            // `StructArray::from` can't infer the correct row count
            // if we have zero fields
            return Ok(Arc::new(StructArray::new_empty_fields(len, nulls)));
        }

        let struct_array = if skip_validation {
            // safety: flag can only be set via unsafe code
            unsafe { StructArray::new_unchecked(struct_fields.clone(), struct_arrays, nulls) }
        } else {
            StructArray::try_new(struct_fields.clone(), struct_arrays, nulls)?
        };

        Ok(Arc::new(struct_array))
    }

    /// Reads the correct number of buffers based on list type and null_count, and creates a
    /// list array ref
    fn create_dictionary_array(
        &self,
        field_node: &FieldNode,
        data_type: &DataType,
        buffers: &[Buffer],
        value_array: ArrayRef,
    ) -> Result<ArrayRef, ArrowError> {
        if let Dictionary(_, _) = *data_type {
            let null_buffer = (field_node.null_count() > 0).then_some(buffers[0].clone());
            let builder = ArrayData::builder(data_type.clone())
                .len(field_node.length() as usize)
                .add_buffer(buffers[1].clone())
                .add_child_data(value_array.into_data())
                .null_bit_buffer(null_buffer)
                .null_count(field_node.null_count() as usize);
            self.create_array_from_builder(builder)
        } else {
            unreachable!("Cannot create dictionary array from {:?}", data_type)
        }
    }
}

/// State for decoding Arrow arrays from an [IPC RecordBatch] structure to
/// [`RecordBatch`]
///
/// [IPC RecordBatch]: crate::RecordBatch
///
pub struct RecordBatchDecoder<'a> {
    /// The flatbuffers encoded record batch
    batch: crate::RecordBatch<'a>,
    /// The output schema
    schema: SchemaRef,
    /// Decoded dictionaries indexed by dictionary id
    dictionaries_by_id: &'a HashMap<i64, ArrayRef>,
    /// Optional compression codec
    compression: Option<CompressionCodec>,
    /// Decompression context for reusing zstd decompressor state
    decompression_context: DecompressionContext,
    /// The format version
    version: MetadataVersion,
    /// The raw data buffer
    data: &'a Buffer,
    /// The fields comprising this array
    nodes: VectorIter<'a, FieldNode>,
    /// The buffers comprising this array
    buffers: VectorIter<'a, crate::Buffer>,
    /// Projection (subset of columns) to read, if any
    /// See [`RecordBatchDecoder::with_projection`] for details
    projection: Option<&'a [usize]>,
    /// Are buffers required to already be aligned? See
    /// [`RecordBatchDecoder::with_require_alignment`] for details
    require_alignment: bool,
    /// Should validation be skipped when reading data? Defaults to false.
    ///
    /// See [`FileDecoder::with_skip_validation`] for details.
    skip_validation: UnsafeFlag,
}

impl<'a> RecordBatchDecoder<'a> {
    /// Create a reader for decoding arrays from an encoded [`RecordBatch`]
    pub fn try_new(
        buf: &'a Buffer,
        batch: crate::RecordBatch<'a>,
        schema: SchemaRef,
        dictionaries_by_id: &'a HashMap<i64, ArrayRef>,
        metadata: &'a MetadataVersion,
    ) -> Result<Self, ArrowError> {
        let buffers = batch.buffers().ok_or_else(|| {
            ArrowError::IpcError("Unable to get buffers from IPC RecordBatch".to_string())
        })?;
        let field_nodes = batch.nodes().ok_or_else(|| {
            ArrowError::IpcError("Unable to get field nodes from IPC RecordBatch".to_string())
        })?;

        let batch_compression = batch.compression();
        let compression = batch_compression
            .map(|batch_compression| batch_compression.codec().try_into())
            .transpose()?;

        Ok(Self {
            batch,
            schema,
            dictionaries_by_id,
            compression,
            decompression_context: DecompressionContext::new(),
            version: *metadata,
            data: buf,
            nodes: field_nodes.iter(),
            buffers: buffers.iter(),
            projection: None,
            require_alignment: false,
            skip_validation: UnsafeFlag::new(),
        })
    }

    /// Set the projection (default: None)
    ///
    /// If set, the projection is the list  of column indices
    /// that will be read
    pub fn with_projection(mut self, projection: Option<&'a [usize]>) -> Self {
        self.projection = projection;
        self
    }

    /// Set require_alignment (default: false)
    ///
    /// If true, buffers must be aligned appropriately or error will
    /// result. If false, buffers will be copied to aligned buffers
    /// if necessary.
    pub fn with_require_alignment(mut self, require_alignment: bool) -> Self {
        self.require_alignment = require_alignment;
        self
    }

    /// Specifies if validation should be skipped when reading data (defaults to `false`)
    ///
    /// When enabled, the following checks are bypassed:
    /// - Offset bounds (e.g. list/string offsets pointing past the end of their value buffer)
    /// - UTF-8 validity of string columns (`Utf8` / `LargeUtf8`)
    /// - Null count consistency and buffer length checks
    ///
    /// # Undefined behavior
    ///
    /// Relies on the caller only passing a flag with `true` value if they are
    /// certain that the data is valid. Invalid data that bypasses these checks
    /// may cause undefined behavior when the arrays are later accessed.
    pub fn with_skip_validation(mut self, skip_validation: UnsafeFlag) -> Self {
        self.skip_validation = skip_validation;
        self
    }

    /// Read the record batch, consuming the reader
    ///
    /// # Errors
    ///
    /// Returns an error if the message does not describe a batch matching the schema,
    /// for example if it declares more variadic buffer counts than the schema uses.
    pub fn read_record_batch(mut self) -> Result<RecordBatch, ArrowError> {
        let mut variadic_counts: VecDeque<i64> = self
            .batch
            .variadicBufferCounts()
            .into_iter()
            .flatten()
            .collect();

        let options = RecordBatchOptions::new().with_row_count(Some(self.batch.length() as usize));

        let schema = Arc::clone(&self.schema);
        if let Some(projection) = self.projection {
            let mut arrays = Vec::with_capacity(projection.len());
            // project fields
            for (idx, field) in schema.fields().iter().enumerate() {
                // A projected field can appear more than once, so collect all matching positions.
                let mut decoded = None;
                for (proj_idx, projected_idx) in projection.iter().enumerate() {
                    if *projected_idx == idx {
                        // Reuse the decoded array for duplicate projection entries.
                        let child = match decoded.clone() {
                            Some(child) => child,
                            None => {
                                let child = self.create_array(field, &mut variadic_counts)?;
                                decoded = Some(Arc::clone(&child));
                                child
                            }
                        };
                        arrays.push((proj_idx, child));
                    }
                }

                if decoded.is_none() {
                    self.skip_field(field, &mut variadic_counts)?;
                }
            }

            arrays.sort_by_key(|t| t.0);

            let schema = Arc::new(schema.project(projection)?);
            let columns = arrays.into_iter().map(|t| t.1).collect::<Vec<_>>();

            if self.skip_validation.get() {
                // Safety: setting `skip_validation` requires `unsafe`, user assures data is valid
                unsafe {
                    Ok(RecordBatch::new_unchecked(
                        schema,
                        columns,
                        self.batch.length() as usize,
                    ))
                }
            } else {
                check_variadic_counts_consumed(&variadic_counts)?;
                RecordBatch::try_new_with_options(schema, columns, &options)
            }
        } else {
            let mut children = Vec::with_capacity(schema.fields().len());
            // keep track of index as lists require more than one node
            for field in schema.fields() {
                let child = self.create_array(field, &mut variadic_counts)?;
                children.push(child);
            }

            if self.skip_validation.get() {
                // Safety: setting `skip_validation` requires `unsafe`, user assures data is valid
                unsafe {
                    Ok(RecordBatch::new_unchecked(
                        schema,
                        children,
                        self.batch.length() as usize,
                    ))
                }
            } else {
                check_variadic_counts_consumed(&variadic_counts)?;
                RecordBatch::try_new_with_options(schema, children, &options)
            }
        }
    }

    fn next_buffer(&mut self) -> Result<Buffer, ArrowError> {
        let buffer = self.buffers.next().ok_or_else(|| {
            ArrowError::IpcError("Buffer count mismatched with metadata".to_string())
        })?;
        read_buffer(
            buffer,
            self.data,
            self.compression,
            &mut self.decompression_context,
        )
    }

    fn skip_buffer(&mut self) -> Result<(), ArrowError> {
        self.buffers.next().ok_or_else(|| {
            ArrowError::IpcError("Buffer count mismatched with metadata".to_string())
        })?;
        Ok(())
    }

    fn next_node(&mut self, field: &Field) -> Result<&'a FieldNode, ArrowError> {
        self.nodes.next().ok_or_else(|| {
            ArrowError::SchemaError(format!(
                "Invalid data for schema. {field} refers to node not found in schema",
            ))
        })
    }

    fn skip_field(
        &mut self,
        field: &Field,
        variadic_count: &mut VecDeque<i64>,
    ) -> Result<(), ArrowError> {
        self.next_node(field)?;

        match field.data_type() {
            Utf8 | Binary | LargeBinary | LargeUtf8 => {
                for _ in 0..3 {
                    self.skip_buffer()?;
                }
            }
            Utf8View | BinaryView => {
                let count = self.next_variadic_buffer_count(variadic_count, field.data_type())?;
                for _ in 0..count {
                    self.skip_buffer()?;
                }
            }
            FixedSizeBinary(_) => {
                self.skip_buffer()?;
                self.skip_buffer()?;
            }
            List(list_field) | LargeList(list_field) | Map(list_field, _) => {
                self.skip_buffer()?;
                self.skip_buffer()?;
                self.skip_field(list_field, variadic_count)?;
            }
            ListView(list_field) | LargeListView(list_field) => {
                self.skip_buffer()?; // Null buffer
                self.skip_buffer()?; // Offsets
                self.skip_buffer()?; // Sizes
                self.skip_field(list_field, variadic_count)?;
            }
            FixedSizeList(list_field, _) => {
                self.skip_buffer()?;
                self.skip_field(list_field, variadic_count)?;
            }
            Struct(struct_fields) => {
                self.skip_buffer()?;

                // skip for each field
                for struct_field in struct_fields {
                    self.skip_field(struct_field, variadic_count)?
                }
            }
            RunEndEncoded(run_ends_field, values_field) => {
                self.skip_field(run_ends_field, variadic_count)?;
                self.skip_field(values_field, variadic_count)?;
            }
            Dictionary(_, _) => {
                self.skip_buffer()?; // Nulls
                self.skip_buffer()?; // Indices
            }
            Union(fields, mode) => {
                if self.version < MetadataVersion::V5 {
                    self.skip_buffer()?; // Null buffer
                }
                self.skip_buffer()?; // Type ids

                match mode {
                    UnionMode::Dense => self.skip_buffer()?, // Offsets
                    UnionMode::Sparse => {}
                }

                for (_, field) in fields.iter() {
                    self.skip_field(field, variadic_count)?
                }
            }
            // Null has no buffers to skip
            Null => {}

            // Fixed-width and boolean types: skip null buffer + values buffer
            Boolean
            | Int8
            | Int16
            | Int32
            | Int64
            | UInt8
            | UInt16
            | UInt32
            | UInt64
            | Float16
            | Float32
            | Float64
            | Timestamp(_, _)
            | Date32
            | Date64
            | Time32(_)
            | Time64(_)
            | Duration(_)
            | Interval(_)
            | Decimal32(_, _)
            | Decimal64(_, _)
            | Decimal128(_, _)
            | Decimal256(_, _) => {
                self.skip_buffer()?;
                self.skip_buffer()?;
            }
        }
        Ok(())
    }
}

impl RecordBatchDecoder<'_> {
    /// Takes the number of variadic buffers declared for one `BinaryView` or `Utf8View`
    /// column, and returns the total number of buffers to read for it.
    ///
    /// The count comes from the IPC message, so it may be missing, negative, or larger
    /// than the number of buffers the message actually has.
    fn next_variadic_buffer_count(
        &self,
        variadic_counts: &mut VecDeque<i64>,
        data_type: &DataType,
    ) -> Result<usize, ArrowError> {
        let count = variadic_counts.pop_front().ok_or_else(|| {
            ArrowError::IpcError(format!("Missing variadic count for {data_type} column"))
        })?;

        let remaining = self.buffers.len();

        // The view buffer and the null buffer are not counted as variadic.
        usize::try_from(count)
            .ok()
            .and_then(|count| count.checked_add(2))
            .filter(|total| *total <= remaining)
            .ok_or_else(|| {
                ArrowError::IpcError(format!(
                    "Invalid variadic count {count} for {data_type} column, \
                     with {remaining} buffer(s) left in the message"
                ))
            })
    }
}

/// Creates a record batch from binary data using the `crate::RecordBatch` indexes and the `Schema`.
///
/// If `require_alignment` is true, this function will return an error if any array data in the
/// input `buf` is not properly aligned.
/// Under the hood it will use [`arrow_data::ArrayDataBuilder::build`] to construct [`arrow_data::ArrayData`].
///
/// If `require_alignment` is false, this function will automatically allocate a new aligned buffer
/// and copy over the data if any array data in the input `buf` is not properly aligned.
/// (Properly aligned array data will remain zero-copy.)
/// Under the hood it will use [`arrow_data::ArrayDataBuilder::align_buffers`] to construct [`arrow_data::ArrayData`].
pub fn read_record_batch(
    buf: &Buffer,
    batch: crate::RecordBatch,
    schema: SchemaRef,
    dictionaries_by_id: &HashMap<i64, ArrayRef>,
    projection: Option<&[usize]>,
    metadata: &MetadataVersion,
) -> Result<RecordBatch, ArrowError> {
    RecordBatchDecoder::try_new(buf, batch, schema, dictionaries_by_id, metadata)?
        .with_projection(projection)
        .with_require_alignment(false)
        .read_record_batch()
}

/// Read the dictionary from the buffer and provided metadata,
/// updating the `dictionaries_by_id` with the resulting dictionary
pub fn read_dictionary(
    buf: &Buffer,
    batch: crate::DictionaryBatch,
    schema: &Schema,
    dictionaries_by_id: &mut HashMap<i64, ArrayRef>,
    metadata: &MetadataVersion,
) -> Result<(), ArrowError> {
    read_dictionary_impl(
        buf,
        batch,
        schema,
        dictionaries_by_id,
        metadata,
        false,
        UnsafeFlag::new(),
    )
}

/// Low-level version of [`read_dictionary`] with alignment and validation controls
pub fn read_dictionary_impl(
    buf: &Buffer,
    batch: crate::DictionaryBatch,
    schema: &Schema,
    dictionaries_by_id: &mut HashMap<i64, ArrayRef>,
    metadata: &MetadataVersion,
    require_alignment: bool,
    skip_validation: UnsafeFlag,
) -> Result<(), ArrowError> {
    let id = batch.id();

    let dictionary_values = get_dictionary_values(
        buf,
        batch,
        schema,
        dictionaries_by_id,
        metadata,
        require_alignment,
        skip_validation,
    )?;

    update_dictionaries(dictionaries_by_id, batch.isDelta(), id, dictionary_values)?;

    Ok(())
}

/// Updates the `dictionaries_by_id` with the provided dictionary values and id.
///
/// # Errors
/// - If `is_delta` is true and there is no existing dictionary for the given
///   `dict_id`
/// - If `is_delta` is true and the concatenation of the existing and new
///   dictionary fails. This usually signals a type mismatch between the old and
///   new values.
fn update_dictionaries(
    dictionaries_by_id: &mut HashMap<i64, ArrayRef>,
    is_delta: bool,
    dict_id: i64,
    dict_values: ArrayRef,
) -> Result<(), ArrowError> {
    if !is_delta {
        // We don't currently record the isOrdered field. This could be general
        // attributes of arrays.
        // Add (possibly multiple) array refs to the dictionaries array.
        dictionaries_by_id.insert(dict_id, dict_values.clone());
        return Ok(());
    }

    let existing = dictionaries_by_id.get(&dict_id).ok_or_else(|| {
        ArrowError::InvalidArgumentError(format!(
            "No existing dictionary for delta dictionary with id '{dict_id}'"
        ))
    })?;

    let combined = concat::concat(&[existing, &dict_values]).map_err(|e| {
        ArrowError::InvalidArgumentError(format!("Failed to concat delta dictionary: {e}"))
    })?;

    dictionaries_by_id.insert(dict_id, combined);

    Ok(())
}

/// Returns the first field with dictionary `id`, in [`Schema::fields_with_dict_id`] order.
///
/// Unlike that method this stops at the first match without collecting every nested field,
/// which is significant when a stream has many dictionary batches.
fn first_field_with_dict_id(fields: &Fields, id: i64) -> Option<&Field> {
    fields
        .iter()
        .find_map(|field| field_with_dict_id(field, id))
}

fn field_with_dict_id(field: &Field, id: i64) -> Option<&Field> {
    #[expect(deprecated)]
    let dict_id = field.dict_id();
    if dict_id == Some(id) {
        return Some(field);
    }
    child_field_with_dict_id(field.data_type(), id)
}

fn child_field_with_dict_id(data_type: &DataType, id: i64) -> Option<&Field> {
    match data_type {
        DataType::Struct(fields) => first_field_with_dict_id(fields, id),
        DataType::Union(fields, _) => fields.iter().find_map(|(_, f)| field_with_dict_id(f, id)),
        DataType::List(field)
        | DataType::LargeList(field)
        | DataType::ListView(field)
        | DataType::LargeListView(field)
        | DataType::FixedSizeList(field, _)
        | DataType::Map(field, _)
        | DataType::RunEndEncoded(_, field) => field_with_dict_id(field, id),
        DataType::Dictionary(_, value_type) => child_field_with_dict_id(value_type, id),
        _ => None,
    }
}

/// Given a dictionary batch IPC message/body along with the full state of a
/// stream including schema, dictionary cache, metadata, and other flags, this
/// function will parse the buffer into an array of dictionary values.
fn get_dictionary_values(
    buf: &Buffer,
    batch: crate::DictionaryBatch,
    schema: &Schema,
    dictionaries_by_id: &HashMap<i64, ArrayRef>,
    metadata: &MetadataVersion,
    require_alignment: bool,
    skip_validation: UnsafeFlag,
) -> Result<ArrayRef, ArrowError> {
    let id = batch.id();
    let first_field = first_field_with_dict_id(schema.fields(), id).ok_or_else(|| {
        ArrowError::InvalidArgumentError(format!("dictionary id {id} not found in schema"))
    })?;

    // As the dictionary batch does not contain the type of the
    // values array, we need to retrieve this from the schema.
    // Get an array representing this dictionary's values.
    let dictionary_values: ArrayRef = match first_field.data_type() {
        DataType::Dictionary(_, value_type) => {
            // Make a fake schema for the dictionary batch.
            let value = value_type.as_ref().clone();
            let schema = Schema::new(vec![Field::new("", value, true)]);
            // Read a single column
            let Some(data) = batch.data() else {
                return Err(ArrowError::ParseError(
                    "Dictionary batch is missing its data".to_string(),
                ));
            };
            let record_batch = RecordBatchDecoder::try_new(
                buf,
                data,
                Arc::new(schema),
                dictionaries_by_id,
                metadata,
            )?
            .with_require_alignment(require_alignment)
            .with_skip_validation(skip_validation)
            .read_record_batch()?;

            Some(record_batch.column(0).clone())
        }
        _ => None,
    }
    .ok_or_else(|| {
        ArrowError::InvalidArgumentError(format!("dictionary id {id} not found in schema"))
    })?;

    Ok(dictionary_values)
}

/// Read the data for a given block
fn read_block<R: Read + Seek>(mut reader: R, block: &Block) -> Result<Buffer, ArrowError> {
    reader.seek(SeekFrom::Start(block.offset() as u64))?;
    let body_len = block.bodyLength().to_usize().unwrap();
    let metadata_len = block.metaDataLength().to_usize().unwrap();
    let total_len = body_len.checked_add(metadata_len).unwrap();

    let mut buf = MutableBuffer::try_from_len_zeroed(total_len)
        .map_err(|e| ArrowError::MemoryError(e.to_string()))?;
    reader.read_exact(&mut buf)?;
    Ok(buf.into())
}

/// One variadic buffer count is consumed per `BinaryView` or `Utf8View` column in
/// the schema, so any count left over means the message and the schema disagree.
///
/// The opposite case, too few counts, is reported by [`RecordBatchDecoder::create_array`].
fn check_variadic_counts_consumed(variadic_counts: &VecDeque<i64>) -> Result<(), ArrowError> {
    if variadic_counts.is_empty() {
        Ok(())
    } else {
        Err(ArrowError::IpcError(format!(
            "Mismatch between schema and data: the IPC message declares {} more variadic \
             buffer count(s) than the schema has BinaryView or Utf8View columns",
            variadic_counts.len()
        )))
    }
}

/// Parse an encapsulated message
///
/// <https://arrow.apache.org/docs/format/Columnar.html#encapsulated-message-format>
fn parse_message(buf: &[u8]) -> Result<Message::Message<'_>, ArrowError> {
    let buf = match buf[..4] == CONTINUATION_MARKER {
        true => &buf[8..],
        false => &buf[4..],
    };
    crate::root_as_message(buf)
        .map_err(|err| ArrowError::ParseError(format!("Unable to get root as message: {err:?}")))
}

/// Read the footer length from the last 10 bytes of an Arrow IPC file
///
/// Expects a 4 byte footer length followed by `b"ARROW1"`
pub fn read_footer_length(buf: [u8; 10]) -> Result<usize, ArrowError> {
    if buf[4..] != super::ARROW_MAGIC {
        return Err(ArrowError::ParseError(
            "Arrow file does not contain correct footer".to_string(),
        ));
    }

    // read footer length
    let footer_len = i32::from_le_bytes([buf[0], buf[1], buf[2], buf[3]]);
    footer_len
        .try_into()
        .map_err(|_| ArrowError::ParseError(format!("Invalid footer length: {footer_len}")))
}

/// A low-level, push-based interface for reading an IPC file
///
/// For a higher-level interface see [`FileReader`]
///
/// For an example of using this API with `mmap` see the [`zero_copy_ipc`] example.
///
/// [`zero_copy_ipc`]: https://github.com/apache/arrow-rs/blob/main/arrow/examples/zero_copy_ipc.rs
///
/// ```
/// # use std::sync::Arc;
/// # use arrow_array::*;
/// # use arrow_array::types::Int32Type;
/// # use arrow_buffer::Buffer;
/// # use arrow_ipc::convert::try_fb_to_schema;
/// # use arrow_ipc::reader::{FileDecoder, read_footer_length};
/// # use arrow_ipc::root_as_footer;
/// # use arrow_ipc::writer::FileWriter;
/// // Write an IPC file
///
/// let batch = RecordBatch::try_from_iter([
///     ("a", Arc::new(Int32Array::from(vec![1, 2, 3])) as _),
///     ("b", Arc::new(Int32Array::from(vec![1, 2, 3])) as _),
///     ("c", Arc::new(DictionaryArray::<Int32Type>::from_iter(["hello", "hello", "world"])) as _),
/// ]).unwrap();
///
/// let schema = batch.schema();
///
/// let mut out = Vec::with_capacity(1024);
/// let mut writer = FileWriter::try_new(&mut out, schema.as_ref()).unwrap();
/// writer.write(&batch).unwrap();
/// writer.finish().unwrap();
///
/// drop(writer);
///
/// // Read IPC file
///
/// let buffer = Buffer::from_vec(out);
/// let trailer_start = buffer.len() - 10;
/// let footer_len = read_footer_length(buffer[trailer_start..].try_into().unwrap()).unwrap();
/// let footer = root_as_footer(&buffer[trailer_start - footer_len..trailer_start]).unwrap();
///
/// let back = try_fb_to_schema(footer.schema().unwrap()).unwrap();
/// assert_eq!(&back, schema.as_ref());
///
/// let mut decoder = FileDecoder::new(schema, footer.version());
///
/// // Read dictionaries
/// for block in footer.dictionaries().iter().flatten() {
///     let block_len = block.bodyLength() as usize + block.metaDataLength() as usize;
///     let data = buffer.slice_with_length(block.offset() as _, block_len);
///     decoder.read_dictionary(&block, &data).unwrap();
/// }
///
/// // Read record batch
/// let batches = footer.recordBatches().unwrap();
/// assert_eq!(batches.len(), 1); // Only wrote a single batch
///
/// let block = batches.get(0);
/// let block_len = block.bodyLength() as usize + block.metaDataLength() as usize;
/// let data = buffer.slice_with_length(block.offset() as _, block_len);
/// let back = decoder.read_record_batch(block, &data).unwrap().unwrap();
///
/// assert_eq!(batch, back);
/// ```
#[derive(Debug)]
pub struct FileDecoder {
    schema: SchemaRef,
    dictionaries: HashMap<i64, ArrayRef>,
    version: MetadataVersion,
    projection: Option<Vec<usize>>,
    require_alignment: bool,
    skip_validation: UnsafeFlag,
}

impl FileDecoder {
    /// Create a new [`FileDecoder`] with the given schema and version
    pub fn new(schema: SchemaRef, version: MetadataVersion) -> Self {
        Self {
            schema,
            version,
            dictionaries: Default::default(),
            projection: None,
            require_alignment: false,
            skip_validation: UnsafeFlag::new(),
        }
    }

    /// Specify a projection
    pub fn with_projection(mut self, projection: Vec<usize>) -> Self {
        self.projection = Some(projection);
        self
    }

    /// Specifies if the array data in input buffers is required to be properly aligned.
    ///
    /// If `require_alignment` is true, this decoder will return an error if any array data in the
    /// input `buf` is not properly aligned.
    /// Under the hood it will use [`arrow_data::ArrayDataBuilder::build`] to construct
    /// [`arrow_data::ArrayData`].
    ///
    /// If `require_alignment` is false (the default), this decoder will automatically allocate a
    /// new aligned buffer and copy over the data if any array data in the input `buf` is not
    /// properly aligned. (Properly aligned array data will remain zero-copy.)
    /// Under the hood it will use [`arrow_data::ArrayDataBuilder::align_buffers`] to construct
    /// [`arrow_data::ArrayData`].
    pub fn with_require_alignment(mut self, require_alignment: bool) -> Self {
        self.require_alignment = require_alignment;
        self
    }

    /// Specifies if validation should be skipped when reading data (defaults to `false`)
    ///
    /// # Safety
    ///
    /// This flag must only be set to `true` when you trust the input data and are sure the data you are
    /// reading is a valid Arrow IPC file, otherwise undefined behavior may
    /// result.
    ///
    /// For example, some programs may wish to trust reading IPC files written
    /// by the same process that created the files.
    pub unsafe fn with_skip_validation(mut self, skip_validation: bool) -> Self {
        unsafe { self.skip_validation.set(skip_validation) };
        self
    }

    fn read_message<'a>(&self, buf: &'a [u8]) -> Result<Message::Message<'a>, ArrowError> {
        let message = parse_message(buf)?;

        // some old test data's footer metadata is not set, so we account for that
        if self.version != MetadataVersion::V1 && message.version() != self.version {
            return Err(ArrowError::IpcError(
                "Could not read IPC message as metadata versions mismatch".to_string(),
            ));
        }
        Ok(message)
    }

    /// Read the dictionary with the given block and data buffer
    pub fn read_dictionary(&mut self, block: &Block, buf: &Buffer) -> Result<(), ArrowError> {
        let message = self.read_message(buf)?;
        match message.header_type() {
            crate::MessageHeader::DictionaryBatch => {
                let batch = message.header_as_dictionary_batch().unwrap();
                read_dictionary_impl(
                    &buf.slice(block.metaDataLength() as _),
                    batch,
                    &self.schema,
                    &mut self.dictionaries,
                    &message.version(),
                    self.require_alignment,
                    self.skip_validation.clone(),
                )
            }
            t => Err(ArrowError::ParseError(format!(
                "Expecting DictionaryBatch in dictionary blocks, found {t:?}."
            ))),
        }
    }

    /// Read the RecordBatch with the given block and data buffer
    pub fn read_record_batch(
        &self,
        block: &Block,
        buf: &Buffer,
    ) -> Result<Option<RecordBatch>, ArrowError> {
        let message = self.read_message(buf)?;
        match message.header_type() {
            crate::MessageHeader::Schema => Err(ArrowError::IpcError(
                "Not expecting a schema when messages are read".to_string(),
            )),
            crate::MessageHeader::RecordBatch => {
                let batch = message.header_as_record_batch().ok_or_else(|| {
                    ArrowError::IpcError("Unable to read IPC message as record batch".to_string())
                })?;
                // read the block that makes up the record batch into a buffer
                RecordBatchDecoder::try_new(
                    &buf.slice(block.metaDataLength() as _),
                    batch,
                    self.schema.clone(),
                    &self.dictionaries,
                    &message.version(),
                )?
                .with_projection(self.projection.as_deref())
                .with_require_alignment(self.require_alignment)
                .with_skip_validation(self.skip_validation.clone())
                .read_record_batch()
                .map(Some)
            }
            crate::MessageHeader::NONE => Ok(None),
            t => Err(ArrowError::InvalidArgumentError(format!(
                "Reading types other than record batches not yet supported, unable to read {t:?}"
            ))),
        }
    }
}

/// Build an Arrow [`FileReader`] with custom options.
#[derive(Debug)]
pub struct FileReaderBuilder {
    /// Optional projection for which columns to load (zero-based column indices)
    projection: Option<Vec<usize>>,
    /// Passed through to construct [`VerifierOptions`]
    max_footer_fb_tables: usize,
    /// Passed through to construct [`VerifierOptions`]
    max_footer_fb_depth: usize,
}

impl Default for FileReaderBuilder {
    fn default() -> Self {
        let verifier_options = VerifierOptions::default();
        Self {
            max_footer_fb_tables: verifier_options.max_tables,
            max_footer_fb_depth: verifier_options.max_depth,
            projection: None,
        }
    }
}

impl FileReaderBuilder {
    /// Options for creating a new [`FileReader`].
    ///
    /// To convert a builder into a reader, call [`FileReaderBuilder::build`].
    pub fn new() -> Self {
        Self::default()
    }

    /// Optional projection for which columns to load (zero-based column indices).
    pub fn with_projection(mut self, projection: Vec<usize>) -> Self {
        self.projection = Some(projection);
        self
    }

    /// Flatbuffers option for parsing the footer. Controls the max number of fields and
    /// metadata key-value pairs that can be parsed from the schema of the footer.
    ///
    /// By default this is set to `1_000_000` which roughly translates to a schema with
    /// no metadata key-value pairs but 499,999 fields.
    ///
    /// This default limit is enforced to protect against malicious files with a massive
    /// amount of flatbuffer tables which could cause a denial of service attack.
    ///
    /// If you need to ingest a trusted file with a massive number of fields and/or
    /// metadata key-value pairs and are facing the error `"Unable to get root as
    /// footer: TooManyTables"` then increase this parameter as necessary.
    pub fn with_max_footer_fb_tables(mut self, max_footer_fb_tables: usize) -> Self {
        self.max_footer_fb_tables = max_footer_fb_tables;
        self
    }

    /// Flatbuffers option for parsing the footer. Controls the max depth for schemas with
    /// nested fields parsed from the footer.
    ///
    /// By default this is set to `64` which roughly translates to a schema with
    /// a field nested 60 levels down through other struct fields.
    ///
    /// This default limit is enforced to protect against malicious files with a extremely
    /// deep flatbuffer structure which could cause a denial of service attack.
    ///
    /// If you need to ingest a trusted file with a deeply nested field and are facing the
    /// error `"Unable to get root as footer: DepthLimitReached"` then increase this
    /// parameter as necessary.
    pub fn with_max_footer_fb_depth(mut self, max_footer_fb_depth: usize) -> Self {
        self.max_footer_fb_depth = max_footer_fb_depth;
        self
    }

    /// Build [`FileReader`] with given reader.
    pub fn build<R: Read + Seek>(self, mut reader: R) -> Result<FileReader<R>, ArrowError> {
        // Space for ARROW_MAGIC (6 bytes) and length (4 bytes)
        let mut buffer = [0; 10];
        reader.seek(SeekFrom::End(-10))?;
        reader.read_exact(&mut buffer)?;

        let footer_len = read_footer_length(buffer)?;

        // read footer
        let mut footer_data = vec![0; footer_len];
        reader.seek(SeekFrom::End(-10 - footer_len as i64))?;
        reader.read_exact(&mut footer_data)?;

        let verifier_options = VerifierOptions {
            max_tables: self.max_footer_fb_tables,
            max_depth: self.max_footer_fb_depth,
            ..Default::default()
        };
        let footer = crate::root_as_footer_with_opts(&verifier_options, &footer_data[..]).map_err(
            |err| ArrowError::ParseError(format!("Unable to get root as footer: {err:?}")),
        )?;

        let blocks = footer.recordBatches().ok_or_else(|| {
            ArrowError::ParseError("Unable to get record batches from IPC Footer".to_string())
        })?;

        let total_blocks = blocks.len();

        let ipc_schema = footer.schema().ok_or_else(|| {
            ArrowError::ParseError("Unable to get schema from IPC Footer".to_string())
        })?;
        if !ipc_schema.endianness().equals_to_target_endianness() {
            return Err(ArrowError::IpcError(
                "the endianness of the source system does not match the endianness of the target system.".to_owned()
            ));
        }

        let schema = Arc::new(crate::convert::try_fb_to_schema(ipc_schema)?);

        let projected_schema = match &self.projection {
            Some(projection) => Arc::new(schema.project(projection)?),
            None => schema.clone(),
        };

        let mut custom_metadata = HashMap::new();
        if let Some(fb_custom_metadata) = footer.custom_metadata() {
            for kv in fb_custom_metadata {
                let (Some(key), Some(value)) = (kv.key(), kv.value()) else {
                    return Err(ArrowError::ParseError(
                        "Custom metadata in the IPC footer is missing a key or a value".to_string(),
                    ));
                };
                custom_metadata.insert(key.to_string(), value.to_string());
            }
        }

        let mut decoder = FileDecoder::new(schema, footer.version());
        if let Some(projection) = self.projection {
            decoder = decoder.with_projection(projection)
        }

        // Create an array of optional dictionary value arrays, one per field.
        if let Some(dictionaries) = footer.dictionaries() {
            for block in dictionaries {
                let buf = read_block(&mut reader, block)?;
                decoder.read_dictionary(block, &buf)?;
            }
        }

        Ok(FileReader {
            reader,
            blocks: blocks.iter().copied().collect(),
            current_block: 0,
            total_blocks,
            decoder,
            schema: projected_schema,
            custom_metadata,
        })
    }
}

/// Arrow File Reader
///
/// Reads Arrow [`RecordBatch`]es from bytes in the [IPC File Format],
/// providing random access to the record batches.
///
/// # See Also
///
/// * [`Self::set_index`] for random access
/// * [`StreamReader`] for reading streaming data
///
/// # Example: Reading from a `File`
/// ```
/// # use std::io::Cursor;
/// use arrow_array::record_batch;
/// # use arrow_ipc::reader::FileReader;
/// # use arrow_ipc::writer::FileWriter;
/// # let batch = record_batch!(("a", Int32, [1, 2, 3])).unwrap();
/// # let mut file = vec![]; // mimic a stream for the example
/// # {
/// #  let mut writer = FileWriter::try_new(&mut file, &batch.schema()).unwrap();
/// #  writer.write(&batch).unwrap();
/// #  writer.write(&batch).unwrap();
/// #  writer.finish().unwrap();
/// # }
/// # let mut file = Cursor::new(&file);
/// let projection = None; // read all columns
/// let mut reader = FileReader::try_new(&mut file, projection).unwrap();
/// // Position the reader to the second batch
/// reader.set_index(1).unwrap();
/// // read batches from the reader using the Iterator trait
/// let mut num_rows = 0;
/// for batch in reader {
///    let batch = batch.unwrap();
///    num_rows += batch.num_rows();
/// }
/// assert_eq!(num_rows, 3);
/// ```
/// # Example: Reading from `mmap`ed file
///
/// For an example creating Arrays without copying using  memory mapped (`mmap`)
/// files see the [`zero_copy_ipc`] example.
///
/// [IPC File Format]: https://arrow.apache.org/docs/format/Columnar.html#ipc-file-format
/// [`zero_copy_ipc`]: https://github.com/apache/arrow-rs/blob/main/arrow/examples/zero_copy_ipc.rs
pub struct FileReader<R> {
    /// File reader that supports reading and seeking
    reader: R,

    /// The decoder
    decoder: FileDecoder,

    /// Schema of the record batches produced by this reader
    schema: SchemaRef,

    /// The blocks in the file
    ///
    /// A block indicates the regions in the file to read to get data
    blocks: Vec<Block>,

    /// A counter to keep track of the current block that should be read
    current_block: usize,

    /// The total number of blocks, which may contain record batches and other types
    total_blocks: usize,

    /// User defined metadata
    custom_metadata: HashMap<String, String>,
}

impl<R> fmt::Debug for FileReader<R> {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> Result<(), fmt::Error> {
        f.debug_struct("FileReader<R>")
            .field("decoder", &self.decoder)
            .field("blocks", &self.blocks)
            .field("current_block", &self.current_block)
            .field("total_blocks", &self.total_blocks)
            .finish_non_exhaustive()
    }
}

impl<R: Read + Seek> FileReader<BufReader<R>> {
    /// Try to create a new file reader with the reader wrapped in a BufReader.
    ///
    /// See [`FileReader::try_new`] for an unbuffered version.
    pub fn try_new_buffered(reader: R, projection: Option<Vec<usize>>) -> Result<Self, ArrowError> {
        Self::try_new(BufReader::new(reader), projection)
    }
}

impl<R: Read + Seek> FileReader<R> {
    /// Try to create a new file reader.
    ///
    /// There is no internal buffering. If buffered reads are needed you likely want to use
    /// [`FileReader::try_new_buffered`] instead.
    ///
    /// # Errors
    ///
    /// An [`Err`] may be returned if:
    /// - the file does not meet the Arrow Format footer requirements,
    /// - file endianness does not match the target endianness, or
    /// - the projection contains an index outside the file schema.
    pub fn try_new(reader: R, projection: Option<Vec<usize>>) -> Result<Self, ArrowError> {
        let builder = FileReaderBuilder {
            projection,
            ..Default::default()
        };
        builder.build(reader)
    }

    /// Return user defined customized metadata
    pub fn custom_metadata(&self) -> &HashMap<String, String> {
        &self.custom_metadata
    }

    /// Return the number of batches in the file
    pub fn num_batches(&self) -> usize {
        self.total_blocks
    }

    /// Return the schema of the record batches produced by this reader
    pub fn schema(&self) -> SchemaRef {
        self.schema.clone()
    }

    /// See to a specific [`RecordBatch`]
    ///
    /// Sets the current block to the index, allowing random reads
    pub fn set_index(&mut self, index: usize) -> Result<(), ArrowError> {
        if index >= self.total_blocks {
            Err(ArrowError::InvalidArgumentError(format!(
                "Cannot set batch to index {} from {} total batches",
                index, self.total_blocks
            )))
        } else {
            self.current_block = index;
            Ok(())
        }
    }

    fn maybe_next(&mut self) -> Result<Option<RecordBatch>, ArrowError> {
        let block = &self.blocks[self.current_block];
        self.current_block += 1;

        // read length
        let buffer = read_block(&mut self.reader, block)?;
        self.decoder.read_record_batch(block, &buffer)
    }

    /// Gets a reference to the underlying reader.
    ///
    /// It is inadvisable to directly read from the underlying reader.
    pub fn get_ref(&self) -> &R {
        &self.reader
    }

    /// Gets a mutable reference to the underlying reader.
    ///
    /// It is inadvisable to directly read from the underlying reader.
    pub fn get_mut(&mut self) -> &mut R {
        &mut self.reader
    }

    /// Specifies if validation should be skipped when reading data (defaults to `false`)
    ///
    /// # Safety
    ///
    /// See [`FileDecoder::with_skip_validation`]
    pub unsafe fn with_skip_validation(mut self, skip_validation: bool) -> Self {
        self.decoder = unsafe { self.decoder.with_skip_validation(skip_validation) };
        self
    }
}

impl<R: Read + Seek> Iterator for FileReader<R> {
    type Item = Result<RecordBatch, ArrowError>;

    fn next(&mut self) -> Option<Self::Item> {
        // get current block
        if self.current_block < self.total_blocks {
            self.maybe_next().transpose()
        } else {
            None
        }
    }
}

impl<R: Read + Seek> RecordBatchReader for FileReader<R> {
    fn schema(&self) -> SchemaRef {
        self.schema()
    }
}

/// Arrow Stream Reader
///
/// Reads Arrow [`RecordBatch`]es from bytes in the [IPC Streaming Format].
///
/// # See Also
///
/// * [`FileReader`] for random access.
///
/// # Example
/// ```
/// # use arrow_array::record_batch;
/// # use arrow_ipc::reader::StreamReader;
/// # use arrow_ipc::writer::StreamWriter;
/// # let batch = record_batch!(("a", Int32, [1, 2, 3])).unwrap();
/// # let mut stream = vec![]; // mimic a stream for the example
/// # {
/// #  let mut writer = StreamWriter::try_new(&mut stream, &batch.schema()).unwrap();
/// #  writer.write(&batch).unwrap();
/// #  writer.finish().unwrap();
/// # }
/// # let stream = stream.as_slice();
/// let projection = None; // read all columns
/// let mut reader = StreamReader::try_new(stream, projection).unwrap();
/// // read batches from the reader using the Iterator trait
/// let mut num_rows = 0;
/// for batch in reader {
///    let batch = batch.unwrap();
///    num_rows += batch.num_rows();
/// }
/// assert_eq!(num_rows, 3);
/// ```
///
/// [IPC Streaming Format]: https://arrow.apache.org/docs/format/Columnar.html#ipc-streaming-format
pub struct StreamReader<R> {
    /// Stream reader
    reader: MessageReader<R>,

    /// The schema that is read from the stream's first message
    schema: SchemaRef,

    /// Optional dictionaries for each schema field.
    ///
    /// Dictionaries may be appended to in the streaming format.
    dictionaries_by_id: HashMap<i64, ArrayRef>,

    /// An indicator of whether the stream is complete.
    ///
    /// This value is set to `true` the first time the reader's `next()` returns `None`.
    finished: bool,

    /// Optional projection: column indices and the resulting projected schema
    projection: Option<(Vec<usize>, SchemaRef)>,

    /// Should validation be skipped when reading data? Defaults to false.
    ///
    /// See [`FileDecoder::with_skip_validation`] for details.
    skip_validation: UnsafeFlag,
}

impl<R> fmt::Debug for StreamReader<R> {
    fn fmt(&self, f: &mut fmt::Formatter<'_>) -> std::result::Result<(), fmt::Error> {
        f.debug_struct("StreamReader<R>")
            .field("reader", &"R")
            .field("schema", &self.schema)
            .field("dictionaries_by_id", &self.dictionaries_by_id)
            .field("finished", &self.finished)
            .field("projection", &self.projection)
            .finish()
    }
}

impl<R: Read> StreamReader<BufReader<R>> {
    /// Try to create a new stream reader with the reader wrapped in a BufReader.
    ///
    /// See [`StreamReader::try_new`] for an unbuffered version.
    pub fn try_new_buffered(reader: R, projection: Option<Vec<usize>>) -> Result<Self, ArrowError> {
        Self::try_new(BufReader::new(reader), projection)
    }
}

impl<R: Read> StreamReader<R> {
    /// Try to create a new stream reader.
    ///
    /// To check if the reader is done, use [`is_finished(self)`](StreamReader::is_finished).
    ///
    /// There is no internal buffering. If buffered reads are needed you likely want to use
    /// [`StreamReader::try_new_buffered`] instead.
    ///
    /// # Errors
    ///
    /// An [`Err`] may be returned if the reader does not encounter a schema
    /// as the first message in the stream.
    pub fn try_new(
        reader: R,
        projection: Option<Vec<usize>>,
    ) -> Result<StreamReader<R>, ArrowError> {
        let mut msg_reader = MessageReader::new(reader);
        let message = msg_reader.maybe_next()?;
        let Some((message, _)) = message else {
            return Err(ArrowError::IpcError(
                "Expected schema message, found empty stream.".to_string(),
            ));
        };

        if message.header_type() != Message::MessageHeader::Schema {
            return Err(ArrowError::IpcError(format!(
                "Expected a schema as the first message in the stream, got: {:?}",
                message.header_type()
            )));
        }

        let schema = message.header_as_schema().ok_or_else(|| {
            ArrowError::ParseError("Failed to parse schema from message header".to_string())
        })?;
        let schema = crate::convert::try_fb_to_schema(schema)?;

        // Create an array of optional dictionary value arrays, one per field.
        let dictionaries_by_id = HashMap::new();

        let projection = match projection {
            Some(projection_indices) => {
                let schema = Arc::new(schema.project(&projection_indices)?);
                Some((projection_indices, schema))
            }
            _ => None,
        };

        Ok(Self {
            reader: msg_reader,
            schema: Arc::new(schema),
            finished: false,
            dictionaries_by_id,
            projection,
            skip_validation: UnsafeFlag::new(),
        })
    }

    /// Return the schema of the record batches produced by this reader
    pub fn schema(&self) -> SchemaRef {
        match &self.projection {
            Some((_, projected_schema)) => projected_schema.clone(),
            None => self.schema.clone(),
        }
    }

    /// Check if the stream is finished
    pub fn is_finished(&self) -> bool {
        self.finished
    }

    fn maybe_next(&mut self) -> Result<Option<RecordBatch>, ArrowError> {
        if self.finished {
            return Ok(None);
        }

        // Read messages until we get a record batch or end of stream
        loop {
            let message = self.next_ipc_message()?;
            let Some(message) = message else {
                // If the message is None, we have reached the end of the stream.
                self.finished = true;
                return Ok(None);
            };

            match message {
                IpcMessage::Schema(_) => {
                    return Err(ArrowError::IpcError(
                        "Expected a record batch, but found a schema".to_string(),
                    ));
                }
                IpcMessage::RecordBatch(record_batch) => {
                    return Ok(Some(record_batch));
                }
                IpcMessage::DictionaryBatch { .. } => {}
            }
        }
    }

    /// Reads and fully parses the next IPC message from the stream. Whereas
    /// [`Self::maybe_next`] is a higher level method focused on reading
    /// `RecordBatch`es, this method returns the individual fully parsed IPC
    /// messages from the underlying stream.
    ///
    /// This is useful primarily for testing reader/writer behaviors as it
    /// allows a full view into the messages that have been written to a stream.
    pub(crate) fn next_ipc_message(&mut self) -> Result<Option<IpcMessage>, ArrowError> {
        let message = self.reader.maybe_next()?;
        let Some((message, body)) = message else {
            // If the message is None, we have reached the end of the stream.
            return Ok(None);
        };

        let ipc_message = match message.header_type() {
            Message::MessageHeader::Schema => {
                let schema = message.header_as_schema().ok_or_else(|| {
                    ArrowError::ParseError("Failed to parse schema from message header".to_string())
                })?;
                let arrow_schema = crate::convert::try_fb_to_schema(schema)?;
                IpcMessage::Schema(arrow_schema)
            }
            Message::MessageHeader::RecordBatch => {
                let batch = message.header_as_record_batch().ok_or_else(|| {
                    ArrowError::IpcError("Unable to read IPC message as record batch".to_string())
                })?;

                let version = message.version();
                let schema = self.schema.clone();
                let record_batch = RecordBatchDecoder::try_new(
                    &body.into(),
                    batch,
                    schema,
                    &self.dictionaries_by_id,
                    &version,
                )?
                .with_projection(self.projection.as_ref().map(|x| x.0.as_ref()))
                .with_require_alignment(false)
                .with_skip_validation(self.skip_validation.clone())
                .read_record_batch()?;
                IpcMessage::RecordBatch(record_batch)
            }
            Message::MessageHeader::DictionaryBatch => {
                let dict = message.header_as_dictionary_batch().ok_or_else(|| {
                    ArrowError::ParseError(
                        "Failed to parse dictionary batch from message header".to_string(),
                    )
                })?;

                let version = message.version();
                let dict_values = get_dictionary_values(
                    &body.into(),
                    dict,
                    &self.schema,
                    &self.dictionaries_by_id,
                    &version,
                    false,
                    self.skip_validation.clone(),
                )?;

                update_dictionaries(
                    &mut self.dictionaries_by_id,
                    dict.isDelta(),
                    dict.id(),
                    dict_values.clone(),
                )?;

                IpcMessage::DictionaryBatch {
                    id: dict.id(),
                    is_delta: (dict.isDelta()),
                    values: (dict_values),
                }
            }
            x => {
                return Err(ArrowError::ParseError(format!(
                    "Unsupported message header type in IPC stream: '{x:?}'"
                )));
            }
        };

        Ok(Some(ipc_message))
    }

    /// Gets a reference to the underlying reader.
    ///
    /// It is inadvisable to directly read from the underlying reader.
    pub fn get_ref(&self) -> &R {
        self.reader.inner()
    }

    /// Gets a mutable reference to the underlying reader.
    ///
    /// It is inadvisable to directly read from the underlying reader.
    pub fn get_mut(&mut self) -> &mut R {
        self.reader.inner_mut()
    }

    /// Specifies if validation should be skipped when reading data (defaults to `false`)
    ///
    /// # Safety
    ///
    /// See [`FileDecoder::with_skip_validation`]
    pub unsafe fn with_skip_validation(mut self, skip_validation: bool) -> Self {
        unsafe { self.skip_validation.set(skip_validation) };
        self
    }
}

impl<R: Read> Iterator for StreamReader<R> {
    type Item = Result<RecordBatch, ArrowError>;

    fn next(&mut self) -> Option<Self::Item> {
        self.maybe_next().transpose()
    }
}

impl<R: Read> RecordBatchReader for StreamReader<R> {
    fn schema(&self) -> SchemaRef {
        self.schema()
    }
}

/// Representation of a fully parsed IpcMessage from the underlying stream.
/// Parsing this kind of message is done by higher level constructs such as
/// [`StreamReader`], because fully interpreting the messages into a record
/// batch or dictionary batch requires access to stream state such as schema
/// and the full dictionary cache.
#[derive(Debug)]
#[expect(dead_code)]
pub(crate) enum IpcMessage {
    Schema(arrow_schema::Schema),
    RecordBatch(RecordBatch),
    DictionaryBatch {
        id: i64,
        is_delta: bool,
        values: ArrayRef,
    },
}

/// Upper bound on how much memory a single message length is allowed to reserve before any
/// of the bytes it promises have been read.
///
/// Message lengths come from the stream itself, so they cannot be trusted: a truncated or
/// corrupted stream can declare a body of arbitrary size. Reserving that up front turns a
/// malformed input into an allocation failure, which aborts the process rather than
/// returning an error the caller can handle. Beyond this size the buffer grows as the data
/// arrives instead, so an implausible length costs one bounded allocation and then fails as
/// a short read.
///
/// The value trades the size of that bounded allocation against how large a body still gets
/// read in a single allocation: bodies up to this size behave exactly as before, larger ones
/// start here and the buffer doubles as the data arrives, so a body of `n` bytes costs
/// `log2(n / MAX_PREALLOC_BYTES)` reallocations and the buffer never runs more than a factor
/// of two ahead of the bytes actually received.
pub(crate) const MAX_PREALLOC_BYTES: usize = 64 * 1024 * 1024;

/// Reads exactly `len` bytes of message body, without reserving `len` before reading it.
fn read_body_bounded<R: Read>(reader: &mut R, len: usize) -> Result<MutableBuffer, ArrowError> {
    let mut buf = MutableBuffer::try_from_len_zeroed(len.min(MAX_PREALLOC_BYTES))
        .map_err(|e| ArrowError::MemoryError(e.to_string()))?;
    let mut filled = 0;
    while filled < len {
        let target = buf.len();
        reader.read_exact(&mut buf.as_slice_mut()[filled..target])?;
        filled = target;
        if filled < len {
            buf.try_resize(len.min(target.saturating_mul(2)), 0)
                .map_err(|e| ArrowError::MemoryError(e.to_string()))?;
        }
    }
    Ok(buf)
}

/// A low-level construct that reads [`Message::Message`]s from a reader while
/// re-using a buffer for metadata. This is composed into [`StreamReader`].
struct MessageReader<R> {
    reader: R,
    buf: Vec<u8>,
}

impl<R: Read> MessageReader<R> {
    fn new(reader: R) -> Self {
        Self {
            reader,
            buf: Vec::new(),
        }
    }

    /// Reads the entire next message from the underlying reader which includes
    /// the metadata length, the metadata, and the body.
    ///
    /// # Returns
    /// - `Ok(None)` if the the reader signals the end of stream with EOF on
    ///   the first read
    /// - `Err(_)` if the reader returns an error other than on the first
    ///   read, or if the metadata length is invalid
    /// - `Ok(Some(_))` with the Message and buffer containing the
    ///   body bytes otherwise.
    fn maybe_next(&mut self) -> Result<Option<(Message::Message<'_>, MutableBuffer)>, ArrowError> {
        let meta_len = self.read_meta_len()?;
        let Some(meta_len) = meta_len else {
            return Ok(None);
        };

        // `read_to_end` on a `Take` grows with the bytes that arrive, so an implausible
        // `meta_len` costs a short read rather than the allocation it asks for.
        self.buf.clear();
        let read = (&mut self.reader)
            .take(meta_len as u64)
            .read_to_end(&mut self.buf)?;
        if read != meta_len {
            return Err(ArrowError::ParseError(format!(
                "Unexpected end of stream: expected {meta_len} metadata bytes, got {read}"
            )));
        }

        let message = crate::root_as_message(self.buf.as_slice()).map_err(|err| {
            ArrowError::ParseError(format!("Unable to get root as message: {err:?}"))
        })?;

        let body_len = usize::try_from(message.bodyLength()).map_err(|_| {
            ArrowError::ParseError(format!(
                "Invalid IPC message body length: {}",
                message.bodyLength()
            ))
        })?;
        let buf = read_body_bounded(&mut self.reader, body_len)?;

        Ok(Some((message, buf)))
    }

    /// Get a mutable reference to the underlying reader.
    fn inner_mut(&mut self) -> &mut R {
        &mut self.reader
    }

    /// Get an immutable reference to the underlying reader.
    fn inner(&self) -> &R {
        &self.reader
    }

    /// Read the metadata length for the next message from the underlying stream.
    ///
    /// # Returns
    /// - `Ok(None)` if the the reader signals the end of stream with EOF on
    ///   the first read
    /// - `Err(_)` if the reader returns an error other than on the first
    ///   read, or if the metadata length is less than 0.
    /// - `Ok(Some(_))` with the length otherwise.
    pub fn read_meta_len(&mut self) -> Result<Option<usize>, ArrowError> {
        let mut meta_len: [u8; 4] = [0; 4];
        match self.reader.read_exact(&mut meta_len) {
            Ok(()) => {}
            Err(e) => {
                return if e.kind() == std::io::ErrorKind::UnexpectedEof {
                    // Handle EOF without the "0xFFFFFFFF 0x00000000"
                    // valid according to:
                    // https://arrow.apache.org/docs/format/Columnar.html#ipc-streaming-format
                    Ok(None)
                } else {
                    Err(ArrowError::from(e))
                };
            }
        }

        let meta_len = {
            // If a continuation marker is encountered, skip over it and read
            // the size from the next four bytes.
            if meta_len == CONTINUATION_MARKER {
                self.reader.read_exact(&mut meta_len)?;
            }

            i32::from_le_bytes(meta_len)
        };

        if meta_len == 0 {
            return Ok(None);
        }

        let meta_len = usize::try_from(meta_len)
            .map_err(|_| ArrowError::ParseError(format!("Invalid metadata length: {meta_len}")))?;

        Ok(Some(meta_len))
    }
}

#[cfg(test)]
mod tests {
    use std::io::Cursor;

    use crate::convert::try_fb_to_schema;
    use crate::writer::{
        DictionaryTracker, IpcDataGenerator, IpcWriteOptions, unslice_run_array, write_message,
    };

    use super::*;

    use crate::{root_as_footer, root_as_message, size_prefixed_root_as_message};
    use arrow_array::builder::{PrimitiveRunBuilder, UnionBuilder};
    use arrow_array::types::*;
    use arrow_buffer::{NullBuffer, OffsetBuffer};
    use arrow_data::ArrayDataBuilder;

    fn create_test_projection_schema() -> Schema {
        // define field types
        let list_data_type = DataType::List(Arc::new(Field::new_list_field(DataType::Int32, true)));

        let fixed_size_list_data_type =
            DataType::FixedSizeList(Arc::new(Field::new_list_field(DataType::Int32, false)), 3);

        let union_fields = UnionFields::from_fields(vec![
            Field::new("a", DataType::Int32, false),
            Field::new("b", DataType::Float64, false),
        ]);

        let union_data_type = DataType::Union(union_fields, UnionMode::Dense);

        let struct_fields = Fields::from(vec![
            Field::new("id", DataType::Int32, false),
            Field::new_list("list", Field::new_list_field(DataType::Int8, true), false),
        ]);
        let struct_data_type = DataType::Struct(struct_fields);

        let run_encoded_data_type = DataType::RunEndEncoded(
            Arc::new(Field::new(
                Field::REE_RUN_ENDS_FIELD_DEFAULT_NAME,
                DataType::Int16,
                false,
            )),
            Arc::new(Field::new(
                Field::REE_VALUES_FIELD_DEFAULT_NAME,
                DataType::Int32,
                true,
            )),
        );

        // define schema
        Schema::new(vec![
            Field::new("f0", DataType::UInt32, false),
            Field::new("f1", DataType::Utf8, false),
            Field::new("f2", DataType::Boolean, false),
            Field::new("f3", union_data_type, true),
            Field::new("f4", DataType::Null, true),
            Field::new("f5", DataType::Float64, true),
            Field::new("f6", list_data_type, false),
            Field::new("f7", DataType::FixedSizeBinary(3), true),
            Field::new("f8", fixed_size_list_data_type, false),
            Field::new("f9", struct_data_type, false),
            Field::new("f10", run_encoded_data_type, false),
            Field::new("f11", DataType::Boolean, false),
            Field::new_dictionary("f12", DataType::Int8, DataType::Utf8, false),
            Field::new("f13", DataType::Utf8, false),
        ])
    }

    fn create_test_projection_batch_data(schema: &Schema) -> RecordBatch {
        // set test data for each column
        let array0 = UInt32Array::from(vec![1, 2, 3]);
        let array1 = StringArray::from(vec!["foo", "bar", "baz"]);
        let array2 = BooleanArray::from(vec![true, false, true]);

        let mut union_builder = UnionBuilder::new_dense();
        union_builder.append::<Int32Type>("a", 1).unwrap();
        union_builder.append::<Float64Type>("b", 10.1).unwrap();
        union_builder.append_null::<Float64Type>("b").unwrap();
        let array3 = union_builder.build().unwrap();

        let array4 = NullArray::new(3);
        let array5 = Float64Array::from(vec![Some(1.1), None, Some(3.3)]);
        let array6_values = vec![
            Some(vec![Some(10), Some(10), Some(10)]),
            Some(vec![Some(20), Some(20), Some(20)]),
            Some(vec![Some(30), Some(30)]),
        ];
        let array6 = ListArray::from_iter_primitive::<Int32Type, _, _>(array6_values);
        let array7_values = vec![vec![11, 12, 13], vec![22, 23, 24], vec![33, 34, 35]];
        let array7 = FixedSizeBinaryArray::try_from_iter(array7_values.into_iter()).unwrap();

        let array8_values = ArrayData::builder(DataType::Int32)
            .len(9)
            .add_buffer(Buffer::from_slice_ref([40, 41, 42, 43, 44, 45, 46, 47, 48]))
            .build()
            .unwrap();
        let array8_data = ArrayData::builder(schema.field(8).data_type().clone())
            .len(3)
            .add_child_data(array8_values)
            .build()
            .unwrap();
        let array8 = FixedSizeListArray::from(array8_data);

        let array9_id: ArrayRef = Arc::new(Int32Array::from(vec![1001, 1002, 1003]));
        let array9_list: ArrayRef =
            Arc::new(ListArray::from_iter_primitive::<Int8Type, _, _>(vec![
                Some(vec![Some(-10)]),
                Some(vec![Some(-20), Some(-20), Some(-20)]),
                Some(vec![Some(-30)]),
            ]));
        let array9 = ArrayDataBuilder::new(schema.field(9).data_type().clone())
            .add_child_data(array9_id.into_data())
            .add_child_data(array9_list.into_data())
            .len(3)
            .build()
            .unwrap();
        let array9 = StructArray::from(array9);

        let array10_input = vec![Some(1_i32), None, None];
        let mut array10_builder = PrimitiveRunBuilder::<Int16Type, Int32Type>::new();
        array10_builder.extend(array10_input);
        let array10 = array10_builder.finish();

        let array11 = BooleanArray::from(vec![false, false, true]);

        let array12_values = StringArray::from(vec!["x", "yy", "zzz"]);
        let array12_keys = Int8Array::from_iter_values([1, 1, 2]);
        let array12 = DictionaryArray::new(array12_keys, Arc::new(array12_values));

        let array13 = StringArray::from(vec!["a", "bb", "ccc"]);

        // create record batch
        RecordBatch::try_new(
            Arc::new(schema.clone()),
            vec![
                Arc::new(array0),
                Arc::new(array1),
                Arc::new(array2),
                Arc::new(array3),
                Arc::new(array4),
                Arc::new(array5),
                Arc::new(array6),
                Arc::new(array7),
                Arc::new(array8),
                Arc::new(array9),
                Arc::new(array10),
                Arc::new(array11),
                Arc::new(array12),
                Arc::new(array13),
            ],
        )
        .unwrap()
    }

    #[test]
    fn test_negative_meta_len_start_stream() {
        let bytes = i32::to_le_bytes(-1);
        let mut buf = vec![];
        buf.extend(CONTINUATION_MARKER);
        buf.extend(bytes);

        let reader_err = StreamReader::try_new(Cursor::new(buf), None).err();
        assert!(reader_err.is_some());
        assert_eq!(
            reader_err.unwrap().to_string(),
            "Parser error: Invalid metadata length: -1"
        );
    }

    #[test]
    fn test_negative_meta_len_mid_stream() {
        let schema = Schema::new(vec![Field::new("a", DataType::Int32, false)]);
        let mut buf = Vec::new();
        {
            let mut writer = crate::writer::StreamWriter::try_new(&mut buf, &schema).unwrap();
            let batch =
                RecordBatch::try_new(Arc::new(schema), vec![Arc::new(Int32Array::from(vec![1]))])
                    .unwrap();
            writer.write(&batch).unwrap();
        }

        let bytes = i32::to_le_bytes(-1);
        buf.extend(CONTINUATION_MARKER);
        buf.extend(bytes);

        let mut reader = StreamReader::try_new(Cursor::new(buf), None).unwrap();
        // Read the valid value
        assert!(reader.maybe_next().is_ok());
        // Read the invalid meta len
        let batch_err = reader.maybe_next().err();
        assert!(batch_err.is_some());
        assert_eq!(
            batch_err.unwrap().to_string(),
            "Parser error: Invalid metadata length: -1"
        );
    }

    #[test]
    fn test_invalid_dictionary_batch_without_data() {
        use crate::r#gen::Message::*;
        use flatbuffers::FlatBufferBuilder;

        let schema = Schema::new(vec![Field::new(
            "col",
            DataType::Dictionary(Box::new(DataType::Int8), Box::new(DataType::Utf8)),
            true,
        )]);

        // DictionaryBatch.data is optional in the flatbuffer grammar and
        // required by the format
        let mut fbb = FlatBufferBuilder::new();
        let batch_offset = DictionaryBatch::create(
            &mut fbb,
            &DictionaryBatchArgs {
                id: 0,
                data: None,
                isDelta: false,
            },
        );
        fbb.finish_minimal(batch_offset);
        let batch_bytes = fbb.finished_data().to_vec();
        let batch = flatbuffers::root::<DictionaryBatch>(&batch_bytes).unwrap();

        let data_buffer = Buffer::from(vec![0u8; 0]);
        let mut dictionaries: HashMap<i64, ArrayRef> = HashMap::new();

        let err = read_dictionary(
            &data_buffer,
            batch,
            &schema,
            &mut dictionaries,
            &MetadataVersion::V5,
        )
        .unwrap_err();

        assert_eq!(
            err.to_string(),
            "Parser error: Dictionary batch is missing its data"
        );
    }

    #[test]
    fn test_missing_buffer_metadata_error() {
        use crate::r#gen::Message::*;
        use flatbuffers::FlatBufferBuilder;

        let schema = Arc::new(Schema::new(vec![Field::new("col", DataType::Int32, true)]));

        // create RecordBatch buffer metadata with invalid buffer count
        // Int32Array needs 2 buffers (validity + data) but we provide only 1
        let mut fbb = FlatBufferBuilder::new();
        let nodes = fbb.create_vector(&[FieldNode::new(2, 0)]);
        let buffers = fbb.create_vector(&[crate::Buffer::new(0, 8)]);
        let batch_offset = RecordBatch::create(
            &mut fbb,
            &RecordBatchArgs {
                length: 2,
                nodes: Some(nodes),
                buffers: Some(buffers),
                compression: None,
                variadicBufferCounts: None,
            },
        );
        fbb.finish_minimal(batch_offset);
        let batch_bytes = fbb.finished_data().to_vec();
        let batch = flatbuffers::root::<RecordBatch>(&batch_bytes).unwrap();

        let data_buffer = Buffer::from(vec![0u8; 8]);
        let dictionaries: HashMap<i64, ArrayRef> = HashMap::new();
        let metadata = MetadataVersion::V5;

        let decoder = RecordBatchDecoder::try_new(
            &data_buffer,
            batch,
            schema.clone(),
            &dictionaries,
            &metadata,
        )
        .unwrap();

        let result = decoder.read_record_batch();

        match result {
            Err(ArrowError::IpcError(msg)) => {
                assert_eq!(msg, "Buffer count mismatched with metadata");
            }
            other => panic!("unexpected error: {other:?}"),
        }
    }

    #[test]
    fn test_read_record_batch_buffer_out_of_bounds() {
        use crate::r#gen::Message::*;
        use flatbuffers::FlatBufferBuilder;

        let schema = Arc::new(Schema::new(vec![Field::new("col", DataType::Int32, false)]));
        let body = Buffer::from(vec![0u8; 8]);
        let mut fbb = FlatBufferBuilder::new();
        let nodes = fbb.create_vector(&[FieldNode::new(2, 0)]);
        // Empty validity buffer followed by two Int32 values whose range exceeds the body.
        let buffers = fbb.create_vector(&[crate::Buffer::new(0, 0), crate::Buffer::new(4, 8)]);
        let batch = RecordBatch::create(
            &mut fbb,
            &RecordBatchArgs {
                length: 2,
                nodes: Some(nodes),
                buffers: Some(buffers),
                ..Default::default()
            },
        );
        fbb.finish_minimal(batch);
        let batch = flatbuffers::root::<RecordBatch>(fbb.finished_data()).unwrap();

        read_record_batch(
            &body,
            batch,
            schema,
            &HashMap::new(),
            None,
            &MetadataVersion::V5,
        )
        .expect_err("invalid IPC buffer range must return an error");
    }

    #[test]
    fn test_union_short_buffers() {
        use crate::r#gen::Message::*;

        let fields = UnionFields::from_fields([Field::new("child", DataType::Null, true)]);
        // All buffer ranges fit in the body. Only the union length is inconsistent:
        // sparse has one type id for two rows, dense has one offset for two rows
        let cases = [
            (UnionMode::Sparse, vec![crate::Buffer::new(0, 1)]),
            (
                UnionMode::Dense,
                vec![crate::Buffer::new(0, 2), crate::Buffer::new(8, 4)],
            ),
        ];
        for (mode, buffers) in cases {
            let schema = Arc::new(Schema::new(vec![Field::new(
                "union",
                DataType::Union(fields.clone(), mode),
                false,
            )]));

            let mut fbb = flatbuffers::FlatBufferBuilder::new();
            let nodes = fbb.create_vector(&[FieldNode::new(2, 0), FieldNode::new(2, 2)]);
            let buffers = fbb.create_vector(&buffers);
            let batch = RecordBatch::create(
                &mut fbb,
                &RecordBatchArgs {
                    length: 2,
                    nodes: Some(nodes),
                    buffers: Some(buffers),
                    ..Default::default()
                },
            );
            fbb.finish_minimal(batch);
            let batch = flatbuffers::root::<RecordBatch>(fbb.finished_data()).unwrap();

            read_record_batch(
                &Buffer::from(vec![0u8; 16]),
                batch,
                schema,
                &HashMap::new(),
                None,
                &MetadataVersion::V5,
            )
            .expect_err("a short union buffer should return an error");
        }
    }

    #[test]
    fn test_offsets_buffer_with_trailing_bytes() {
        use crate::r#gen::Message::*;
        use flatbuffers::FlatBufferBuilder;

        let schema = Arc::new(Schema::new(vec![Field::new("col", DataType::Utf8, false)]));

        // The offsets buffer is 14 bytes: three i32 offsets and two trailing
        // bytes, which is more than the array needs but not a whole number
        // of offsets.
        let mut fbb = FlatBufferBuilder::new();
        let nodes = fbb.create_vector(&[FieldNode::new(2, 0)]);
        let buffers = fbb.create_vector(&[
            crate::Buffer::new(0, 0),
            crate::Buffer::new(0, 14),
            crate::Buffer::new(16, 5),
        ]);
        let batch_offset = RecordBatch::create(
            &mut fbb,
            &RecordBatchArgs {
                length: 2,
                nodes: Some(nodes),
                buffers: Some(buffers),
                compression: None,
                variadicBufferCounts: None,
            },
        );
        fbb.finish_minimal(batch_offset);
        let batch_bytes = fbb.finished_data().to_vec();
        let batch = flatbuffers::root::<RecordBatch>(&batch_bytes).unwrap();

        let mut body = Buffer::from_slice_ref([0_i32, 2, 5]).as_slice().to_vec();
        body.extend_from_slice(&[0; 4]);
        body.extend_from_slice(b"hello\0\0\0");
        let data_buffer = Buffer::from_slice_ref(&body);
        let dictionaries: HashMap<i64, ArrayRef> = HashMap::new();
        let metadata = MetadataVersion::V5;

        let decoder =
            RecordBatchDecoder::try_new(&data_buffer, batch, schema, &dictionaries, &metadata)
                .unwrap();
        let batch = decoder.read_record_batch().unwrap();
        let col = batch.column(0).as_any().downcast_ref::<StringArray>();
        assert_eq!(col.unwrap(), &StringArray::from(vec!["he", "llo"]));
    }

    /// A `Utf8View` batch whose variadic buffer count is `count`, with two buffers
    /// in the message. Returns the error from reading it with the given projection.
    fn read_batch_with_variadic_count(count: i64, projection: Option<&[usize]>) -> ArrowError {
        use crate::r#gen::Message::*;
        use flatbuffers::FlatBufferBuilder;

        let schema = Arc::new(Schema::new(vec![Field::new(
            "col",
            DataType::Utf8View,
            true,
        )]));

        let mut fbb = FlatBufferBuilder::new();
        let nodes = fbb.create_vector(&[FieldNode::new(1, 0)]);
        let buffers = fbb.create_vector(&[crate::Buffer::new(0, 8), crate::Buffer::new(8, 8)]);
        let variadic_buffer_counts = fbb.create_vector(&[count]);
        let batch_offset = RecordBatch::create(
            &mut fbb,
            &RecordBatchArgs {
                length: 1,
                nodes: Some(nodes),
                buffers: Some(buffers),
                compression: None,
                variadicBufferCounts: Some(variadic_buffer_counts),
            },
        );
        fbb.finish_minimal(batch_offset);
        let batch_bytes = fbb.finished_data().to_vec();
        let batch = flatbuffers::root::<RecordBatch>(&batch_bytes).unwrap();

        let data_buffer = Buffer::from(vec![0u8; 16]);
        let dictionaries: HashMap<i64, ArrayRef> = HashMap::new();

        RecordBatchDecoder::try_new(
            &data_buffer,
            batch,
            schema,
            &dictionaries,
            &MetadataVersion::V5,
        )
        .unwrap()
        .with_projection(projection)
        .read_record_batch()
        .expect_err("should get error")
    }

    /// A variadic count the message cannot honour used to panic while slicing the
    /// buffers it did not read, both when reading the column and when skipping it.
    #[test]
    fn test_invalid_variadic_buffer_count_error() {
        // -2 leaves no buffers at all, -1 leaves too few, and 1 asks for more than the
        // message has. The projection selects nothing, so the column is skipped instead.
        for count in [-2, -1, 1, i64::MAX] {
            for projection in [None, Some([].as_slice())] {
                let err = read_batch_with_variadic_count(count, projection);
                assert_eq!(
                    err.to_string(),
                    format!(
                        "Ipc error: Invalid variadic count {count} for Utf8View column, \
                         with 2 buffer(s) left in the message"
                    ),
                    "count {count}, projection {projection:?}"
                );
            }
        }
    }

    /// The valid count for a message with two buffers is zero.
    #[test]
    fn test_valid_variadic_buffer_count_is_accepted() {
        let err = read_batch_with_variadic_count(0, None);
        assert!(!err.to_string().contains("Invalid variadic count"), "{err}");
    }

    #[test]
    fn test_missing_footer_schema_error() {
        use crate::r#gen::File::{Footer, FooterArgs};
        use flatbuffers::FlatBufferBuilder;

        // a footer that verifies but has no schema table. record batches present
        // (so the earlier ok_or_else passes) but schema absent, which used to panic.
        let mut fbb = FlatBufferBuilder::new();
        let record_batches = fbb.create_vector::<Block>(&[]);
        let footer = Footer::create(
            &mut fbb,
            &FooterArgs {
                version: MetadataVersion::V5,
                schema: None,
                dictionaries: None,
                recordBatches: Some(record_batches),
                custom_metadata: None,
            },
        );
        fbb.finish(footer, None);
        let footer_data = fbb.finished_data();

        // assemble a minimal IPC file: magic header, footer, footer length, magic trailer
        let mut buf = Vec::new();
        buf.extend_from_slice(&crate::ARROW_MAGIC);
        buf.extend_from_slice(footer_data);
        buf.extend_from_slice(&(footer_data.len() as i32).to_le_bytes());
        buf.extend_from_slice(&crate::ARROW_MAGIC);

        let err = FileReader::try_new(Cursor::new(buf), None)
            .expect_err("expected an error, not a panic");
        assert!(
            matches!(err, ArrowError::ParseError(_)),
            "expected ParseError, got {err:?}"
        );
    }

    /// Test that the reader can read legacy files where empty list arrays were written with a 0-byte offsets buffer.
    #[test]
    fn test_read_legacy_empty_list_without_offsets_buffer() {
        use crate::r#gen::Message::*;
        use flatbuffers::FlatBufferBuilder;

        let schema = Arc::new(Schema::new(vec![Field::new_list(
            "items",
            Field::new_list_field(DataType::Int32, true),
            true,
        )]));

        // Legacy arrow-rs versions wrote empty offsets buffers for empty list arrays.
        // Keep reader compatibility with such files by accepting a 0-byte offsets buffer.
        let mut fbb = FlatBufferBuilder::new();
        let nodes = fbb.create_vector(&[
            FieldNode::new(0, 0), // list node
            FieldNode::new(0, 0), // child int32 node
        ]);
        let buffers = fbb.create_vector(&[
            crate::Buffer::new(0, 0), // list validity
            crate::Buffer::new(0, 0), // list offsets (legacy empty buffer)
            crate::Buffer::new(0, 0), // child validity
            crate::Buffer::new(0, 0), // child values
        ]);
        let batch_offset = RecordBatch::create(
            &mut fbb,
            &RecordBatchArgs {
                length: 0,
                nodes: Some(nodes),
                buffers: Some(buffers),
                compression: None,
                variadicBufferCounts: None,
            },
        );
        fbb.finish_minimal(batch_offset);
        let batch_bytes = fbb.finished_data().to_vec();
        let batch = flatbuffers::root::<RecordBatch>(&batch_bytes).unwrap();

        let body = Buffer::from(Vec::<u8>::new());
        let dictionaries: HashMap<i64, ArrayRef> = HashMap::new();
        let metadata = MetadataVersion::V5;

        let decoder =
            RecordBatchDecoder::try_new(&body, batch, schema.clone(), &dictionaries, &metadata)
                .unwrap();

        let read_batch = decoder.read_record_batch().unwrap();
        assert_eq!(read_batch.num_rows(), 0);

        let list = read_batch
            .column(0)
            .as_any()
            .downcast_ref::<ListArray>()
            .unwrap();
        assert_eq!(list.len(), 0);
        assert_eq!(list.values().len(), 0);
    }

    /// Test that the reader can read legacy files where empty Utf8/Binary arrays were written with a 0-byte offsets buffer.
    #[test]
    fn test_read_legacy_empty_utf8_and_binary_without_offsets_buffer() {
        use crate::r#gen::Message::*;
        use flatbuffers::FlatBufferBuilder;

        let schema = Arc::new(Schema::new(vec![
            Field::new("name", DataType::Utf8, true),
            Field::new("payload", DataType::Binary, true),
        ]));

        // Legacy arrow-rs versions wrote empty offsets buffers for empty Utf8/Binary arrays.
        // Keep reader compatibility with such files by accepting 0-byte offsets buffers.
        let mut fbb = FlatBufferBuilder::new();
        let nodes = fbb.create_vector(&[
            FieldNode::new(0, 0), // utf8 node
            FieldNode::new(0, 0), // binary node
        ]);
        let buffers = fbb.create_vector(&[
            crate::Buffer::new(0, 0), // utf8 validity
            crate::Buffer::new(0, 0), // utf8 offsets (legacy empty buffer)
            crate::Buffer::new(0, 0), // utf8 values
            crate::Buffer::new(0, 0), // binary validity
            crate::Buffer::new(0, 0), // binary offsets (legacy empty buffer)
            crate::Buffer::new(0, 0), // binary values
        ]);
        let batch_offset = RecordBatch::create(
            &mut fbb,
            &RecordBatchArgs {
                length: 0,
                nodes: Some(nodes),
                buffers: Some(buffers),
                compression: None,
                variadicBufferCounts: None,
            },
        );
        fbb.finish_minimal(batch_offset);
        let batch_bytes = fbb.finished_data().to_vec();
        let batch = flatbuffers::root::<RecordBatch>(&batch_bytes).unwrap();

        let body = Buffer::from(Vec::<u8>::new());
        let dictionaries: HashMap<i64, ArrayRef> = HashMap::new();
        let metadata = MetadataVersion::V5;

        let decoder =
            RecordBatchDecoder::try_new(&body, batch, schema.clone(), &dictionaries, &metadata)
                .unwrap();

        let read_batch = decoder.read_record_batch().unwrap();
        assert_eq!(read_batch.num_rows(), 0);

        let utf8 = read_batch
            .column(0)
            .as_any()
            .downcast_ref::<StringArray>()
            .unwrap();
        assert_eq!(utf8.len(), 0);
        assert_eq!(utf8.value_offsets(), [0]);

        let binary = read_batch
            .column(1)
            .as_any()
            .downcast_ref::<BinaryArray>()
            .unwrap();
        assert_eq!(binary.len(), 0);
        assert_eq!(binary.value_offsets(), [0]);
    }

    #[test]
    fn test_projection_array_values() {
        // define schema
        let schema = create_test_projection_schema();

        // create record batch with test data
        let batch = create_test_projection_batch_data(&schema);

        // write record batch in IPC format
        let mut buf = Vec::new();
        {
            let mut writer = crate::writer::FileWriter::try_new(&mut buf, &schema).unwrap();
            writer.write(&batch).unwrap();
            writer.finish().unwrap();
        }

        // read record batch with projection
        for index in 0..12 {
            let projection = vec![index];
            let reader = FileReader::try_new(std::io::Cursor::new(buf.clone()), Some(projection));
            let read_batch = reader.unwrap().next().unwrap().unwrap();
            let projected_column = read_batch.column(0);
            let expected_column = batch.column(index);

            // check the projected column equals the expected column
            assert_eq!(projected_column.as_ref(), expected_column.as_ref());
        }

        {
            // read record batch with reversed projection
            let reader =
                FileReader::try_new(std::io::Cursor::new(buf.clone()), Some(vec![3, 2, 1]));
            let read_batch = reader.unwrap().next().unwrap().unwrap();
            let expected_batch = batch.project(&[3, 2, 1]).unwrap();
            assert_eq!(read_batch, expected_batch);
        }
    }

    #[test]
    fn test_file_reader_projected_schema_matches_batch_schema() {
        let schema = create_test_projection_schema();
        let batch = create_test_projection_batch_data(&schema);

        let mut buf = Vec::new();
        {
            let mut writer = crate::writer::FileWriter::try_new(&mut buf, &schema).unwrap();
            writer.write(&batch).unwrap();
            writer.finish().unwrap();
        }

        let projection = vec![3, 2, 1];
        let mut reader = FileReader::try_new(Cursor::new(buf), Some(projection)).unwrap();
        let reader_schema = RecordBatchReader::schema(&reader);
        let read_batch = reader.next().unwrap().unwrap();

        assert_eq!(reader_schema, read_batch.schema());
    }

    #[test]
    fn test_stream_reader_projected_schema_matches_batch_schema() {
        let schema = create_test_projection_schema();
        let batch = create_test_projection_batch_data(&schema);

        let mut buf = Vec::new();
        {
            let mut writer = crate::writer::StreamWriter::try_new(&mut buf, &schema).unwrap();
            writer.write(&batch).unwrap();
            writer.finish().unwrap();
        }

        let projection = vec![3, 2, 1];
        let mut reader = StreamReader::try_new(Cursor::new(buf), Some(projection)).unwrap();
        let reader_schema = RecordBatchReader::schema(&reader);
        let read_batch = reader.next().unwrap().unwrap();

        assert_eq!(reader_schema, read_batch.schema());
    }

    #[test]
    fn test_file_reader_rejects_invalid_projection() {
        let schema = create_test_projection_schema();
        let batch = create_test_projection_batch_data(&schema);

        let mut buf = Vec::new();
        {
            let mut writer = crate::writer::FileWriter::try_new(&mut buf, &schema).unwrap();
            writer.write(&batch).unwrap();
            writer.finish().unwrap();
        }

        let result = FileReader::try_new(Cursor::new(buf), Some(vec![schema.fields().len()]));

        assert!(matches!(result, Err(ArrowError::SchemaError(_))));
    }

    #[test]
    fn test_projection_duplicate_indices() {
        let schema = create_test_projection_schema();
        let batch = create_test_projection_batch_data(&schema);

        // Write the batch to IPC
        let mut buf = Vec::new();
        {
            let mut writer = crate::writer::FileWriter::try_new(&mut buf, &schema).unwrap();
            writer.write(&batch).unwrap();
            writer.finish().unwrap();
        }

        // Verify duplicate([1, 1]) and reordered([2, 0, 2]) projection indices
        for projection in [vec![1, 1], vec![2, 0, 2]] {
            let reader =
                FileReader::try_new(std::io::Cursor::new(buf.clone()), Some(projection.clone()));
            let read_batch = reader.unwrap().next().unwrap().unwrap();

            let expected_batch = batch.project(&projection).unwrap();
            assert_eq!(read_batch, expected_batch);
        }
    }

    #[test]
    fn test_arrow_single_float_row() {
        let schema = Schema::new(vec![
            Field::new("a", DataType::Float32, false),
            Field::new("b", DataType::Float32, false),
            Field::new("c", DataType::Int32, false),
            Field::new("d", DataType::Int32, false),
        ]);
        let arrays = vec![
            Arc::new(Float32Array::from(vec![1.23])) as ArrayRef,
            Arc::new(Float32Array::from(vec![-6.50])) as ArrayRef,
            Arc::new(Int32Array::from(vec![2])) as ArrayRef,
            Arc::new(Int32Array::from(vec![1])) as ArrayRef,
        ];
        let batch = RecordBatch::try_new(Arc::new(schema.clone()), arrays).unwrap();
        // create stream writer
        let mut file = tempfile::tempfile().unwrap();
        let mut stream_writer = crate::writer::StreamWriter::try_new(&mut file, &schema).unwrap();
        stream_writer.write(&batch).unwrap();
        stream_writer.finish().unwrap();

        drop(stream_writer);

        file.rewind().unwrap();

        // read stream back
        let reader = StreamReader::try_new(&mut file, None).unwrap();

        reader.for_each(|batch| {
            let batch = batch.unwrap();
            assert!(
                batch
                    .column(0)
                    .as_any()
                    .downcast_ref::<Float32Array>()
                    .unwrap()
                    .value(0)
                    != 0.0
            );
            assert!(
                batch
                    .column(1)
                    .as_any()
                    .downcast_ref::<Float32Array>()
                    .unwrap()
                    .value(0)
                    != 0.0
            );
        });

        file.rewind().unwrap();

        // Read with projection
        let reader = StreamReader::try_new(file, Some(vec![0, 3])).unwrap();

        reader.for_each(|batch| {
            let batch = batch.unwrap();
            assert_eq!(batch.schema().fields().len(), 2);
            assert_eq!(batch.schema().fields()[0].data_type(), &DataType::Float32);
            assert_eq!(batch.schema().fields()[1].data_type(), &DataType::Int32);
        });
    }

    /// Write the record batch to an in-memory buffer in IPC File format
    fn write_ipc(rb: &RecordBatch) -> Vec<u8> {
        let mut buf = Vec::new();
        let mut writer = crate::writer::FileWriter::try_new(&mut buf, rb.schema_ref()).unwrap();
        writer.write(rb).unwrap();
        writer.finish().unwrap();
        buf
    }

    /// Return the first record batch read from the IPC File buffer
    fn read_ipc(buf: &[u8]) -> Result<RecordBatch, ArrowError> {
        let mut reader = FileReader::try_new(std::io::Cursor::new(buf), None)?;
        reader.next().unwrap()
    }

    /// Return the first record batch read from the IPC File buffer, disabling
    /// validation
    fn read_ipc_skip_validation(buf: &[u8]) -> Result<RecordBatch, ArrowError> {
        let mut reader = unsafe {
            FileReader::try_new(std::io::Cursor::new(buf), None)?.with_skip_validation(true)
        };
        reader.next().unwrap()
    }

    fn roundtrip_ipc(rb: &RecordBatch) -> RecordBatch {
        let buf = write_ipc(rb);
        read_ipc(&buf).unwrap()
    }

    /// Return the first record batch read from the IPC File buffer
    /// using the FileDecoder API
    fn read_ipc_with_decoder(buf: Vec<u8>) -> Result<RecordBatch, ArrowError> {
        read_ipc_with_decoder_inner(buf, false)
    }

    /// Return the first record batch read from the IPC File buffer
    /// using the FileDecoder API, disabling validation
    fn read_ipc_with_decoder_skip_validation(buf: Vec<u8>) -> Result<RecordBatch, ArrowError> {
        read_ipc_with_decoder_inner(buf, true)
    }

    fn read_ipc_with_decoder_inner(
        buf: Vec<u8>,
        skip_validation: bool,
    ) -> Result<RecordBatch, ArrowError> {
        let buffer = Buffer::from_vec(buf);
        let trailer_start = buffer.len() - 10;
        let footer_len = read_footer_length(buffer[trailer_start..].try_into().unwrap())?;
        let footer = root_as_footer(&buffer[trailer_start - footer_len..trailer_start])
            .map_err(|e| ArrowError::InvalidArgumentError(format!("Invalid footer: {e}")))?;

        let schema = try_fb_to_schema(footer.schema().unwrap()).unwrap();

        let mut decoder = unsafe {
            FileDecoder::new(Arc::new(schema), footer.version())
                .with_skip_validation(skip_validation)
        };
        // Read dictionaries
        for block in footer.dictionaries().iter().flatten() {
            let block_len = block.bodyLength() as usize + block.metaDataLength() as usize;
            let data = buffer.slice_with_length(block.offset() as _, block_len);
            decoder.read_dictionary(block, &data)?
        }

        // Read record batch
        let batches = footer.recordBatches().unwrap();
        assert_eq!(batches.len(), 1); // Only wrote a single batch

        let block = batches.get(0);
        let block_len = block.bodyLength() as usize + block.metaDataLength() as usize;
        let data = buffer.slice_with_length(block.offset() as _, block_len);
        Ok(decoder.read_record_batch(block, &data)?.unwrap())
    }

    /// Write the record batch to an in-memory buffer in IPC Stream format
    fn write_stream(rb: &RecordBatch) -> Vec<u8> {
        let mut buf = Vec::new();
        let mut writer = crate::writer::StreamWriter::try_new(&mut buf, rb.schema_ref()).unwrap();
        writer.write(rb).unwrap();
        writer.finish().unwrap();
        buf
    }

    /// Return the first record batch read from the IPC Stream buffer
    fn read_stream(buf: &[u8]) -> Result<RecordBatch, ArrowError> {
        let mut reader = StreamReader::try_new(std::io::Cursor::new(buf), None)?;
        reader.next().unwrap()
    }

    /// Return the first record batch read from the IPC Stream buffer,
    /// disabling validation
    fn read_stream_skip_validation(buf: &[u8]) -> Result<RecordBatch, ArrowError> {
        let mut reader = unsafe {
            StreamReader::try_new(std::io::Cursor::new(buf), None)?.with_skip_validation(true)
        };
        reader.next().unwrap()
    }

    /// Return the first record batch read from the IPC Stream buffer
    /// using the StreamDecoder API
    fn read_stream_with_decoder(
        buf: &[u8],
        skip_validation: bool,
    ) -> Result<RecordBatch, ArrowError> {
        let mut decoder = unsafe { StreamDecoder::new().with_skip_validation(skip_validation) };
        Ok(decoder.decode(&mut Buffer::from(buf))?.unwrap())
    }

    fn roundtrip_ipc_stream(rb: &RecordBatch) -> RecordBatch {
        let buf = write_stream(rb);
        read_stream(&buf).unwrap()
    }

    #[test]
    fn test_roundtrip_with_custom_metadata() {
        let schema = Schema::new(vec![Field::new("dummy", DataType::Float64, false)]);
        let mut buf = Vec::new();
        let mut writer = crate::writer::FileWriter::try_new(&mut buf, &schema).unwrap();
        let mut test_metadata = HashMap::new();
        test_metadata.insert("abc".to_string(), "abc".to_string());
        test_metadata.insert("def".to_string(), "def".to_string());
        for (k, v) in &test_metadata {
            writer.write_metadata(k, v);
        }
        writer.finish().unwrap();
        drop(writer);

        let reader = crate::reader::FileReader::try_new(std::io::Cursor::new(buf), None).unwrap();
        assert_eq!(reader.custom_metadata(), &test_metadata);
    }

    #[test]
    fn test_roundtrip_nested_dict() {
        let inner: DictionaryArray<Int32Type> = vec!["a", "b", "a"].into_iter().collect();

        let array = Arc::new(inner) as ArrayRef;

        let dctfield = Arc::new(Field::new("dict", array.data_type().clone(), false));

        let s = StructArray::from(vec![(dctfield, array)]);
        let struct_array = Arc::new(s) as ArrayRef;

        let schema = Arc::new(Schema::new(vec![Field::new(
            "struct",
            struct_array.data_type().clone(),
            false,
        )]));

        let batch = RecordBatch::try_new(schema, vec![struct_array]).unwrap();

        assert_eq!(batch, roundtrip_ipc(&batch));
    }

    #[test]
    fn test_roundtrip_nested_dict_no_preserve_dict_id() {
        let inner: DictionaryArray<Int32Type> = vec!["a", "b", "a"].into_iter().collect();

        let array = Arc::new(inner) as ArrayRef;

        let dctfield = Arc::new(Field::new("dict", array.data_type().clone(), false));

        let s = StructArray::from(vec![(dctfield, array)]);
        let struct_array = Arc::new(s) as ArrayRef;

        let schema = Arc::new(Schema::new(vec![Field::new(
            "struct",
            struct_array.data_type().clone(),
            false,
        )]));

        let batch = RecordBatch::try_new(schema, vec![struct_array]).unwrap();

        let mut buf = Vec::new();
        let mut writer = crate::writer::FileWriter::try_new_with_options(
            &mut buf,
            batch.schema_ref(),
            IpcWriteOptions::default(),
        )
        .unwrap();
        writer.write(&batch).unwrap();
        writer.finish().unwrap();
        drop(writer);

        let mut reader = FileReader::try_new(std::io::Cursor::new(buf), None).unwrap();

        assert_eq!(batch, reader.next().unwrap().unwrap());
    }

    fn check_union_with_builder(mut builder: UnionBuilder) {
        builder.append::<Int32Type>("a", 1).unwrap();
        builder.append_null::<Int32Type>("a").unwrap();
        builder.append::<Float64Type>("c", 3.0).unwrap();
        builder.append::<Int32Type>("a", 4).unwrap();
        builder.append::<Int64Type>("d", 11).unwrap();
        let union = builder.build().unwrap();

        let schema = Arc::new(Schema::new(vec![Field::new(
            "union",
            union.data_type().clone(),
            false,
        )]));

        let union_array = Arc::new(union) as ArrayRef;

        let rb = RecordBatch::try_new(schema, vec![union_array]).unwrap();
        let rb2 = roundtrip_ipc(&rb);
        // TODO: equality not yet implemented for union, so we check that the length of the array is
        // the same and that all of the buffers are the same instead.
        assert_eq!(rb.schema(), rb2.schema());
        assert_eq!(rb.num_columns(), rb2.num_columns());
        assert_eq!(rb.num_rows(), rb2.num_rows());
        let union1 = rb.column(0);
        let union2 = rb2.column(0);

        assert_eq!(union1, union2);
    }

    #[test]
    fn test_roundtrip_dense_union() {
        check_union_with_builder(UnionBuilder::new_dense());
    }

    #[test]
    fn test_roundtrip_sparse_union() {
        check_union_with_builder(UnionBuilder::new_sparse());
    }

    #[test]
    fn test_roundtrip_struct_empty_fields() {
        let nulls = NullBuffer::from(&[true, true, false]);
        let rb = RecordBatch::try_from_iter([(
            "",
            Arc::new(StructArray::new_empty_fields(nulls.len(), Some(nulls))) as _,
        )])
        .unwrap();
        let rb2 = roundtrip_ipc(&rb);
        assert_eq!(rb, rb2);
    }

    #[test]
    fn test_roundtrip_stream_run_array_sliced() {
        let run_array_1: Int32RunArray = vec!["a", "a", "a", "b", "b", "c", "c", "c"]
            .into_iter()
            .collect();
        let run_array_1_sliced = run_array_1.slice(2, 5);

        let run_array_2_inupt = vec![Some(1_i32), None, None, Some(2), Some(2)];
        let mut run_array_2_builder = PrimitiveRunBuilder::<Int16Type, Int32Type>::new();
        run_array_2_builder.extend(run_array_2_inupt);
        let run_array_2 = run_array_2_builder.finish();

        let schema = Arc::new(Schema::new(vec![
            Field::new(
                "run_array_1_sliced",
                run_array_1_sliced.data_type().clone(),
                false,
            ),
            Field::new("run_array_2", run_array_2.data_type().clone(), false),
        ]));
        let input_batch = RecordBatch::try_new(
            schema,
            vec![Arc::new(run_array_1_sliced.clone()), Arc::new(run_array_2)],
        )
        .unwrap();
        let output_batch = roundtrip_ipc_stream(&input_batch);

        // As partial comparison not yet supported for run arrays, the sliced run array
        // has to be unsliced before comparing with the output. the second run array
        // can be compared as such.
        assert_eq!(input_batch.column(1), output_batch.column(1));

        let run_array_1_unsliced = unslice_run_array(run_array_1_sliced.into_data()).unwrap();
        assert_eq!(run_array_1_unsliced, output_batch.column(0).into_data());
    }

    #[test]
    fn test_roundtrip_stream_nested_dict() {
        let xs = vec!["AA", "BB", "AA", "CC", "BB"];
        let dict = Arc::new(
            xs.clone()
                .into_iter()
                .collect::<DictionaryArray<Int8Type>>(),
        );
        let string_array: ArrayRef = Arc::new(StringArray::from(xs.clone()));
        let struct_array = StructArray::from(vec![
            (
                Arc::new(Field::new("f2.1", DataType::Utf8, false)),
                string_array,
            ),
            (
                Arc::new(Field::new("f2.2_struct", dict.data_type().clone(), false)),
                dict.clone() as ArrayRef,
            ),
        ]);
        let schema = Arc::new(Schema::new(vec![
            Field::new("f1_string", DataType::Utf8, false),
            Field::new("f2_struct", struct_array.data_type().clone(), false),
        ]));
        let input_batch = RecordBatch::try_new(
            schema,
            vec![
                Arc::new(StringArray::from(xs.clone())),
                Arc::new(struct_array),
            ],
        )
        .unwrap();
        let output_batch = roundtrip_ipc_stream(&input_batch);
        assert_eq!(input_batch, output_batch);
    }

    #[test]
    fn test_ipc_writers_reject_dictionary_of_dictionary_schema() {
        let values = Arc::new(StringArray::from(vec![Some("a"), Some("b")])) as ArrayRef;
        let inner = Arc::new(DictionaryArray::new(
            UInt32Array::from_iter_values([0, 1]),
            values,
        )) as ArrayRef;
        let outer = Arc::new(DictionaryArray::new(
            UInt32Array::from_iter_values([0, 1, 0]),
            inner,
        )) as ArrayRef;

        let schema = Arc::new(Schema::new(vec![Field::new(
            "f1",
            outer.data_type().clone(),
            false,
        )]));
        let batch = RecordBatch::try_new(schema, vec![outer]).unwrap();

        let mut stream = Vec::new();
        let Err(err) = crate::writer::StreamWriter::try_new(&mut stream, batch.schema_ref()) else {
            panic!("IPC stream writer should reject dictionary-of-dictionary schemas");
        };
        assert!(stream.is_empty());

        assert!(
            err.to_string().contains("dictionary-of-dictionary values"),
            "unexpected error: {err}"
        );

        let mut file = Vec::new();
        let Err(err) = crate::writer::FileWriter::try_new(&mut file, batch.schema_ref()) else {
            panic!("IPC file writer should reject dictionary-of-dictionary schemas");
        };
        assert!(file.is_empty());

        assert!(
            err.to_string().contains("dictionary-of-dictionary values"),
            "unexpected error: {err}"
        );
    }

    #[test]
    fn test_roundtrip_stream_nested_dict_of_map_of_dict() {
        let values = StringArray::from(vec![Some("a"), None, Some("b"), Some("c")]);
        let values = Arc::new(values) as ArrayRef;
        let value_dict_keys = Int8Array::from_iter_values([0, 1, 1, 2, 3, 1]);
        let value_dict_array = DictionaryArray::new(value_dict_keys, values.clone());

        let key_dict_keys = Int8Array::from_iter_values([0, 0, 2, 2, 2, 3]);
        let key_dict_array = DictionaryArray::new(key_dict_keys, values);

        #[expect(deprecated)]
        let keys_field = Arc::new(Field::new_dict(
            Field::MAP_KEY_FIELD_DEFAULT_NAME,
            DataType::Dictionary(Box::new(DataType::Int8), Box::new(DataType::Utf8)),
            false,
            1,
            false,
        ));
        #[expect(deprecated)]
        let values_field = Arc::new(Field::new_dict(
            Field::MAP_VALUE_FIELD_DEFAULT_NAME,
            DataType::Dictionary(Box::new(DataType::Int8), Box::new(DataType::Utf8)),
            true,
            2,
            false,
        ));
        let entry_struct = StructArray::from(vec![
            (keys_field, make_array(key_dict_array.into_data())),
            (values_field, make_array(value_dict_array.into_data())),
        ]);
        let map_data_type = DataType::Map(
            Arc::new(Field::new(
                Field::MAP_ENTRIES_FIELD_DEFAULT_NAME,
                entry_struct.data_type().clone(),
                false,
            )),
            false,
        );

        let entry_offsets = Buffer::from_slice_ref([0, 2, 4, 6]);
        let map_data = ArrayData::builder(map_data_type)
            .len(3)
            .add_buffer(entry_offsets)
            .add_child_data(entry_struct.into_data())
            .build()
            .unwrap();
        let map_array = MapArray::from(map_data);

        let dict_keys = Int8Array::from_iter_values([0, 1, 1, 2, 2, 1]);
        let dict_dict_array = DictionaryArray::new(dict_keys, Arc::new(map_array));

        let schema = Arc::new(Schema::new(vec![Field::new(
            "f1",
            dict_dict_array.data_type().clone(),
            false,
        )]));
        let input_batch = RecordBatch::try_new(schema, vec![Arc::new(dict_dict_array)]).unwrap();
        let output_batch = roundtrip_ipc_stream(&input_batch);
        assert_eq!(input_batch, output_batch);
    }

    fn test_roundtrip_stream_dict_of_list_of_dict_impl<
        OffsetSize: OffsetSizeTrait,
        U: ArrowNativeType,
    >(
        list_data_type: DataType,
        offsets: &[U; 5],
    ) {
        let values = StringArray::from(vec![Some("a"), None, Some("c"), None]);
        let keys = Int8Array::from_iter_values([0, 0, 1, 2, 0, 1, 3]);
        let dict_array = DictionaryArray::new(keys, Arc::new(values));
        let dict_data = dict_array.to_data();

        let value_offsets = Buffer::from_slice_ref(offsets);

        let list_data = ArrayData::builder(list_data_type)
            .len(4)
            .add_buffer(value_offsets)
            .add_child_data(dict_data)
            .build()
            .unwrap();
        let list_array = GenericListArray::<OffsetSize>::from(list_data);

        let keys_for_dict = Int8Array::from_iter_values([0, 3, 0, 1, 1, 2, 0, 1, 3]);
        let dict_dict_array = DictionaryArray::new(keys_for_dict, Arc::new(list_array));

        let schema = Arc::new(Schema::new(vec![Field::new(
            "f1",
            dict_dict_array.data_type().clone(),
            false,
        )]));
        let input_batch = RecordBatch::try_new(schema, vec![Arc::new(dict_dict_array)]).unwrap();
        let output_batch = roundtrip_ipc_stream(&input_batch);
        assert_eq!(input_batch, output_batch);
    }

    #[test]
    fn test_roundtrip_stream_dict_of_list_of_dict() {
        // list
        #[expect(deprecated)]
        let list_data_type = DataType::List(Arc::new(Field::new_dict(
            "item",
            DataType::Dictionary(Box::new(DataType::Int8), Box::new(DataType::Utf8)),
            true,
            1,
            false,
        )));
        let offsets: &[i32; 5] = &[0, 2, 4, 4, 6];
        test_roundtrip_stream_dict_of_list_of_dict_impl::<i32, i32>(list_data_type, offsets);

        // large list
        #[expect(deprecated)]
        let list_data_type = DataType::LargeList(Arc::new(Field::new_dict(
            "item",
            DataType::Dictionary(Box::new(DataType::Int8), Box::new(DataType::Utf8)),
            true,
            1,
            false,
        )));
        let offsets: &[i64; 5] = &[0, 2, 4, 4, 7];
        test_roundtrip_stream_dict_of_list_of_dict_impl::<i64, i64>(list_data_type, offsets);
    }

    #[test]
    fn test_roundtrip_stream_dict_of_fixed_size_list_of_dict() {
        let values = StringArray::from(vec![Some("a"), None, Some("c"), None]);
        let keys = Int8Array::from_iter_values([0, 0, 1, 2, 0, 1, 3, 1, 2]);
        let dict_array = DictionaryArray::new(keys, Arc::new(values));
        let dict_data = dict_array.into_data();

        #[expect(deprecated)]
        let list_data_type = DataType::FixedSizeList(
            Arc::new(Field::new_dict(
                "item",
                DataType::Dictionary(Box::new(DataType::Int8), Box::new(DataType::Utf8)),
                true,
                1,
                false,
            )),
            3,
        );
        let list_data = ArrayData::builder(list_data_type)
            .len(3)
            .add_child_data(dict_data)
            .build()
            .unwrap();
        let list_array = FixedSizeListArray::from(list_data);

        let keys_for_dict = Int8Array::from_iter_values([0, 1, 0, 1, 1, 2, 0, 1, 2]);
        let dict_dict_array = DictionaryArray::new(keys_for_dict, Arc::new(list_array));

        let schema = Arc::new(Schema::new(vec![Field::new(
            "f1",
            dict_dict_array.data_type().clone(),
            false,
        )]));
        let input_batch = RecordBatch::try_new(schema, vec![Arc::new(dict_dict_array)]).unwrap();
        let output_batch = roundtrip_ipc_stream(&input_batch);
        assert_eq!(input_batch, output_batch);
    }

    const LONG_TEST_STRING: &str =
        "This is a long string to make sure binary view array handles it";

    #[test]
    fn test_roundtrip_view_types() {
        let schema = Schema::new(vec![
            Field::new("field_1", DataType::BinaryView, true),
            Field::new("field_2", DataType::Utf8, true),
            Field::new("field_3", DataType::Utf8View, true),
        ]);
        let bin_values: Vec<Option<&[u8]>> = vec![
            Some(b"foo"),
            None,
            Some(b"bar"),
            Some(LONG_TEST_STRING.as_bytes()),
        ];
        let utf8_values: Vec<Option<&str>> =
            vec![Some("foo"), None, Some("bar"), Some(LONG_TEST_STRING)];
        let bin_view_array = BinaryViewArray::from_iter(bin_values);
        let utf8_array = StringArray::from_iter(utf8_values.iter());
        let utf8_view_array = StringViewArray::from_iter(utf8_values);
        let record_batch = RecordBatch::try_new(
            Arc::new(schema.clone()),
            vec![
                Arc::new(bin_view_array),
                Arc::new(utf8_array),
                Arc::new(utf8_view_array),
            ],
        )
        .unwrap();

        assert_eq!(record_batch, roundtrip_ipc(&record_batch));
        assert_eq!(record_batch, roundtrip_ipc_stream(&record_batch));

        let sliced_batch = record_batch.slice(1, 2);
        assert_eq!(sliced_batch, roundtrip_ipc(&sliced_batch));
        assert_eq!(sliced_batch, roundtrip_ipc_stream(&sliced_batch));
    }

    #[test]
    fn test_roundtrip_view_types_nested_dict() {
        let bin_values: Vec<Option<&[u8]>> = vec![
            Some(b"foo"),
            None,
            Some(b"bar"),
            Some(LONG_TEST_STRING.as_bytes()),
            Some(b"field"),
        ];
        let utf8_values: Vec<Option<&str>> = vec![
            Some("foo"),
            None,
            Some("bar"),
            Some(LONG_TEST_STRING),
            Some("field"),
        ];
        let bin_view_array = Arc::new(BinaryViewArray::from_iter(bin_values));
        let utf8_view_array = Arc::new(StringViewArray::from_iter(utf8_values));

        let key_dict_keys = Int8Array::from_iter_values([0, 0, 2, 2, 0, 2, 3]);
        let key_dict_array = DictionaryArray::new(key_dict_keys, utf8_view_array.clone());
        #[expect(deprecated)]
        let keys_field = Arc::new(Field::new_dict(
            Field::MAP_KEY_FIELD_DEFAULT_NAME,
            DataType::Dictionary(Box::new(DataType::Int8), Box::new(DataType::Utf8View)),
            false,
            1,
            false,
        ));

        let value_dict_keys = Int8Array::from_iter_values([0, 3, 0, 1, 2, 0, 1]);
        let value_dict_array = DictionaryArray::new(value_dict_keys, bin_view_array);
        #[expect(deprecated)]
        let values_field = Arc::new(Field::new_dict(
            Field::MAP_VALUE_FIELD_DEFAULT_NAME,
            DataType::Dictionary(Box::new(DataType::Int8), Box::new(DataType::BinaryView)),
            true,
            2,
            false,
        ));
        let entry_struct = StructArray::from(vec![
            (keys_field, make_array(key_dict_array.into_data())),
            (values_field, make_array(value_dict_array.into_data())),
        ]);

        let map_data_type = DataType::Map(
            Arc::new(Field::new(
                Field::MAP_ENTRIES_FIELD_DEFAULT_NAME,
                entry_struct.data_type().clone(),
                false,
            )),
            false,
        );
        let entry_offsets = Buffer::from_slice_ref([0, 2, 4, 7]);
        let map_data = ArrayData::builder(map_data_type)
            .len(3)
            .add_buffer(entry_offsets)
            .add_child_data(entry_struct.into_data())
            .build()
            .unwrap();
        let map_array = MapArray::from(map_data);

        let dict_keys = Int8Array::from_iter_values([0, 1, 0, 1, 1, 2, 0, 1, 2]);
        let dict_dict_array = DictionaryArray::new(dict_keys, Arc::new(map_array));
        let schema = Arc::new(Schema::new(vec![Field::new(
            "f1",
            dict_dict_array.data_type().clone(),
            false,
        )]));
        let batch = RecordBatch::try_new(schema, vec![Arc::new(dict_dict_array)]).unwrap();
        assert_eq!(batch, roundtrip_ipc(&batch));
        assert_eq!(batch, roundtrip_ipc_stream(&batch));

        let sliced_batch = batch.slice(1, 2);
        assert_eq!(sliced_batch, roundtrip_ipc(&sliced_batch));
        assert_eq!(sliced_batch, roundtrip_ipc_stream(&sliced_batch));
    }

    #[test]
    fn test_no_columns_batch() {
        let schema = Arc::new(Schema::empty());
        let options = RecordBatchOptions::new()
            .with_match_field_names(true)
            .with_row_count(Some(10));
        let input_batch = RecordBatch::try_new_with_options(schema, vec![], &options).unwrap();
        let output_batch = roundtrip_ipc_stream(&input_batch);
        assert_eq!(input_batch, output_batch);
    }

    #[test]
    fn test_unaligned() {
        let batch = RecordBatch::try_from_iter(vec![(
            "i32",
            Arc::new(Int32Array::from(vec![1, 2, 3, 4])) as _,
        )])
        .unwrap();

        let r#gen = IpcDataGenerator {};
        let mut dict_tracker = DictionaryTracker::new(false);
        let (_, encoded) = r#gen
            .encode(
                &batch,
                &mut dict_tracker,
                &Default::default(),
                &mut Default::default(),
            )
            .unwrap();

        let message = root_as_message(&encoded.ipc_message).unwrap();

        // Construct an unaligned buffer
        let mut buffer = MutableBuffer::with_capacity(encoded.arrow_data.len() + 1);
        buffer.push(0_u8);
        buffer.extend_from_slice(&encoded.arrow_data);
        let b = Buffer::from(buffer).slice(1);
        assert_ne!(b.as_ptr().align_offset(8), 0);

        let ipc_batch = message.header_as_record_batch().unwrap();
        let roundtrip = RecordBatchDecoder::try_new(
            &b,
            ipc_batch,
            batch.schema(),
            &Default::default(),
            &message.version(),
        )
        .unwrap()
        .with_require_alignment(false)
        .read_record_batch()
        .unwrap();
        assert_eq!(batch, roundtrip);
    }

    #[test]
    fn test_unaligned_throws_error_with_require_alignment() {
        let batch = RecordBatch::try_from_iter(vec![(
            "i32",
            Arc::new(Int32Array::from(vec![1, 2, 3, 4])) as _,
        )])
        .unwrap();

        let r#gen = IpcDataGenerator {};
        let mut dict_tracker = DictionaryTracker::new(false);
        let (_, encoded) = r#gen
            .encode(
                &batch,
                &mut dict_tracker,
                &Default::default(),
                &mut Default::default(),
            )
            .unwrap();

        let message = root_as_message(&encoded.ipc_message).unwrap();

        // Construct an unaligned buffer
        let mut buffer = MutableBuffer::with_capacity(encoded.arrow_data.len() + 1);
        buffer.push(0_u8);
        buffer.extend_from_slice(&encoded.arrow_data);
        let b = Buffer::from(buffer).slice(1);
        assert_ne!(b.as_ptr().align_offset(8), 0);

        let ipc_batch = message.header_as_record_batch().unwrap();
        let result = RecordBatchDecoder::try_new(
            &b,
            ipc_batch,
            batch.schema(),
            &Default::default(),
            &message.version(),
        )
        .unwrap()
        .with_require_alignment(true)
        .read_record_batch();

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

    /// Verify that misaligned IPC buffers are caught by `require_alignment = true`.
    ///
    /// For each array type we shift the IPC body by every byte offset in 0..OFFSET_RANGE and
    /// assert that the decoder errors exactly when the offset violates the type's
    /// minimum alignment requirement.  The Arrow columnar spec permits alignment to
    /// any multiple of 8 or 64 bytes, so multiples of 8 (which include every multiple
    /// of 64) must succeed; everything else must return a "Misaligned buffers" error.
    /// See <https://arrow.apache.org/docs/format/Columnar.html#buffer-alignment-and-padding>.
    #[test]
    fn test_misaligned_buffers_error() {
        const OFFSET_RANGE: usize = 128;

        // (array, minimum required alignment in bytes)
        // Fixed-width: alignment == element byte width.
        // Variable-width (e.g. StringArray): the offsets buffer drives alignment (Int32 → 4).
        let cases: Vec<(ArrayRef, usize)> = vec![
            (Arc::new(Int32Array::from_iter(0i32..100)) as _, 4),
            (Arc::new(Int64Array::from_iter(0i64..100)) as _, 8),
            (
                Arc::new(StringArray::from_iter_values(
                    (0..100).map(|i| i.to_string()),
                )) as _,
                4,
            ),
            (
                Arc::new(LargeStringArray::from_iter_values(
                    (0..100).map(|i| i.to_string()),
                )) as _,
                8,
            ),
        ];

        for (array, alignment) in cases {
            let batch = RecordBatch::try_from_iter(vec![("col", Arc::clone(&array))]).unwrap();
            let encoder = IpcDataGenerator {};
            let mut dict_tracker = DictionaryTracker::new(false);
            let (_, encoded) = encoder
                .encode(
                    &batch,
                    &mut dict_tracker,
                    &Default::default(),
                    &mut Default::default(),
                )
                .unwrap();
            let message = root_as_message(&encoded.ipc_message).unwrap();
            let ipc_batch = message.header_as_record_batch().unwrap();

            for offset in 0..OFFSET_RANGE {
                // MutableBuffer always allocates at a 64-byte aligned base address.
                // Slicing by `offset` bytes yields a pointer at `base_ptr + offset`,
                // whose alignment is gcd(64, offset).  That makes `offset` the sole
                // determinant of the resulting alignment — without this guarantee the
                // test would be non-deterministic depending on what the allocator returns.
                let mut storage = MutableBuffer::with_capacity(encoded.arrow_data.len() + offset);
                for _ in 0..offset {
                    storage.push(0_u8);
                }
                storage.extend_from_slice(&encoded.arrow_data);
                let buf = Buffer::from(storage).slice(offset);

                let result = RecordBatchDecoder::try_new(
                    &buf,
                    ipc_batch,
                    batch.schema(),
                    &Default::default(),
                    &message.version(),
                )
                .unwrap()
                .with_require_alignment(true)
                .read_record_batch();

                if offset % alignment == 0 {
                    assert!(
                        result.is_ok(),
                        "type={} offset={offset}: expected Ok but got {:?}",
                        array.data_type(),
                        result.unwrap_err(),
                    );
                } else {
                    let err = result
                        .expect_err(&format!(
                            "type={} offset={offset}: expected Err for misaligned buffer",
                            array.data_type()
                        ))
                        .to_string();
                    assert!(
                        err.contains("Misaligned buffers"),
                        "type={} offset={offset}: unexpected error: {err}",
                        array.data_type(),
                    );
                }
            }
        }
    }

    #[test]
    #[cfg_attr(miri, ignore)] // Takes too long
    fn test_file_with_massive_column_count() {
        // 499_999 is upper limit for default settings (1_000_000)
        let limit = 600_000;

        let fields = (0..limit)
            .map(|i| Field::new(format!("{i}"), DataType::Boolean, false))
            .collect::<Vec<_>>();
        let schema = Arc::new(Schema::new(fields));
        let batch = RecordBatch::new_empty(schema);

        let mut buf = Vec::new();
        let mut writer = crate::writer::FileWriter::try_new(&mut buf, batch.schema_ref()).unwrap();
        writer.write(&batch).unwrap();
        writer.finish().unwrap();
        drop(writer);

        let mut reader = FileReaderBuilder::new()
            .with_max_footer_fb_tables(1_500_000)
            .build(std::io::Cursor::new(buf))
            .unwrap();
        let roundtrip_batch = reader.next().unwrap().unwrap();

        assert_eq!(batch, roundtrip_batch);
    }

    #[test]
    fn test_file_with_deeply_nested_columns() {
        // 60 is upper limit for default settings (64)
        let limit = 61;

        let fields = (0..limit).fold(
            vec![Field::new("leaf", DataType::Boolean, false)],
            |field, index| vec![Field::new_struct(format!("{index}"), field, false)],
        );
        let schema = Arc::new(Schema::new(fields));
        let batch = RecordBatch::new_empty(schema);

        let mut buf = Vec::new();
        let mut writer = crate::writer::FileWriter::try_new(&mut buf, batch.schema_ref()).unwrap();
        writer.write(&batch).unwrap();
        writer.finish().unwrap();
        drop(writer);

        let mut reader = FileReaderBuilder::new()
            .with_max_footer_fb_depth(65)
            .build(std::io::Cursor::new(buf))
            .unwrap();
        let roundtrip_batch = reader.next().unwrap().unwrap();

        assert_eq!(batch, roundtrip_batch);
    }

    #[test]
    fn test_stream_reader_rejects_short_validity_buffer() {
        // Reproduce #7124: serialize an Int32Array with too few validity bits.
        let data = ArrayDataBuilder::new(DataType::Int32)
            .len(8000)
            .add_buffer(ScalarBuffer::<i32>::from_iter(0..8000).into())
            .nulls(Some(NullBuffer::from(&[true, false, true, false])));
        let array: ArrayRef = unsafe { Arc::new(Int32Array::from(data.build_unchecked())) };
        let batch = RecordBatch::try_from_iter([("a", array)]).unwrap();

        let mut stream = Vec::new();
        let mut writer =
            crate::writer::StreamWriter::try_new(&mut stream, &batch.schema()).unwrap();
        writer.write(&batch).unwrap();
        writer.finish().unwrap();

        let mut reader = StreamReader::try_new(Cursor::new(stream), None).unwrap();
        let err = reader.next().unwrap().unwrap_err();
        assert_eq!(
            err.to_string(),
            "Invalid argument error: null_bit_buffer size too small. got 1 needed 1000"
        );
    }

    #[test]
    fn test_stream_reader_rejects_short_struct_validity_buffer() {
        // Serialize a StructArray with too few validity bits.
        let fields = Fields::from(vec![Field::new("b", DataType::Int32, false)]);
        let child: ArrayRef = Arc::new(Int32Array::from_iter_values(0..8000));
        let nulls = NullBuffer::from(&[true, false, true, false]);
        let array: ArrayRef =
            unsafe { Arc::new(StructArray::new_unchecked(fields, vec![child], Some(nulls))) };
        let batch = RecordBatch::try_from_iter([("a", array)]).unwrap();

        let mut stream = Vec::new();
        let mut writer =
            crate::writer::StreamWriter::try_new(&mut stream, &batch.schema()).unwrap();
        writer.write(&batch).unwrap();
        writer.finish().unwrap();

        let mut reader = StreamReader::try_new(Cursor::new(stream), None).unwrap();
        let err = reader.next().unwrap().unwrap_err();
        assert_eq!(
            err.to_string(),
            "Invalid argument error: null_bit_buffer size too small. got 1 needed 1000"
        );
    }

    #[test]
    fn test_invalid_struct_array_ipc_read_errors() {
        let a_field = Field::new("a", DataType::Int32, false);
        let b_field = Field::new("b", DataType::Int32, false);
        let struct_fields = Fields::from(vec![a_field.clone(), b_field.clone()]);

        let a_array_data = ArrayData::builder(a_field.data_type().clone())
            .len(4)
            .add_buffer(Buffer::from_slice_ref([1, 2, 3, 4]))
            .build()
            .unwrap();
        let b_array_data = ArrayData::builder(b_field.data_type().clone())
            .len(3)
            .add_buffer(Buffer::from_slice_ref([5, 6, 7]))
            .build()
            .unwrap();

        let invalid_struct_arr = unsafe {
            StructArray::new_unchecked(
                struct_fields,
                vec![make_array(a_array_data), make_array(b_array_data)],
                None,
            )
        };

        expect_ipc_validation_error(
            Arc::new(invalid_struct_arr),
            "Invalid argument error: Incorrect array length for StructArray field \"b\", expected 4 got 3",
        );
    }

    #[test]
    fn test_invalid_nested_array_ipc_read_errors() {
        // one of the nested arrays has invalid data
        let a_field = Field::new("a", DataType::Int32, false);
        let b_field = Field::new("b", DataType::Utf8, false);

        let schema = Arc::new(Schema::new(vec![Field::new_struct(
            "s",
            vec![a_field.clone(), b_field.clone()],
            false,
        )]));

        let a_array_data = ArrayData::builder(a_field.data_type().clone())
            .len(4)
            .add_buffer(Buffer::from_slice_ref([1, 2, 3, 4]))
            .build()
            .unwrap();
        // invalid nested child array -- length is correct, but has invalid utf8 data
        let b_array_data = {
            let valid: &[u8] = b"   ";
            let mut invalid = vec![];
            invalid.extend_from_slice(b"ValidString");
            invalid.extend_from_slice(INVALID_UTF8_FIRST_CHAR);
            let binary_array =
                BinaryArray::from_iter(vec![None, Some(valid), None, Some(&invalid)]);
            let array = unsafe {
                StringArray::new_unchecked(
                    binary_array.offsets().clone(),
                    binary_array.values().clone(),
                    binary_array.nulls().cloned(),
                )
            };
            array.into_data()
        };
        let struct_data_type = schema.field(0).data_type();

        let invalid_struct_arr = unsafe {
            make_array(
                ArrayData::builder(struct_data_type.clone())
                    .len(4)
                    .add_child_data(a_array_data)
                    .add_child_data(b_array_data)
                    .build_unchecked(),
            )
        };
        expect_ipc_validation_error(
            invalid_struct_arr,
            "Invalid argument error: Invalid UTF8 sequence at string index 3 (3..18): invalid utf-8 sequence of 1 bytes from index 11",
        );
    }

    #[test]
    #[expect(deprecated)]
    fn test_first_field_with_dict_id_matches_fields_with_dict_id() {
        let dict = |name: &str, id: i64| {
            let data_type =
                DataType::Dictionary(Box::new(DataType::Int32), Box::new(DataType::Utf8));
            Arc::new(Field::new_dict(name, data_type, true, id, false))
        };
        let dict_of_list_of_dict = |name: &str, outer_id: i64, inner_id: i64| {
            let inner = DataType::List(dict("inner", inner_id));
            let data_type = DataType::Dictionary(Box::new(DataType::Int32), Box::new(inner));
            Field::new_dict(name, data_type, true, outer_id, false)
        };
        // Duplicate IDs check that the first match in pre-order wins.
        let schema = Schema::new(vec![
            // dict_id is ignored on non-dictionary fields.
            Field::new_dict("not_a_dictionary", DataType::Utf8, true, 3, false),
            Field::new_struct(
                "struct",
                vec![dict("struct.a", 1), dict("struct.b", 2)],
                true,
            ),
            Field::new(
                "union",
                DataType::Union(
                    UnionFields::try_new(vec![0, 1], vec![dict("union.a", 3), dict("union.b", 1)])
                        .unwrap(),
                    UnionMode::Dense,
                ),
                true,
            ),
            Field::new("list", DataType::List(dict("list.item", 4)), true),
            Field::new_map(
                "map",
                "entries",
                dict("key", 5),
                dict("value", 2),
                false,
                true,
            ),
            Field::new(
                "ree",
                DataType::RunEndEncoded(
                    Arc::new(Field::new("run_ends", DataType::Int32, false)),
                    dict("ree.values", 6),
                ),
                true,
            ),
            dict_of_list_of_dict("dict_of_list_of_dict", 7, 8),
            dict_of_list_of_dict("same_outer_and_inner_id", 9, 9),
        ]);

        for id in -1..=10 {
            let expected = schema.fields_with_dict_id(id).first().map(|f| f.name());
            let actual = first_field_with_dict_id(schema.fields(), id).map(|f| f.name());
            assert_eq!(actual, expected, "dictionary id {id}");
        }
    }

    #[test]
    fn test_same_dict_id_without_preserve() {
        let batch = RecordBatch::try_new(
            Arc::new(Schema::new(
                ["a", "b"]
                    .iter()
                    .map(|name| {
                        #[expect(deprecated)]
                        Field::new_dict(
                            name.to_string(),
                            DataType::Dictionary(
                                Box::new(DataType::Int32),
                                Box::new(DataType::Utf8),
                            ),
                            true,
                            0,
                            false,
                        )
                    })
                    .collect::<Vec<Field>>(),
            )),
            vec![
                Arc::new(
                    vec![Some("c"), Some("d")]
                        .into_iter()
                        .collect::<DictionaryArray<Int32Type>>(),
                ) as ArrayRef,
                Arc::new(
                    vec![Some("e"), Some("f")]
                        .into_iter()
                        .collect::<DictionaryArray<Int32Type>>(),
                ) as ArrayRef,
            ],
        )
        .expect("Failed to create RecordBatch");

        // serialize the record batch as an IPC stream
        let mut buf = vec![];
        {
            let mut writer = crate::writer::StreamWriter::try_new_with_options(
                &mut buf,
                batch.schema().as_ref(),
                crate::writer::IpcWriteOptions::default(),
            )
            .expect("Failed to create StreamWriter");
            writer.write(&batch).expect("Failed to write RecordBatch");
            writer.finish().expect("Failed to finish StreamWriter");
        }

        StreamReader::try_new(std::io::Cursor::new(buf), None)
            .expect("Failed to create StreamReader")
            .for_each(|decoded_batch| {
                assert_eq!(decoded_batch.expect("Failed to read RecordBatch"), batch);
            });
    }

    #[test]
    fn test_validation_of_invalid_list_array() {
        // ListArray with invalid offsets
        let array = unsafe {
            let values = Int32Array::from(vec![1, 2, 3]);
            let bad_offsets = ScalarBuffer::<i32>::from(vec![0, 2, 4, 2]); // offsets can't go backwards
            let offsets = OffsetBuffer::new_unchecked(bad_offsets); // INVALID array created
            let field = Field::new_list_field(DataType::Int32, true);
            let nulls = None;
            ListArray::new(Arc::new(field), offsets, Arc::new(values), nulls)
        };

        expect_ipc_validation_error(
            Arc::new(array),
            "Invalid argument error: Offset invariant failure: offset at position 2 out of bounds: 4 > 2",
        );
    }

    #[test]
    fn test_validation_of_invalid_string_array() {
        let valid: &[u8] = b"   ";
        let mut invalid = vec![];
        invalid.extend_from_slice(b"ThisStringIsCertainlyLongerThan12Bytes");
        invalid.extend_from_slice(INVALID_UTF8_FIRST_CHAR);
        let binary_array = BinaryArray::from_iter(vec![None, Some(valid), None, Some(&invalid)]);
        // data is not valid utf8 we can not construct a correct StringArray
        // safely, so purposely create an invalid StringArray
        let array = unsafe {
            StringArray::new_unchecked(
                binary_array.offsets().clone(),
                binary_array.values().clone(),
                binary_array.nulls().cloned(),
            )
        };
        expect_ipc_validation_error(
            Arc::new(array),
            "Invalid argument error: Invalid UTF8 sequence at string index 3 (3..45): invalid utf-8 sequence of 1 bytes from index 38",
        );
    }

    #[test]
    fn test_validation_of_invalid_string_view_array() {
        let valid: &[u8] = b"   ";
        let mut invalid = vec![];
        invalid.extend_from_slice(b"ThisStringIsCertainlyLongerThan12Bytes");
        invalid.extend_from_slice(INVALID_UTF8_FIRST_CHAR);
        let binary_view_array =
            BinaryViewArray::from_iter(vec![None, Some(valid), None, Some(&invalid)]);
        // data is not valid utf8 we can not construct a correct StringArray
        // safely, so purposely create an invalid StringArray
        let array = unsafe {
            StringViewArray::new_unchecked(
                binary_view_array.views().clone(),
                Arc::clone(binary_view_array.data_buffers()),
                binary_view_array.nulls().cloned(),
            )
        };
        expect_ipc_validation_error(
            Arc::new(array),
            "Invalid argument error: Encountered non-UTF-8 data at index 3: invalid utf-8 sequence of 1 bytes from index 38",
        );
    }

    /// return an invalid dictionary array (key is larger than values)
    /// ListArray with invalid offsets
    #[test]
    fn test_validation_of_invalid_dictionary_array() {
        let array = unsafe {
            let values = StringArray::from_iter_values(["a", "b", "c"]);
            let keys = Int32Array::from(vec![1, 200]); // keys are not valid for values
            DictionaryArray::new_unchecked(keys, Arc::new(values))
        };

        expect_ipc_validation_error(
            Arc::new(array),
            "Invalid argument error: Value at position 1 out of bounds: 200 (should be in [0, 2])",
        );
    }

    #[test]
    fn test_validation_of_invalid_union_array() {
        let array = unsafe {
            let fields = UnionFields::try_new(
                vec![1, 3], // typeids : type id 2 is not valid
                vec![
                    Field::new("a", DataType::Int32, false),
                    Field::new("b", DataType::Utf8, false),
                ],
            )
            .unwrap();
            let type_ids = ScalarBuffer::from(vec![1i8, 2, 3]); // 2 is invalid
            let offsets = None;
            let children: Vec<ArrayRef> = vec![
                Arc::new(Int32Array::from(vec![10, 20, 30])),
                Arc::new(StringArray::from(vec![Some("a"), Some("b"), Some("c")])),
            ];

            UnionArray::new_unchecked(fields, type_ids, offsets, children)
        };

        expect_ipc_validation_error(
            Arc::new(array),
            "Invalid argument error: Type Ids values must match one of the field type ids",
        );
    }

    /// Invalid Utf-8 sequence in the first character
    /// <https://stackoverflow.com/questions/1301402/example-invalid-utf8-string>
    const INVALID_UTF8_FIRST_CHAR: &[u8] = &[0xa0, 0xa1, 0x20, 0x20];

    /// Expect an error when reading the record batch using IPC or IPC Streams
    fn expect_ipc_validation_error(array: ArrayRef, expected_err: &str) {
        let rb = RecordBatch::try_from_iter([("a", array)]).unwrap();

        // IPC Stream format
        let buf = write_stream(&rb); // write is ok
        read_stream_skip_validation(&buf).unwrap();
        let err = read_stream(&buf).unwrap_err();
        assert_eq!(err.to_string(), expected_err);

        // IPC Stream format with StreamDecoder
        read_stream_with_decoder(&buf, true).unwrap();
        let err = read_stream_with_decoder(&buf, false).unwrap_err();
        assert_eq!(err.to_string(), expected_err);

        // IPC File format
        let buf = write_ipc(&rb); // write is ok
        read_ipc_skip_validation(&buf).unwrap();
        let err = read_ipc(&buf).unwrap_err();
        assert_eq!(err.to_string(), expected_err);

        // IPC Format with FileDecoder
        read_ipc_with_decoder_skip_validation(buf.clone()).unwrap();
        let err = read_ipc_with_decoder(buf).unwrap_err();
        assert_eq!(err.to_string(), expected_err);
    }

    #[test]
    fn test_roundtrip_schema() {
        let schema = Schema::new(vec![
            Field::new(
                "a",
                DataType::Dictionary(Box::new(DataType::UInt16), Box::new(DataType::Utf8)),
                false,
            ),
            Field::new(
                "b",
                DataType::Dictionary(Box::new(DataType::UInt16), Box::new(DataType::Utf8)),
                false,
            ),
        ]);

        let options = IpcWriteOptions::default();
        let data_gen = IpcDataGenerator::default();
        let mut dict_tracker = DictionaryTracker::new(false);
        let encoded_data =
            data_gen.schema_to_bytes_with_dictionary_tracker(&schema, &mut dict_tracker, &options);
        let mut schema_bytes = vec![];
        write_message(&mut schema_bytes, encoded_data, &options).expect("write_message");

        let begin_offset: usize = if schema_bytes[0..4].eq(&CONTINUATION_MARKER) {
            4
        } else {
            0
        };

        size_prefixed_root_as_message(&schema_bytes[begin_offset..])
            .expect_err("size_prefixed_root_as_message");

        let msg = parse_message(&schema_bytes).expect("parse_message");
        let ipc_schema = msg.header_as_schema().expect("header_as_schema");
        let new_schema = try_fb_to_schema(ipc_schema).unwrap();

        assert_eq!(schema, new_schema);
    }

    #[test]
    fn test_negative_meta_len() {
        let bytes = i32::to_le_bytes(-1);
        let mut buf = vec![];
        buf.extend(CONTINUATION_MARKER);
        buf.extend(bytes);

        let reader = StreamReader::try_new(Cursor::new(buf), None);
        assert!(reader.is_err());
    }

    /// Per the IPC specification, dictionary batches may be omitted for
    /// dictionary-encoded columns where all values are null.  The C++
    /// implementation relies on this and does not emit a dictionary batch
    /// in that case.  Verify that the Rust reader handles such streams
    /// by synthesizing an empty dictionary instead of returning an error.
    #[test]
    fn test_read_null_dict_without_dictionary_batch() {
        // Build an all-null dictionary-encoded column.
        let keys = Int32Array::new_null(4);
        let values: ArrayRef = new_empty_array(&DataType::Utf8);
        let dict_array = DictionaryArray::new(keys, values);

        let schema = Arc::new(Schema::new(vec![Field::new(
            "d",
            dict_array.data_type().clone(),
            true,
        )]));
        let batch = RecordBatch::try_new(schema.clone(), vec![Arc::new(dict_array)]).unwrap();

        // Write a normal IPC stream (which includes the dictionary batch).
        let full_stream = write_stream(&batch);

        // Parse the stream into individual messages and reconstruct it
        // without the DictionaryBatch message, simulating what C++ emits
        // for an all-null dictionary column.
        let mut stripped = Vec::new();
        let mut cursor = Cursor::new(&full_stream);
        loop {
            // Each message is: [continuation (4 bytes)] [meta_len (4 bytes)]
            //                   [metadata (meta_len bytes)] [body (bodyLength bytes)]
            let mut header = [0u8; 4];
            if cursor.read_exact(&mut header).is_err() {
                break;
            }
            if header == CONTINUATION_MARKER && cursor.read_exact(&mut header).is_err() {
                break;
            }
            let meta_len = u32::from_le_bytes(header) as usize;
            if meta_len == 0 {
                // EOS marker — write it through.
                stripped.extend_from_slice(&CONTINUATION_MARKER);
                stripped.extend_from_slice(&0u32.to_le_bytes());
                break;
            }
            let mut meta_buf = vec![0u8; meta_len];
            cursor.read_exact(&mut meta_buf).unwrap();

            let message = root_as_message(&meta_buf).unwrap();
            let body_len = message.bodyLength() as usize;
            let mut body_buf = vec![0u8; body_len];
            cursor.read_exact(&mut body_buf).unwrap();

            if message.header_type() == crate::MessageHeader::DictionaryBatch {
                // Skip the dictionary batch — this is what C++ does for
                // all-null dictionary columns.
                continue;
            }
            stripped.extend_from_slice(&CONTINUATION_MARKER);
            stripped.extend_from_slice(&(meta_len as u32).to_le_bytes());
            stripped.extend_from_slice(&meta_buf);
            stripped.extend_from_slice(&body_buf);
        }

        // Reading the stripped stream must succeed.
        let result = read_stream(&stripped).unwrap();
        assert_eq!(result.num_rows(), 4);
        assert_eq!(result.num_columns(), 1);

        let col = result.column(0);
        assert_eq!(col.null_count(), 4);
        assert_eq!(col.len(), 4);
        // The result must be a dictionary-typed array.
        assert!(matches!(col.data_type(), DataType::Dictionary(_, _)));
    }

    // Tests projected reads where a ListView column is skipped before another column.
    // This catches cases where skipping the ListView consumes the wrong number of buffers.
    #[test]
    fn test_projection_skip_list_view() {
        use crate::reader::FileReader;
        use crate::writer::FileWriter;
        use arrow_array::{
            GenericListViewArray, Int32Array, RecordBatch,
            builder::{GenericListViewBuilder, UInt32Builder},
        };
        use arrow_schema::{DataType, Field, Schema};
        use std::sync::Arc;

        // Build a small ListView column with a mix of valid and null entries
        let mut builder = GenericListViewBuilder::<i32, _>::new(UInt32Builder::new());

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

        builder.append(false);

        builder.values().append_value(3);
        builder.values().append_value(4);
        builder.append(true);

        let list_view: GenericListViewArray<i32> = builder.finish();

        // Second column with simple values
        let values = Int32Array::from(vec![10, 20, 30]);

        // Schema: first column is ListView, second is Int32
        let schema = Arc::new(Schema::new(vec![
            Field::new("a", list_view.data_type().clone(), true),
            Field::new("b", DataType::Int32, false),
        ]));
        // Create a batch with both columns
        let batch =
            RecordBatch::try_new(schema, vec![Arc::new(list_view), Arc::new(values.clone())])
                .unwrap();

        // Write the batch to IPC
        let mut buf = Vec::new();
        {
            let mut writer = FileWriter::try_new(&mut buf, &batch.schema()).unwrap();
            writer.write(&batch).unwrap();
            writer.finish().unwrap();
        }

        // Skip ListView column and Project only column "b"
        let mut reader = FileReader::try_new(std::io::Cursor::new(buf), Some(vec![1])).unwrap();
        let read_batch = reader.next().unwrap().unwrap();

        // Verify that the projected column is read correctly
        assert_eq!(read_batch.num_columns(), 1);
        assert_eq!(read_batch.column(0).as_ref(), &values);
    }

    // Tests reading a column when a preceding V4 Union column is skipped.
    // V4 Union columns include a null buffer and type ids (and offsets for dense unions).
    #[test]
    fn test_projection_skip_union_v4() {
        use crate::MetadataVersion;
        use crate::reader::FileReader;
        use crate::writer::{FileWriter, IpcWriteOptions};
        use arrow_array::{
            ArrayRef, Int32Array, RecordBatch, builder::UnionBuilder, types::Int32Type,
        };
        use arrow_schema::{DataType, Field, Schema};
        use std::sync::Arc;

        // Build a dense Union column with simple Int32 values
        let mut builder = UnionBuilder::new_dense();
        builder.append::<Int32Type>("a", 1).unwrap();
        builder.append::<Int32Type>("a", 2).unwrap();
        builder.append::<Int32Type>("a", 3).unwrap();
        let union = builder.build().unwrap();

        // Second column with known values to verify correctness after projection
        let values = Int32Array::from(vec![10, 20, 30]);

        // Schema: first column is Union (to be skipped), second is Int32 (to be read)
        let schema = Arc::new(Schema::new(vec![
            Field::new("union", union.data_type().clone(), false),
            Field::new("values", DataType::Int32, false),
        ]));

        // Create a batch containing both columns
        let batch = RecordBatch::try_new(
            schema,
            vec![Arc::new(union) as ArrayRef, Arc::new(values.clone())],
        )
        .unwrap();

        // Write IPC using V4 metadata to trigger Union null buffer behavior
        let mut buf = Vec::new();
        {
            let options = IpcWriteOptions::try_new(8, false, MetadataVersion::V4).unwrap();
            let mut writer =
                FileWriter::try_new_with_options(&mut buf, &batch.schema(), options).unwrap();
            writer.write(&batch).unwrap();
            writer.finish().unwrap();
        }
        // Read only the second column (skip the Union column)
        let mut reader = FileReader::try_new(std::io::Cursor::new(buf), Some(vec![1])).unwrap();
        let read_batch = reader.next().unwrap().unwrap();

        // Verify that the projected column is read correctly after skipping Union
        assert_eq!(read_batch.num_columns(), 1);
        assert_eq!(read_batch.column(0).as_ref(), &values);
    }

    // Tests reading a column when preceding fixed-width and boolean columns are skipped.
    // Covers all types that use the same two-buffer layout (null + values).
    // Verifies that skipping these types does not affect subsequent column decoding.
    #[test]
    fn test_projection_skip_fixed_width_types() {
        use std::sync::Arc;

        use arrow_array::{ArrayRef, BooleanArray, Int32Array, RecordBatch, make_array};
        use arrow_buffer::{Buffer, MutableBuffer};
        use arrow_data::ArrayData;
        use arrow_schema::{DataType, Field, IntervalUnit, Schema, TimeUnit};

        use crate::reader::FileReader;
        use crate::writer::FileWriter;

        // Create a minimal array for a given fixed-width or boolean type
        fn make_array_for_type(data_type: DataType) -> ArrayRef {
            let len = 3;

            if matches!(data_type, DataType::Boolean) {
                return Arc::new(BooleanArray::from(vec![true, false, true]));
            }

            let width = data_type.primitive_width().unwrap();
            let data = ArrayData::builder(data_type)
                .len(len)
                .add_buffer(Buffer::from(MutableBuffer::from_len_zeroed(len * width)))
                .build()
                .unwrap();

            make_array(data)
        }

        // List of types that follow the same two-buffer layout (null + values)
        let data_types = vec![
            DataType::Boolean,
            DataType::Int8,
            DataType::Int16,
            DataType::Int32,
            DataType::Int64,
            DataType::UInt8,
            DataType::UInt16,
            DataType::UInt32,
            DataType::UInt64,
            DataType::Float16,
            DataType::Float32,
            DataType::Float64,
            DataType::Timestamp(TimeUnit::Second, None),
            DataType::Date32,
            DataType::Date64,
            DataType::Time32(TimeUnit::Second),
            DataType::Time64(TimeUnit::Microsecond),
            DataType::Duration(TimeUnit::Second),
            DataType::Interval(IntervalUnit::YearMonth),
            DataType::Interval(IntervalUnit::DayTime),
            DataType::Interval(IntervalUnit::MonthDayNano),
            DataType::Decimal32(9, 2),
            DataType::Decimal64(18, 2),
            DataType::Decimal128(38, 2),
            DataType::Decimal256(76, 2),
        ];

        // For each type:
        // - write a batch with [skipped_column, values]
        // - read only the second column
        // - verify the result is correct
        for data_type in data_types {
            let skipped = make_array_for_type(data_type.clone());
            let values = Int32Array::from(vec![10, 20, 30]);

            let schema = Arc::new(Schema::new(vec![
                Field::new("skipped", data_type, false),
                Field::new("values", DataType::Int32, false),
            ]));

            let batch =
                RecordBatch::try_new(schema, vec![skipped, Arc::new(values.clone())]).unwrap();

            // Serialize the batch into IPC format
            let mut buf = Vec::new();
            {
                let mut writer = FileWriter::try_new(&mut buf, &batch.schema()).unwrap();
                writer.write(&batch).unwrap();
                writer.finish().unwrap();
            }

            // Read back only the second column (skip the first)
            let mut reader = FileReader::try_new(std::io::Cursor::new(buf), Some(vec![1])).unwrap();
            let read_batch = reader.next().unwrap().unwrap();

            // Verify that the returned column matches the original values column
            assert_eq!(read_batch.num_columns(), 1);
            assert_eq!(read_batch.column(0).as_ref(), &values);
        }
    }

    /// Builds a stream whose single message declares `body_length` but is followed by only
    /// `body_bytes` bytes of body. The body length is consumed before the header is
    /// interpreted, so the header itself is immaterial.
    fn stream_with_declared_body(body_length: i64, body_bytes: usize) -> Vec<u8> {
        let mut fbb = flatbuffers::FlatBufferBuilder::new();
        let mut message = crate::MessageBuilder::new(&mut fbb);
        message.add_version(crate::MetadataVersion::V5);
        message.add_header_type(crate::MessageHeader::NONE);
        message.add_bodyLength(body_length);
        let root = message.finish();
        fbb.finish(root, None);
        let metadata = fbb.finished_data();

        let mut stream = Vec::new();
        stream.extend_from_slice(&CONTINUATION_MARKER);
        stream.extend_from_slice(&(metadata.len() as i32).to_le_bytes());
        stream.extend_from_slice(metadata);
        stream.resize(stream.len() + body_bytes, 0xAB);
        stream
    }

    /// A message body length is read from the stream before any of the bytes it describes,
    /// so it cannot be trusted. Declaring an implausible one used to reserve it outright,
    /// and the resulting allocation failure aborts the process instead of surfacing an error
    /// the caller can handle.
    #[test]
    fn test_stream_reader_rejects_implausible_body_length() {
        for body_length in [i64::MAX, 1 << 50, -1] {
            let stream = stream_with_declared_body(body_length, 0);
            let err = StreamReader::try_new(std::io::Cursor::new(stream), None).expect_err(
                &format!("a message declaring {body_length} body bytes must not be accepted"),
            );
            let err = err.to_string();
            assert!(
                err.contains("Invalid IPC message body length")
                    || err.contains("Unexpected end of stream")
                    || err.contains("failed to fill whole buffer"),
                "unexpected error for body_length {body_length}: {err}"
            );
        }
    }

    /// A plausible body length whose bytes end early must fail as a short read: before the
    /// first read for a small body, and after at least one growth step for a body larger
    /// than `MAX_PREALLOC_BYTES`.
    #[test]
    #[cfg_attr(miri, ignore)] // Takes too long
    fn test_stream_reader_rejects_truncated_body() {
        let over_prealloc = MAX_PREALLOC_BYTES as i64 + 1;
        for (body_length, body_bytes) in [(1024, 10), (over_prealloc, MAX_PREALLOC_BYTES)] {
            let stream = stream_with_declared_body(body_length, body_bytes);
            let err = StreamReader::try_new(std::io::Cursor::new(stream), None)
                .expect_err("a body backed by fewer bytes than declared must not be accepted");
            assert!(
                err.to_string().contains("failed to fill whole buffer"),
                "unexpected error for body_length {body_length}: {err}"
            );
        }
    }

    /// The metadata length is untrusted for the same reason as the body length.
    #[test]
    fn test_stream_reader_rejects_unbacked_metadata_length() {
        let mut stream = Vec::new();
        stream.extend_from_slice(&CONTINUATION_MARKER);
        stream.extend_from_slice(&i32::MAX.to_le_bytes());
        stream.extend_from_slice(b"not this many bytes");

        let err = StreamReader::try_new(std::io::Cursor::new(stream), None)
            .expect_err("metadata length must be backed by the stream");
        assert!(
            err.to_string().contains("Unexpected end of stream"),
            "unexpected error: {err}"
        );
    }
}