iceberg/arrow/

reader.rs

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

//! Parquet file data reader

use std::collections::{HashMap, HashSet};
use std::ops::Range;
use std::str::FromStr;
use std::sync::Arc;

use arrow_arith::boolean::{and, and_kleene, is_not_null, is_null, not, or, or_kleene};
use arrow_array::{Array, ArrayRef, BooleanArray, Datum as ArrowDatum, RecordBatch, Scalar};
use arrow_cast::cast::cast;
use arrow_ord::cmp::{eq, gt, gt_eq, lt, lt_eq, neq};
use arrow_schema::{
    ArrowError, DataType, FieldRef, Schema as ArrowSchema, SchemaRef as ArrowSchemaRef,
};
use arrow_string::like::starts_with;
use bytes::Bytes;
use fnv::FnvHashSet;
use futures::future::BoxFuture;
use futures::{FutureExt, StreamExt, TryFutureExt, TryStreamExt, try_join};
use parquet::arrow::arrow_reader::{
    ArrowPredicateFn, ArrowReaderOptions, RowFilter, RowSelection, RowSelector,
};
use parquet::arrow::async_reader::AsyncFileReader;
use parquet::arrow::{PARQUET_FIELD_ID_META_KEY, ParquetRecordBatchStreamBuilder, ProjectionMask};
use parquet::file::metadata::{
    PageIndexPolicy, ParquetMetaData, ParquetMetaDataReader, RowGroupMetaData,
};
use parquet::schema::types::{SchemaDescriptor, Type as ParquetType};

use crate::arrow::caching_delete_file_loader::CachingDeleteFileLoader;
use crate::arrow::record_batch_transformer::RecordBatchTransformerBuilder;
use crate::arrow::{arrow_schema_to_schema, get_arrow_datum};
use crate::delete_vector::DeleteVector;
use crate::error::Result;
use crate::expr::visitors::bound_predicate_visitor::{BoundPredicateVisitor, visit};
use crate::expr::visitors::page_index_evaluator::PageIndexEvaluator;
use crate::expr::visitors::row_group_metrics_evaluator::RowGroupMetricsEvaluator;
use crate::expr::{BoundPredicate, BoundReference};
use crate::io::{FileIO, FileMetadata, FileRead};
use crate::metadata_columns::{RESERVED_FIELD_ID_FILE, is_metadata_field};
use crate::scan::{ArrowRecordBatchStream, FileScanTask, FileScanTaskStream};
use crate::spec::{Datum, NameMapping, NestedField, PrimitiveType, Schema, Type};
use crate::utils::available_parallelism;
use crate::{Error, ErrorKind};

/// Builder to create ArrowReader
pub struct ArrowReaderBuilder {
    batch_size: Option<usize>,
    file_io: FileIO,
    concurrency_limit_data_files: usize,
    row_group_filtering_enabled: bool,
    row_selection_enabled: bool,
}

impl ArrowReaderBuilder {
    /// Create a new ArrowReaderBuilder
    pub fn new(file_io: FileIO) -> Self {
        let num_cpus = available_parallelism().get();

        ArrowReaderBuilder {
            batch_size: None,
            file_io,
            concurrency_limit_data_files: num_cpus,
            row_group_filtering_enabled: true,
            row_selection_enabled: false,
        }
    }

    /// Sets the max number of in flight data files that are being fetched
    pub fn with_data_file_concurrency_limit(mut self, val: usize) -> Self {
        self.concurrency_limit_data_files = val;
        self
    }

    /// Sets the desired size of batches in the response
    /// to something other than the default
    pub fn with_batch_size(mut self, batch_size: usize) -> Self {
        self.batch_size = Some(batch_size);
        self
    }

    /// Determines whether to enable row group filtering.
    pub fn with_row_group_filtering_enabled(mut self, row_group_filtering_enabled: bool) -> Self {
        self.row_group_filtering_enabled = row_group_filtering_enabled;
        self
    }

    /// Determines whether to enable row selection.
    pub fn with_row_selection_enabled(mut self, row_selection_enabled: bool) -> Self {
        self.row_selection_enabled = row_selection_enabled;
        self
    }

    /// Build the ArrowReader.
    pub fn build(self) -> ArrowReader {
        ArrowReader {
            batch_size: self.batch_size,
            file_io: self.file_io.clone(),
            delete_file_loader: CachingDeleteFileLoader::new(
                self.file_io.clone(),
                self.concurrency_limit_data_files,
            ),
            concurrency_limit_data_files: self.concurrency_limit_data_files,
            row_group_filtering_enabled: self.row_group_filtering_enabled,
            row_selection_enabled: self.row_selection_enabled,
        }
    }
}

/// Reads data from Parquet files
#[derive(Clone)]
pub struct ArrowReader {
    batch_size: Option<usize>,
    file_io: FileIO,
    delete_file_loader: CachingDeleteFileLoader,

    /// the maximum number of data files that can be fetched at the same time
    concurrency_limit_data_files: usize,

    row_group_filtering_enabled: bool,
    row_selection_enabled: bool,
}

impl ArrowReader {
    /// Take a stream of FileScanTasks and reads all the files.
    /// Returns a stream of Arrow RecordBatches containing the data from the files
    pub fn read(self, tasks: FileScanTaskStream) -> Result<ArrowRecordBatchStream> {
        let file_io = self.file_io.clone();
        let batch_size = self.batch_size;
        let concurrency_limit_data_files = self.concurrency_limit_data_files;
        let row_group_filtering_enabled = self.row_group_filtering_enabled;
        let row_selection_enabled = self.row_selection_enabled;

        // Fast-path for single concurrency to avoid overhead of try_flatten_unordered
        let stream: ArrowRecordBatchStream = if concurrency_limit_data_files == 1 {
            Box::pin(
                tasks
                    .and_then(move |task| {
                        let file_io = file_io.clone();

                        Self::process_file_scan_task(
                            task,
                            batch_size,
                            file_io,
                            self.delete_file_loader.clone(),
                            row_group_filtering_enabled,
                            row_selection_enabled,
                        )
                    })
                    .map_err(|err| {
                        Error::new(ErrorKind::Unexpected, "file scan task generate failed")
                            .with_source(err)
                    })
                    .try_flatten(),
            )
        } else {
            Box::pin(
                tasks
                    .map_ok(move |task| {
                        let file_io = file_io.clone();

                        Self::process_file_scan_task(
                            task,
                            batch_size,
                            file_io,
                            self.delete_file_loader.clone(),
                            row_group_filtering_enabled,
                            row_selection_enabled,
                        )
                    })
                    .map_err(|err| {
                        Error::new(ErrorKind::Unexpected, "file scan task generate failed")
                            .with_source(err)
                    })
                    .try_buffer_unordered(concurrency_limit_data_files)
                    .try_flatten_unordered(concurrency_limit_data_files),
            )
        };

        Ok(stream)
    }

    #[allow(clippy::too_many_arguments)]
    async fn process_file_scan_task(
        task: FileScanTask,
        batch_size: Option<usize>,
        file_io: FileIO,
        delete_file_loader: CachingDeleteFileLoader,
        row_group_filtering_enabled: bool,
        row_selection_enabled: bool,
    ) -> Result<ArrowRecordBatchStream> {
        let should_load_page_index =
            (row_selection_enabled && task.predicate.is_some()) || !task.deletes.is_empty();

        let delete_filter_rx =
            delete_file_loader.load_deletes(&task.deletes, Arc::clone(&task.schema));

        // Migrated tables lack field IDs, requiring us to inspect the schema to choose
        // between field-ID-based or position-based projection
        let initial_stream_builder = Self::create_parquet_record_batch_stream_builder(
            &task.data_file_path,
            file_io.clone(),
            should_load_page_index,
            None,
        )
        .await?;

        // Check if Parquet file has embedded field IDs
        // Corresponds to Java's ParquetSchemaUtil.hasIds()
        // Reference: parquet/src/main/java/org/apache/iceberg/parquet/ParquetSchemaUtil.java:118
        let missing_field_ids = initial_stream_builder
            .schema()
            .fields()
            .iter()
            .next()
            .is_some_and(|f| f.metadata().get(PARQUET_FIELD_ID_META_KEY).is_none());

        // Three-branch schema resolution strategy matching Java's ReadConf constructor
        //
        // Per Iceberg spec Column Projection rules:
        // "Columns in Iceberg data files are selected by field id. The table schema's column
        //  names and order may change after a data file is written, and projection must be done
        //  using field ids."
        // https://iceberg.apache.org/spec/#column-projection
        //
        // When Parquet files lack field IDs (e.g., Hive/Spark migrations via add_files),
        // we must assign field IDs BEFORE reading data to enable correct projection.
        //
        // Java's ReadConf determines field ID strategy:
        // - Branch 1: hasIds(fileSchema) → trust embedded field IDs, use pruneColumns()
        // - Branch 2: nameMapping present → applyNameMapping(), then pruneColumns()
        // - Branch 3: fallback → addFallbackIds(), then pruneColumnsFallback()
        let mut record_batch_stream_builder = if missing_field_ids {
            // Parquet file lacks field IDs - must assign them before reading
            let arrow_schema = if let Some(name_mapping) = &task.name_mapping {
                // Branch 2: Apply name mapping to assign correct Iceberg field IDs
                // Per spec rule #2: "Use schema.name-mapping.default metadata to map field id
                // to columns without field id"
                // Corresponds to Java's ParquetSchemaUtil.applyNameMapping()
                apply_name_mapping_to_arrow_schema(
                    Arc::clone(initial_stream_builder.schema()),
                    name_mapping,
                )?
            } else {
                // Branch 3: No name mapping - use position-based fallback IDs
                // Corresponds to Java's ParquetSchemaUtil.addFallbackIds()
                add_fallback_field_ids_to_arrow_schema(initial_stream_builder.schema())
            };

            let options = ArrowReaderOptions::new().with_schema(arrow_schema);

            Self::create_parquet_record_batch_stream_builder(
                &task.data_file_path,
                file_io.clone(),
                should_load_page_index,
                Some(options),
            )
            .await?
        } else {
            // Branch 1: File has embedded field IDs - trust them
            initial_stream_builder
        };

        // Filter out metadata fields for Parquet projection (they don't exist in files)
        let project_field_ids_without_metadata: Vec<i32> = task
            .project_field_ids
            .iter()
            .filter(|&&id| !is_metadata_field(id))
            .copied()
            .collect();

        // Create projection mask based on field IDs
        // - If file has embedded IDs: field-ID-based projection (missing_field_ids=false)
        // - If name mapping applied: field-ID-based projection (missing_field_ids=true but IDs now match)
        // - If fallback IDs: position-based projection (missing_field_ids=true)
        let projection_mask = Self::get_arrow_projection_mask(
            &project_field_ids_without_metadata,
            &task.schema,
            record_batch_stream_builder.parquet_schema(),
            record_batch_stream_builder.schema(),
            missing_field_ids, // Whether to use position-based (true) or field-ID-based (false) projection
        )?;

        record_batch_stream_builder =
            record_batch_stream_builder.with_projection(projection_mask.clone());

        // RecordBatchTransformer performs any transformations required on the RecordBatches
        // that come back from the file, such as type promotion, default column insertion,
        // column re-ordering, partition constants, and virtual field addition (like _file)
        let mut record_batch_transformer_builder =
            RecordBatchTransformerBuilder::new(task.schema_ref(), task.project_field_ids());

        // Add the _file metadata column if it's in the projected fields
        if task.project_field_ids().contains(&RESERVED_FIELD_ID_FILE) {
            let file_datum = Datum::string(task.data_file_path.clone());
            record_batch_transformer_builder =
                record_batch_transformer_builder.with_constant(RESERVED_FIELD_ID_FILE, file_datum);
        }

        if let (Some(partition_spec), Some(partition_data)) =
            (task.partition_spec.clone(), task.partition.clone())
        {
            record_batch_transformer_builder =
                record_batch_transformer_builder.with_partition(partition_spec, partition_data)?;
        }

        let mut record_batch_transformer = record_batch_transformer_builder.build();

        if let Some(batch_size) = batch_size {
            record_batch_stream_builder = record_batch_stream_builder.with_batch_size(batch_size);
        }

        let delete_filter = delete_filter_rx.await.unwrap()?;
        let delete_predicate = delete_filter.build_equality_delete_predicate(&task).await?;

        // In addition to the optional predicate supplied in the `FileScanTask`,
        // we also have an optional predicate resulting from equality delete files.
        // If both are present, we logical-AND them together to form a single filter
        // predicate that we can pass to the `RecordBatchStreamBuilder`.
        let final_predicate = match (&task.predicate, delete_predicate) {
            (None, None) => None,
            (Some(predicate), None) => Some(predicate.clone()),
            (None, Some(ref predicate)) => Some(predicate.clone()),
            (Some(filter_predicate), Some(delete_predicate)) => {
                Some(filter_predicate.clone().and(delete_predicate))
            }
        };

        // There are three possible sources for potential lists of selected RowGroup indices,
        // and two for `RowSelection`s.
        // Selected RowGroup index lists can come from three sources:
        //   * When task.start and task.length specify a byte range (file splitting);
        //   * When there are equality delete files that are applicable;
        //   * When there is a scan predicate and row_group_filtering_enabled = true.
        // `RowSelection`s can be created in either or both of the following cases:
        //   * When there are positional delete files that are applicable;
        //   * When there is a scan predicate and row_selection_enabled = true
        // Note that row group filtering from predicates only happens when
        // there is a scan predicate AND row_group_filtering_enabled = true,
        // but we perform row selection filtering if there are applicable
        // equality delete files OR (there is a scan predicate AND row_selection_enabled),
        // since the only implemented method of applying positional deletes is
        // by using a `RowSelection`.
        let mut selected_row_group_indices = None;
        let mut row_selection = None;

        // Filter row groups based on byte range from task.start and task.length.
        // If both start and length are 0, read the entire file (backwards compatibility).
        if task.start != 0 || task.length != 0 {
            let byte_range_filtered_row_groups = Self::filter_row_groups_by_byte_range(
                record_batch_stream_builder.metadata(),
                task.start,
                task.length,
            )?;
            selected_row_group_indices = Some(byte_range_filtered_row_groups);
        }

        if let Some(predicate) = final_predicate {
            let (iceberg_field_ids, field_id_map) = Self::build_field_id_set_and_map(
                record_batch_stream_builder.parquet_schema(),
                &predicate,
            )?;

            let row_filter = Self::get_row_filter(
                &predicate,
                record_batch_stream_builder.parquet_schema(),
                &iceberg_field_ids,
                &field_id_map,
            )?;
            record_batch_stream_builder = record_batch_stream_builder.with_row_filter(row_filter);

            if row_group_filtering_enabled {
                let predicate_filtered_row_groups = Self::get_selected_row_group_indices(
                    &predicate,
                    record_batch_stream_builder.metadata(),
                    &field_id_map,
                    &task.schema,
                )?;

                // Merge predicate-based filtering with byte range filtering (if present)
                // by taking the intersection of both filters
                selected_row_group_indices = match selected_row_group_indices {
                    Some(byte_range_filtered) => {
                        // Keep only row groups that are in both filters
                        let intersection: Vec<usize> = byte_range_filtered
                            .into_iter()
                            .filter(|idx| predicate_filtered_row_groups.contains(idx))
                            .collect();
                        Some(intersection)
                    }
                    None => Some(predicate_filtered_row_groups),
                };
            }

            if row_selection_enabled {
                row_selection = Some(Self::get_row_selection_for_filter_predicate(
                    &predicate,
                    record_batch_stream_builder.metadata(),
                    &selected_row_group_indices,
                    &field_id_map,
                    &task.schema,
                )?);
            }
        }

        let positional_delete_indexes = delete_filter.get_delete_vector(&task);

        if let Some(positional_delete_indexes) = positional_delete_indexes {
            let delete_row_selection = {
                let positional_delete_indexes = positional_delete_indexes.lock().unwrap();

                Self::build_deletes_row_selection(
                    record_batch_stream_builder.metadata().row_groups(),
                    &selected_row_group_indices,
                    &positional_delete_indexes,
                )
            }?;

            // merge the row selection from the delete files with the row selection
            // from the filter predicate, if there is one from the filter predicate
            row_selection = match row_selection {
                None => Some(delete_row_selection),
                Some(filter_row_selection) => {
                    Some(filter_row_selection.intersection(&delete_row_selection))
                }
            };
        }

        if let Some(row_selection) = row_selection {
            record_batch_stream_builder =
                record_batch_stream_builder.with_row_selection(row_selection);
        }

        if let Some(selected_row_group_indices) = selected_row_group_indices {
            record_batch_stream_builder =
                record_batch_stream_builder.with_row_groups(selected_row_group_indices);
        }

        // Build the batch stream and send all the RecordBatches that it generates
        // to the requester.
        let record_batch_stream =
            record_batch_stream_builder
                .build()?
                .map(move |batch| match batch {
                    Ok(batch) => {
                        // Process the record batch (type promotion, column reordering, virtual fields, etc.)
                        record_batch_transformer.process_record_batch(batch)
                    }
                    Err(err) => Err(err.into()),
                });

        Ok(Box::pin(record_batch_stream) as ArrowRecordBatchStream)
    }

    pub(crate) async fn create_parquet_record_batch_stream_builder(
        data_file_path: &str,
        file_io: FileIO,
        should_load_page_index: bool,
        arrow_reader_options: Option<ArrowReaderOptions>,
    ) -> Result<ParquetRecordBatchStreamBuilder<ArrowFileReader>> {
        // Get the metadata for the Parquet file we need to read and build
        // a reader for the data within
        let parquet_file = file_io.new_input(data_file_path)?;
        let (parquet_metadata, parquet_reader) =
            try_join!(parquet_file.metadata(), parquet_file.reader())?;
        let parquet_file_reader = ArrowFileReader::new(parquet_metadata, parquet_reader)
            .with_preload_column_index(true)
            .with_preload_offset_index(true)
            .with_preload_page_index(should_load_page_index);

        // Create the record batch stream builder, which wraps the parquet file reader
        let options = arrow_reader_options.unwrap_or_default();
        let record_batch_stream_builder =
            ParquetRecordBatchStreamBuilder::new_with_options(parquet_file_reader, options).await?;
        Ok(record_batch_stream_builder)
    }

    /// computes a `RowSelection` from positional delete indices.
    ///
    /// Using the Parquet page index, we build a `RowSelection` that rejects rows that are indicated
    /// as having been deleted by a positional delete, taking into account any row groups that have
    /// been skipped entirely by the filter predicate
    fn build_deletes_row_selection(
        row_group_metadata_list: &[RowGroupMetaData],
        selected_row_groups: &Option<Vec<usize>>,
        positional_deletes: &DeleteVector,
    ) -> Result<RowSelection> {
        let mut results: Vec<RowSelector> = Vec::new();
        let mut selected_row_groups_idx = 0;
        let mut current_row_group_base_idx: u64 = 0;
        let mut delete_vector_iter = positional_deletes.iter();
        let mut next_deleted_row_idx_opt = delete_vector_iter.next();

        for (idx, row_group_metadata) in row_group_metadata_list.iter().enumerate() {
            let row_group_num_rows = row_group_metadata.num_rows() as u64;
            let next_row_group_base_idx = current_row_group_base_idx + row_group_num_rows;

            // if row group selection is enabled,
            if let Some(selected_row_groups) = selected_row_groups {
                // if we've consumed all the selected row groups, we're done
                if selected_row_groups_idx == selected_row_groups.len() {
                    break;
                }

                if idx == selected_row_groups[selected_row_groups_idx] {
                    // we're in a selected row group. Increment selected_row_groups_idx
                    // so that next time around the for loop we're looking for the next
                    // selected row group
                    selected_row_groups_idx += 1;
                } else {
                    // Advance iterator past all deletes in the skipped row group.
                    // advance_to() positions the iterator to the first delete >= next_row_group_base_idx.
                    // However, if our cached next_deleted_row_idx_opt is in the skipped range,
                    // we need to call next() to update the cache with the newly positioned value.
                    delete_vector_iter.advance_to(next_row_group_base_idx);
                    // Only update the cache if the cached value is stale (in the skipped range)
                    if let Some(cached_idx) = next_deleted_row_idx_opt
                        && cached_idx < next_row_group_base_idx
                    {
                        next_deleted_row_idx_opt = delete_vector_iter.next();
                    }

                    // still increment the current page base index but then skip to the next row group
                    // in the file
                    current_row_group_base_idx += row_group_num_rows;
                    continue;
                }
            }

            let mut next_deleted_row_idx = match next_deleted_row_idx_opt {
                Some(next_deleted_row_idx) => {
                    // if the index of the next deleted row is beyond this row group, add a selection for
                    // the remainder of this row group and skip to the next row group
                    if next_deleted_row_idx >= next_row_group_base_idx {
                        results.push(RowSelector::select(row_group_num_rows as usize));
                        current_row_group_base_idx += row_group_num_rows;
                        continue;
                    }

                    next_deleted_row_idx
                }

                // If there are no more pos deletes, add a selector for the entirety of this row group.
                _ => {
                    results.push(RowSelector::select(row_group_num_rows as usize));
                    current_row_group_base_idx += row_group_num_rows;
                    continue;
                }
            };

            let mut current_idx = current_row_group_base_idx;
            'chunks: while next_deleted_row_idx < next_row_group_base_idx {
                // `select` all rows that precede the next delete index
                if current_idx < next_deleted_row_idx {
                    let run_length = next_deleted_row_idx - current_idx;
                    results.push(RowSelector::select(run_length as usize));
                    current_idx += run_length;
                }

                // `skip` all consecutive deleted rows in the current row group
                let mut run_length = 0;
                while next_deleted_row_idx == current_idx
                    && next_deleted_row_idx < next_row_group_base_idx
                {
                    run_length += 1;
                    current_idx += 1;

                    next_deleted_row_idx_opt = delete_vector_iter.next();
                    next_deleted_row_idx = match next_deleted_row_idx_opt {
                        Some(next_deleted_row_idx) => next_deleted_row_idx,
                        _ => {
                            // We've processed the final positional delete.
                            // Conclude the skip and then break so that we select the remaining
                            // rows in the row group and move on to the next row group
                            results.push(RowSelector::skip(run_length));
                            break 'chunks;
                        }
                    };
                }
                if run_length > 0 {
                    results.push(RowSelector::skip(run_length));
                }
            }

            if current_idx < next_row_group_base_idx {
                results.push(RowSelector::select(
                    (next_row_group_base_idx - current_idx) as usize,
                ));
            }

            current_row_group_base_idx += row_group_num_rows;
        }

        Ok(results.into())
    }

    fn build_field_id_set_and_map(
        parquet_schema: &SchemaDescriptor,
        predicate: &BoundPredicate,
    ) -> Result<(HashSet<i32>, HashMap<i32, usize>)> {
        // Collects all Iceberg field IDs referenced in the filter predicate
        let mut collector = CollectFieldIdVisitor {
            field_ids: HashSet::default(),
        };
        visit(&mut collector, predicate)?;

        let iceberg_field_ids = collector.field_ids();

        // Without embedded field IDs, we fall back to position-based mapping for compatibility
        let field_id_map = match build_field_id_map(parquet_schema)? {
            Some(map) => map,
            None => build_fallback_field_id_map(parquet_schema),
        };

        Ok((iceberg_field_ids, field_id_map))
    }

    /// Recursively extract leaf field IDs because Parquet projection works at the leaf column level.
    /// Nested types (struct/list/map) are flattened in Parquet's columnar format.
    fn include_leaf_field_id(field: &NestedField, field_ids: &mut Vec<i32>) {
        match field.field_type.as_ref() {
            Type::Primitive(_) => {
                field_ids.push(field.id);
            }
            Type::Struct(struct_type) => {
                for nested_field in struct_type.fields() {
                    Self::include_leaf_field_id(nested_field, field_ids);
                }
            }
            Type::List(list_type) => {
                Self::include_leaf_field_id(&list_type.element_field, field_ids);
            }
            Type::Map(map_type) => {
                Self::include_leaf_field_id(&map_type.key_field, field_ids);
                Self::include_leaf_field_id(&map_type.value_field, field_ids);
            }
        }
    }

    fn get_arrow_projection_mask(
        field_ids: &[i32],
        iceberg_schema_of_task: &Schema,
        parquet_schema: &SchemaDescriptor,
        arrow_schema: &ArrowSchemaRef,
        use_fallback: bool, // Whether file lacks embedded field IDs (e.g., migrated from Hive/Spark)
    ) -> Result<ProjectionMask> {
        fn type_promotion_is_valid(
            file_type: Option<&PrimitiveType>,
            projected_type: Option<&PrimitiveType>,
        ) -> bool {
            match (file_type, projected_type) {
                (Some(lhs), Some(rhs)) if lhs == rhs => true,
                (Some(PrimitiveType::Int), Some(PrimitiveType::Long)) => true,
                (Some(PrimitiveType::Float), Some(PrimitiveType::Double)) => true,
                (
                    Some(PrimitiveType::Decimal {
                        precision: file_precision,
                        scale: file_scale,
                    }),
                    Some(PrimitiveType::Decimal {
                        precision: requested_precision,
                        scale: requested_scale,
                    }),
                ) if requested_precision >= file_precision && file_scale == requested_scale => true,
                // Uuid will be store as Fixed(16) in parquet file, so the read back type will be Fixed(16).
                (Some(PrimitiveType::Fixed(16)), Some(PrimitiveType::Uuid)) => true,
                _ => false,
            }
        }

        if field_ids.is_empty() {
            return Ok(ProjectionMask::all());
        }

        if use_fallback {
            // Position-based projection necessary because file lacks embedded field IDs
            Self::get_arrow_projection_mask_fallback(field_ids, parquet_schema)
        } else {
            // Field-ID-based projection using embedded field IDs from Parquet metadata

            // Parquet's columnar format requires leaf-level (not top-level struct/list/map) projection
            let mut leaf_field_ids = vec![];
            for field_id in field_ids {
                let field = iceberg_schema_of_task.field_by_id(*field_id);
                if let Some(field) = field {
                    Self::include_leaf_field_id(field, &mut leaf_field_ids);
                }
            }

            Self::get_arrow_projection_mask_with_field_ids(
                &leaf_field_ids,
                iceberg_schema_of_task,
                parquet_schema,
                arrow_schema,
                type_promotion_is_valid,
            )
        }
    }

    /// Standard projection using embedded field IDs from Parquet metadata.
    /// For iceberg-java compatibility with ParquetSchemaUtil.pruneColumns().
    fn get_arrow_projection_mask_with_field_ids(
        leaf_field_ids: &[i32],
        iceberg_schema_of_task: &Schema,
        parquet_schema: &SchemaDescriptor,
        arrow_schema: &ArrowSchemaRef,
        type_promotion_is_valid: fn(Option<&PrimitiveType>, Option<&PrimitiveType>) -> bool,
    ) -> Result<ProjectionMask> {
        let mut column_map = HashMap::new();
        let fields = arrow_schema.fields();

        // Pre-project only the fields that have been selected, possibly avoiding converting
        // some Arrow types that are not yet supported.
        let mut projected_fields: HashMap<FieldRef, i32> = HashMap::new();
        let projected_arrow_schema = ArrowSchema::new_with_metadata(
            fields.filter_leaves(|_, f| {
                f.metadata()
                    .get(PARQUET_FIELD_ID_META_KEY)
                    .and_then(|field_id| i32::from_str(field_id).ok())
                    .is_some_and(|field_id| {
                        projected_fields.insert((*f).clone(), field_id);
                        leaf_field_ids.contains(&field_id)
                    })
            }),
            arrow_schema.metadata().clone(),
        );
        let iceberg_schema = arrow_schema_to_schema(&projected_arrow_schema)?;

        fields.filter_leaves(|idx, field| {
            let Some(field_id) = projected_fields.get(field).cloned() else {
                return false;
            };

            let iceberg_field = iceberg_schema_of_task.field_by_id(field_id);
            let parquet_iceberg_field = iceberg_schema.field_by_id(field_id);

            if iceberg_field.is_none() || parquet_iceberg_field.is_none() {
                return false;
            }

            if !type_promotion_is_valid(
                parquet_iceberg_field
                    .unwrap()
                    .field_type
                    .as_primitive_type(),
                iceberg_field.unwrap().field_type.as_primitive_type(),
            ) {
                return false;
            }

            column_map.insert(field_id, idx);
            true
        });

        // Schema evolution: New columns may not exist in old Parquet files.
        // We only project existing columns; RecordBatchTransformer adds default/NULL values.
        let mut indices = vec![];
        for field_id in leaf_field_ids {
            if let Some(col_idx) = column_map.get(field_id) {
                indices.push(*col_idx);
            }
        }

        if indices.is_empty() {
            // Edge case: All requested columns are new (don't exist in file).
            // Project all columns so RecordBatchTransformer has a batch to transform.
            Ok(ProjectionMask::all())
        } else {
            Ok(ProjectionMask::leaves(parquet_schema, indices))
        }
    }

    /// Fallback projection for Parquet files without field IDs.
    /// Uses position-based matching: field ID N → column position N-1.
    /// Projects entire top-level columns (including nested content) for iceberg-java compatibility.
    fn get_arrow_projection_mask_fallback(
        field_ids: &[i32],
        parquet_schema: &SchemaDescriptor,
    ) -> Result<ProjectionMask> {
        // Position-based: field_id N → column N-1 (field IDs are 1-indexed)
        let parquet_root_fields = parquet_schema.root_schema().get_fields();
        let mut root_indices = vec![];

        for field_id in field_ids.iter() {
            let parquet_pos = (*field_id - 1) as usize;

            if parquet_pos < parquet_root_fields.len() {
                root_indices.push(parquet_pos);
            }
            // RecordBatchTransformer adds missing columns with NULL values
        }

        if root_indices.is_empty() {
            Ok(ProjectionMask::all())
        } else {
            Ok(ProjectionMask::roots(parquet_schema, root_indices))
        }
    }

    fn get_row_filter(
        predicates: &BoundPredicate,
        parquet_schema: &SchemaDescriptor,
        iceberg_field_ids: &HashSet<i32>,
        field_id_map: &HashMap<i32, usize>,
    ) -> Result<RowFilter> {
        // Collect Parquet column indices from field ids.
        // If the field id is not found in Parquet schema, it will be ignored due to schema evolution.
        let mut column_indices = iceberg_field_ids
            .iter()
            .filter_map(|field_id| field_id_map.get(field_id).cloned())
            .collect::<Vec<_>>();
        column_indices.sort();

        // The converter that converts `BoundPredicates` to `ArrowPredicates`
        let mut converter = PredicateConverter {
            parquet_schema,
            column_map: field_id_map,
            column_indices: &column_indices,
        };

        // After collecting required leaf column indices used in the predicate,
        // creates the projection mask for the Arrow predicates.
        let projection_mask = ProjectionMask::leaves(parquet_schema, column_indices.clone());
        let predicate_func = visit(&mut converter, predicates)?;
        let arrow_predicate = ArrowPredicateFn::new(projection_mask, predicate_func);
        Ok(RowFilter::new(vec![Box::new(arrow_predicate)]))
    }

    fn get_selected_row_group_indices(
        predicate: &BoundPredicate,
        parquet_metadata: &Arc<ParquetMetaData>,
        field_id_map: &HashMap<i32, usize>,
        snapshot_schema: &Schema,
    ) -> Result<Vec<usize>> {
        let row_groups_metadata = parquet_metadata.row_groups();
        let mut results = Vec::with_capacity(row_groups_metadata.len());

        for (idx, row_group_metadata) in row_groups_metadata.iter().enumerate() {
            if RowGroupMetricsEvaluator::eval(
                predicate,
                row_group_metadata,
                field_id_map,
                snapshot_schema,
            )? {
                results.push(idx);
            }
        }

        Ok(results)
    }

    fn get_row_selection_for_filter_predicate(
        predicate: &BoundPredicate,
        parquet_metadata: &Arc<ParquetMetaData>,
        selected_row_groups: &Option<Vec<usize>>,
        field_id_map: &HashMap<i32, usize>,
        snapshot_schema: &Schema,
    ) -> Result<RowSelection> {
        let Some(column_index) = parquet_metadata.column_index() else {
            return Err(Error::new(
                ErrorKind::Unexpected,
                "Parquet file metadata does not contain a column index",
            ));
        };

        let Some(offset_index) = parquet_metadata.offset_index() else {
            return Err(Error::new(
                ErrorKind::Unexpected,
                "Parquet file metadata does not contain an offset index",
            ));
        };

        // If all row groups were filtered out, return an empty RowSelection (select no rows)
        if let Some(selected_row_groups) = selected_row_groups
            && selected_row_groups.is_empty()
        {
            return Ok(RowSelection::from(Vec::new()));
        }

        let mut selected_row_groups_idx = 0;

        let page_index = column_index
            .iter()
            .enumerate()
            .zip(offset_index)
            .zip(parquet_metadata.row_groups());

        let mut results = Vec::new();
        for (((idx, column_index), offset_index), row_group_metadata) in page_index {
            if let Some(selected_row_groups) = selected_row_groups {
                // skip row groups that aren't present in selected_row_groups
                if idx == selected_row_groups[selected_row_groups_idx] {
                    selected_row_groups_idx += 1;
                } else {
                    continue;
                }
            }

            let selections_for_page = PageIndexEvaluator::eval(
                predicate,
                column_index,
                offset_index,
                row_group_metadata,
                field_id_map,
                snapshot_schema,
            )?;

            results.push(selections_for_page);

            if let Some(selected_row_groups) = selected_row_groups
                && selected_row_groups_idx == selected_row_groups.len()
            {
                break;
            }
        }

        Ok(results.into_iter().flatten().collect::<Vec<_>>().into())
    }

    /// Filters row groups by byte range to support Iceberg's file splitting.
    ///
    /// Iceberg splits large files at row group boundaries, so we only read row groups
    /// whose byte ranges overlap with [start, start+length).
    fn filter_row_groups_by_byte_range(
        parquet_metadata: &Arc<ParquetMetaData>,
        start: u64,
        length: u64,
    ) -> Result<Vec<usize>> {
        let row_groups = parquet_metadata.row_groups();
        let mut selected = Vec::new();
        let end = start + length;

        // Row groups are stored sequentially after the 4-byte magic header.
        let mut current_byte_offset = 4u64;

        for (idx, row_group) in row_groups.iter().enumerate() {
            let row_group_size = row_group.compressed_size() as u64;
            let row_group_end = current_byte_offset + row_group_size;

            if current_byte_offset < end && start < row_group_end {
                selected.push(idx);
            }

            current_byte_offset = row_group_end;
        }

        Ok(selected)
    }
}

/// Build the map of parquet field id to Parquet column index in the schema.
/// Returns None if the Parquet file doesn't have field IDs embedded (e.g., migrated tables).
fn build_field_id_map(parquet_schema: &SchemaDescriptor) -> Result<Option<HashMap<i32, usize>>> {
    let mut column_map = HashMap::new();

    for (idx, field) in parquet_schema.columns().iter().enumerate() {
        let field_type = field.self_type();
        match field_type {
            ParquetType::PrimitiveType { basic_info, .. } => {
                if !basic_info.has_id() {
                    return Ok(None);
                }
                column_map.insert(basic_info.id(), idx);
            }
            ParquetType::GroupType { .. } => {
                return Err(Error::new(
                    ErrorKind::DataInvalid,
                    format!(
                        "Leave column in schema should be primitive type but got {field_type:?}"
                    ),
                ));
            }
        };
    }

    Ok(Some(column_map))
}

/// Build a fallback field ID map for Parquet files without embedded field IDs.
/// Position-based (1, 2, 3, ...) for compatibility with iceberg-java migrations.
fn build_fallback_field_id_map(parquet_schema: &SchemaDescriptor) -> HashMap<i32, usize> {
    let mut column_map = HashMap::new();

    // 1-indexed to match iceberg-java's convention
    for (idx, _field) in parquet_schema.columns().iter().enumerate() {
        let field_id = (idx + 1) as i32;
        column_map.insert(field_id, idx);
    }

    column_map
}

/// Apply name mapping to Arrow schema for Parquet files lacking field IDs.
///
/// Assigns Iceberg field IDs based on column names using the name mapping,
/// enabling correct projection on migrated files (e.g., from Hive/Spark via add_files).
///
/// Per Iceberg spec Column Projection rule #2:
/// "Use schema.name-mapping.default metadata to map field id to columns without field id"
/// https://iceberg.apache.org/spec/#column-projection
///
/// Corresponds to Java's ParquetSchemaUtil.applyNameMapping() and ApplyNameMapping visitor.
/// The key difference is Java operates on Parquet MessageType, while we operate on Arrow Schema.
///
/// # Arguments
/// * `arrow_schema` - Arrow schema from Parquet file (without field IDs)
/// * `name_mapping` - Name mapping from table metadata (TableProperties.DEFAULT_NAME_MAPPING)
///
/// # Returns
/// Arrow schema with field IDs assigned based on name mapping
fn apply_name_mapping_to_arrow_schema(
    arrow_schema: ArrowSchemaRef,
    name_mapping: &NameMapping,
) -> Result<Arc<ArrowSchema>> {
    debug_assert!(
        arrow_schema
            .fields()
            .iter()
            .next()
            .is_none_or(|f| f.metadata().get(PARQUET_FIELD_ID_META_KEY).is_none()),
        "Schema already has field IDs - name mapping should not be applied"
    );

    use arrow_schema::Field;

    let fields_with_mapped_ids: Vec<_> = arrow_schema
        .fields()
        .iter()
        .map(|field| {
            // Look up this column name in name mapping to get the Iceberg field ID.
            // Corresponds to Java's ApplyNameMapping visitor which calls
            // nameMapping.find(currentPath()) and returns field.withId() if found.
            //
            // If the field isn't in the mapping, leave it WITHOUT assigning an ID
            // (matching Java's behavior of returning the field unchanged).
            // Later, during projection, fields without IDs are filtered out.
            let mapped_field_opt = name_mapping
                .fields()
                .iter()
                .find(|f| f.names().contains(&field.name().to_string()));

            let mut metadata = field.metadata().clone();

            if let Some(mapped_field) = mapped_field_opt
                && let Some(field_id) = mapped_field.field_id()
            {
                // Field found in mapping with a field_id → assign it
                metadata.insert(PARQUET_FIELD_ID_META_KEY.to_string(), field_id.to_string());
            }
            // If field_id is None, leave the field without an ID (will be filtered by projection)

            Field::new(field.name(), field.data_type().clone(), field.is_nullable())
                .with_metadata(metadata)
        })
        .collect();

    Ok(Arc::new(ArrowSchema::new_with_metadata(
        fields_with_mapped_ids,
        arrow_schema.metadata().clone(),
    )))
}

/// Add position-based fallback field IDs to Arrow schema for Parquet files lacking them.
/// Enables projection on migrated files (e.g., from Hive/Spark).
///
/// Why at schema level (not per-batch): Efficiency - avoids repeated schema modification.
/// Why only top-level: Nested projection uses leaf column indices, not parent struct IDs.
/// Why 1-indexed: Compatibility with iceberg-java's ParquetSchemaUtil.addFallbackIds().
fn add_fallback_field_ids_to_arrow_schema(arrow_schema: &ArrowSchemaRef) -> Arc<ArrowSchema> {
    debug_assert!(
        arrow_schema
            .fields()
            .iter()
            .next()
            .is_none_or(|f| f.metadata().get(PARQUET_FIELD_ID_META_KEY).is_none()),
        "Schema already has field IDs"
    );

    use arrow_schema::Field;

    let fields_with_fallback_ids: Vec<_> = arrow_schema
        .fields()
        .iter()
        .enumerate()
        .map(|(pos, field)| {
            let mut metadata = field.metadata().clone();
            let field_id = (pos + 1) as i32; // 1-indexed for Java compatibility
            metadata.insert(PARQUET_FIELD_ID_META_KEY.to_string(), field_id.to_string());

            Field::new(field.name(), field.data_type().clone(), field.is_nullable())
                .with_metadata(metadata)
        })
        .collect();

    Arc::new(ArrowSchema::new_with_metadata(
        fields_with_fallback_ids,
        arrow_schema.metadata().clone(),
    ))
}

/// A visitor to collect field ids from bound predicates.
struct CollectFieldIdVisitor {
    field_ids: HashSet<i32>,
}

impl CollectFieldIdVisitor {
    fn field_ids(self) -> HashSet<i32> {
        self.field_ids
    }
}

impl BoundPredicateVisitor for CollectFieldIdVisitor {
    type T = ();

    fn always_true(&mut self) -> Result<()> {
        Ok(())
    }

    fn always_false(&mut self) -> Result<()> {
        Ok(())
    }

    fn and(&mut self, _lhs: (), _rhs: ()) -> Result<()> {
        Ok(())
    }

    fn or(&mut self, _lhs: (), _rhs: ()) -> Result<()> {
        Ok(())
    }

    fn not(&mut self, _inner: ()) -> Result<()> {
        Ok(())
    }

    fn is_null(&mut self, reference: &BoundReference, _predicate: &BoundPredicate) -> Result<()> {
        self.field_ids.insert(reference.field().id);
        Ok(())
    }

    fn not_null(&mut self, reference: &BoundReference, _predicate: &BoundPredicate) -> Result<()> {
        self.field_ids.insert(reference.field().id);
        Ok(())
    }

    fn is_nan(&mut self, reference: &BoundReference, _predicate: &BoundPredicate) -> Result<()> {
        self.field_ids.insert(reference.field().id);
        Ok(())
    }

    fn not_nan(&mut self, reference: &BoundReference, _predicate: &BoundPredicate) -> Result<()> {
        self.field_ids.insert(reference.field().id);
        Ok(())
    }

    fn less_than(
        &mut self,
        reference: &BoundReference,
        _literal: &Datum,
        _predicate: &BoundPredicate,
    ) -> Result<()> {
        self.field_ids.insert(reference.field().id);
        Ok(())
    }

    fn less_than_or_eq(
        &mut self,
        reference: &BoundReference,
        _literal: &Datum,
        _predicate: &BoundPredicate,
    ) -> Result<()> {
        self.field_ids.insert(reference.field().id);
        Ok(())
    }

    fn greater_than(
        &mut self,
        reference: &BoundReference,
        _literal: &Datum,
        _predicate: &BoundPredicate,
    ) -> Result<()> {
        self.field_ids.insert(reference.field().id);
        Ok(())
    }

    fn greater_than_or_eq(
        &mut self,
        reference: &BoundReference,
        _literal: &Datum,
        _predicate: &BoundPredicate,
    ) -> Result<()> {
        self.field_ids.insert(reference.field().id);
        Ok(())
    }

    fn eq(
        &mut self,
        reference: &BoundReference,
        _literal: &Datum,
        _predicate: &BoundPredicate,
    ) -> Result<()> {
        self.field_ids.insert(reference.field().id);
        Ok(())
    }

    fn not_eq(
        &mut self,
        reference: &BoundReference,
        _literal: &Datum,
        _predicate: &BoundPredicate,
    ) -> Result<()> {
        self.field_ids.insert(reference.field().id);
        Ok(())
    }

    fn starts_with(
        &mut self,
        reference: &BoundReference,
        _literal: &Datum,
        _predicate: &BoundPredicate,
    ) -> Result<()> {
        self.field_ids.insert(reference.field().id);
        Ok(())
    }

    fn not_starts_with(
        &mut self,
        reference: &BoundReference,
        _literal: &Datum,
        _predicate: &BoundPredicate,
    ) -> Result<()> {
        self.field_ids.insert(reference.field().id);
        Ok(())
    }

    fn r#in(
        &mut self,
        reference: &BoundReference,
        _literals: &FnvHashSet<Datum>,
        _predicate: &BoundPredicate,
    ) -> Result<()> {
        self.field_ids.insert(reference.field().id);
        Ok(())
    }

    fn not_in(
        &mut self,
        reference: &BoundReference,
        _literals: &FnvHashSet<Datum>,
        _predicate: &BoundPredicate,
    ) -> Result<()> {
        self.field_ids.insert(reference.field().id);
        Ok(())
    }
}

/// A visitor to convert Iceberg bound predicates to Arrow predicates.
struct PredicateConverter<'a> {
    /// The Parquet schema descriptor.
    pub parquet_schema: &'a SchemaDescriptor,
    /// The map between field id and leaf column index in Parquet schema.
    pub column_map: &'a HashMap<i32, usize>,
    /// The required column indices in Parquet schema for the predicates.
    pub column_indices: &'a Vec<usize>,
}

impl PredicateConverter<'_> {
    /// When visiting a bound reference, we return index of the leaf column in the
    /// required column indices which is used to project the column in the record batch.
    /// Return None if the field id is not found in the column map, which is possible
    /// due to schema evolution.
    fn bound_reference(&mut self, reference: &BoundReference) -> Result<Option<usize>> {
        // The leaf column's index in Parquet schema.
        if let Some(column_idx) = self.column_map.get(&reference.field().id) {
            if self.parquet_schema.get_column_root(*column_idx).is_group() {
                return Err(Error::new(
                    ErrorKind::DataInvalid,
                    format!(
                        "Leave column `{}` in predicates isn't a root column in Parquet schema.",
                        reference.field().name
                    ),
                ));
            }

            // The leaf column's index in the required column indices.
            let index = self
                .column_indices
                .iter()
                .position(|&idx| idx == *column_idx)
                .ok_or(Error::new(
                    ErrorKind::DataInvalid,
                    format!(
                "Leave column `{}` in predicates cannot be found in the required column indices.",
                reference.field().name
            ),
                ))?;

            Ok(Some(index))
        } else {
            Ok(None)
        }
    }

    /// Build an Arrow predicate that always returns true.
    fn build_always_true(&self) -> Result<Box<PredicateResult>> {
        Ok(Box::new(|batch| {
            Ok(BooleanArray::from(vec![true; batch.num_rows()]))
        }))
    }

    /// Build an Arrow predicate that always returns false.
    fn build_always_false(&self) -> Result<Box<PredicateResult>> {
        Ok(Box::new(|batch| {
            Ok(BooleanArray::from(vec![false; batch.num_rows()]))
        }))
    }
}

/// Gets the leaf column from the record batch for the required column index. Only
/// supports top-level columns for now.
fn project_column(
    batch: &RecordBatch,
    column_idx: usize,
) -> std::result::Result<ArrayRef, ArrowError> {
    let column = batch.column(column_idx);

    match column.data_type() {
        DataType::Struct(_) => Err(ArrowError::SchemaError(
            "Does not support struct column yet.".to_string(),
        )),
        _ => Ok(column.clone()),
    }
}

type PredicateResult =
    dyn FnMut(RecordBatch) -> std::result::Result<BooleanArray, ArrowError> + Send + 'static;

impl BoundPredicateVisitor for PredicateConverter<'_> {
    type T = Box<PredicateResult>;

    fn always_true(&mut self) -> Result<Box<PredicateResult>> {
        self.build_always_true()
    }

    fn always_false(&mut self) -> Result<Box<PredicateResult>> {
        self.build_always_false()
    }

    fn and(
        &mut self,
        mut lhs: Box<PredicateResult>,
        mut rhs: Box<PredicateResult>,
    ) -> Result<Box<PredicateResult>> {
        Ok(Box::new(move |batch| {
            let left = lhs(batch.clone())?;
            let right = rhs(batch)?;
            and_kleene(&left, &right)
        }))
    }

    fn or(
        &mut self,
        mut lhs: Box<PredicateResult>,
        mut rhs: Box<PredicateResult>,
    ) -> Result<Box<PredicateResult>> {
        Ok(Box::new(move |batch| {
            let left = lhs(batch.clone())?;
            let right = rhs(batch)?;
            or_kleene(&left, &right)
        }))
    }

    fn not(&mut self, mut inner: Box<PredicateResult>) -> Result<Box<PredicateResult>> {
        Ok(Box::new(move |batch| {
            let pred_ret = inner(batch)?;
            not(&pred_ret)
        }))
    }

    fn is_null(
        &mut self,
        reference: &BoundReference,
        _predicate: &BoundPredicate,
    ) -> Result<Box<PredicateResult>> {
        if let Some(idx) = self.bound_reference(reference)? {
            Ok(Box::new(move |batch| {
                let column = project_column(&batch, idx)?;
                is_null(&column)
            }))
        } else {
            // A missing column, treating it as null.
            self.build_always_true()
        }
    }

    fn not_null(
        &mut self,
        reference: &BoundReference,
        _predicate: &BoundPredicate,
    ) -> Result<Box<PredicateResult>> {
        if let Some(idx) = self.bound_reference(reference)? {
            Ok(Box::new(move |batch| {
                let column = project_column(&batch, idx)?;
                is_not_null(&column)
            }))
        } else {
            // A missing column, treating it as null.
            self.build_always_false()
        }
    }

    fn is_nan(
        &mut self,
        reference: &BoundReference,
        _predicate: &BoundPredicate,
    ) -> Result<Box<PredicateResult>> {
        if self.bound_reference(reference)?.is_some() {
            self.build_always_true()
        } else {
            // A missing column, treating it as null.
            self.build_always_false()
        }
    }

    fn not_nan(
        &mut self,
        reference: &BoundReference,
        _predicate: &BoundPredicate,
    ) -> Result<Box<PredicateResult>> {
        if self.bound_reference(reference)?.is_some() {
            self.build_always_false()
        } else {
            // A missing column, treating it as null.
            self.build_always_true()
        }
    }

    fn less_than(
        &mut self,
        reference: &BoundReference,
        literal: &Datum,
        _predicate: &BoundPredicate,
    ) -> Result<Box<PredicateResult>> {
        if let Some(idx) = self.bound_reference(reference)? {
            let literal = get_arrow_datum(literal)?;

            Ok(Box::new(move |batch| {
                let left = project_column(&batch, idx)?;
                let literal = try_cast_literal(&literal, left.data_type())?;
                lt(&left, literal.as_ref())
            }))
        } else {
            // A missing column, treating it as null.
            self.build_always_true()
        }
    }

    fn less_than_or_eq(
        &mut self,
        reference: &BoundReference,
        literal: &Datum,
        _predicate: &BoundPredicate,
    ) -> Result<Box<PredicateResult>> {
        if let Some(idx) = self.bound_reference(reference)? {
            let literal = get_arrow_datum(literal)?;

            Ok(Box::new(move |batch| {
                let left = project_column(&batch, idx)?;
                let literal = try_cast_literal(&literal, left.data_type())?;
                lt_eq(&left, literal.as_ref())
            }))
        } else {
            // A missing column, treating it as null.
            self.build_always_true()
        }
    }

    fn greater_than(
        &mut self,
        reference: &BoundReference,
        literal: &Datum,
        _predicate: &BoundPredicate,
    ) -> Result<Box<PredicateResult>> {
        if let Some(idx) = self.bound_reference(reference)? {
            let literal = get_arrow_datum(literal)?;

            Ok(Box::new(move |batch| {
                let left = project_column(&batch, idx)?;
                let literal = try_cast_literal(&literal, left.data_type())?;
                gt(&left, literal.as_ref())
            }))
        } else {
            // A missing column, treating it as null.
            self.build_always_false()
        }
    }

    fn greater_than_or_eq(
        &mut self,
        reference: &BoundReference,
        literal: &Datum,
        _predicate: &BoundPredicate,
    ) -> Result<Box<PredicateResult>> {
        if let Some(idx) = self.bound_reference(reference)? {
            let literal = get_arrow_datum(literal)?;

            Ok(Box::new(move |batch| {
                let left = project_column(&batch, idx)?;
                let literal = try_cast_literal(&literal, left.data_type())?;
                gt_eq(&left, literal.as_ref())
            }))
        } else {
            // A missing column, treating it as null.
            self.build_always_false()
        }
    }

    fn eq(
        &mut self,
        reference: &BoundReference,
        literal: &Datum,
        _predicate: &BoundPredicate,
    ) -> Result<Box<PredicateResult>> {
        if let Some(idx) = self.bound_reference(reference)? {
            let literal = get_arrow_datum(literal)?;

            Ok(Box::new(move |batch| {
                let left = project_column(&batch, idx)?;
                let literal = try_cast_literal(&literal, left.data_type())?;
                eq(&left, literal.as_ref())
            }))
        } else {
            // A missing column, treating it as null.
            self.build_always_false()
        }
    }

    fn not_eq(
        &mut self,
        reference: &BoundReference,
        literal: &Datum,
        _predicate: &BoundPredicate,
    ) -> Result<Box<PredicateResult>> {
        if let Some(idx) = self.bound_reference(reference)? {
            let literal = get_arrow_datum(literal)?;

            Ok(Box::new(move |batch| {
                let left = project_column(&batch, idx)?;
                let literal = try_cast_literal(&literal, left.data_type())?;
                neq(&left, literal.as_ref())
            }))
        } else {
            // A missing column, treating it as null.
            self.build_always_false()
        }
    }

    fn starts_with(
        &mut self,
        reference: &BoundReference,
        literal: &Datum,
        _predicate: &BoundPredicate,
    ) -> Result<Box<PredicateResult>> {
        if let Some(idx) = self.bound_reference(reference)? {
            let literal = get_arrow_datum(literal)?;

            Ok(Box::new(move |batch| {
                let left = project_column(&batch, idx)?;
                let literal = try_cast_literal(&literal, left.data_type())?;
                starts_with(&left, literal.as_ref())
            }))
        } else {
            // A missing column, treating it as null.
            self.build_always_false()
        }
    }

    fn not_starts_with(
        &mut self,
        reference: &BoundReference,
        literal: &Datum,
        _predicate: &BoundPredicate,
    ) -> Result<Box<PredicateResult>> {
        if let Some(idx) = self.bound_reference(reference)? {
            let literal = get_arrow_datum(literal)?;

            Ok(Box::new(move |batch| {
                let left = project_column(&batch, idx)?;
                let literal = try_cast_literal(&literal, left.data_type())?;
                // update here if arrow ever adds a native not_starts_with
                not(&starts_with(&left, literal.as_ref())?)
            }))
        } else {
            // A missing column, treating it as null.
            self.build_always_true()
        }
    }

    fn r#in(
        &mut self,
        reference: &BoundReference,
        literals: &FnvHashSet<Datum>,
        _predicate: &BoundPredicate,
    ) -> Result<Box<PredicateResult>> {
        if let Some(idx) = self.bound_reference(reference)? {
            let literals: Vec<_> = literals
                .iter()
                .map(|lit| get_arrow_datum(lit).unwrap())
                .collect();

            Ok(Box::new(move |batch| {
                // update this if arrow ever adds a native is_in kernel
                let left = project_column(&batch, idx)?;

                let mut acc = BooleanArray::from(vec![false; batch.num_rows()]);
                for literal in &literals {
                    let literal = try_cast_literal(literal, left.data_type())?;
                    acc = or(&acc, &eq(&left, literal.as_ref())?)?
                }

                Ok(acc)
            }))
        } else {
            // A missing column, treating it as null.
            self.build_always_false()
        }
    }

    fn not_in(
        &mut self,
        reference: &BoundReference,
        literals: &FnvHashSet<Datum>,
        _predicate: &BoundPredicate,
    ) -> Result<Box<PredicateResult>> {
        if let Some(idx) = self.bound_reference(reference)? {
            let literals: Vec<_> = literals
                .iter()
                .map(|lit| get_arrow_datum(lit).unwrap())
                .collect();

            Ok(Box::new(move |batch| {
                // update this if arrow ever adds a native not_in kernel
                let left = project_column(&batch, idx)?;
                let mut acc = BooleanArray::from(vec![true; batch.num_rows()]);
                for literal in &literals {
                    let literal = try_cast_literal(literal, left.data_type())?;
                    acc = and(&acc, &neq(&left, literal.as_ref())?)?
                }

                Ok(acc)
            }))
        } else {
            // A missing column, treating it as null.
            self.build_always_true()
        }
    }
}

/// ArrowFileReader is a wrapper around a FileRead that impls parquets AsyncFileReader.
pub struct ArrowFileReader {
    meta: FileMetadata,
    preload_column_index: bool,
    preload_offset_index: bool,
    preload_page_index: bool,
    metadata_size_hint: Option<usize>,
    r: Box<dyn FileRead>,
}

impl ArrowFileReader {
    /// Create a new ArrowFileReader
    pub fn new(meta: FileMetadata, r: Box<dyn FileRead>) -> Self {
        Self {
            meta,
            preload_column_index: false,
            preload_offset_index: false,
            preload_page_index: false,
            metadata_size_hint: None,
            r,
        }
    }

    /// Enable or disable preloading of the column index
    pub fn with_preload_column_index(mut self, preload: bool) -> Self {
        self.preload_column_index = preload;
        self
    }

    /// Enable or disable preloading of the offset index
    pub fn with_preload_offset_index(mut self, preload: bool) -> Self {
        self.preload_offset_index = preload;
        self
    }

    /// Enable or disable preloading of the page index
    pub fn with_preload_page_index(mut self, preload: bool) -> Self {
        self.preload_page_index = preload;
        self
    }

    /// Provide a hint as to the number of bytes to prefetch for parsing the Parquet metadata
    ///
    /// This hint can help reduce the number of fetch requests. For more details see the
    /// [ParquetMetaDataReader documentation](https://docs.rs/parquet/latest/parquet/file/metadata/struct.ParquetMetaDataReader.html#method.with_prefetch_hint).
    pub fn with_metadata_size_hint(mut self, hint: usize) -> Self {
        self.metadata_size_hint = Some(hint);
        self
    }
}

impl AsyncFileReader for ArrowFileReader {
    fn get_bytes(&mut self, range: Range<u64>) -> BoxFuture<'_, parquet::errors::Result<Bytes>> {
        Box::pin(
            self.r
                .read(range.start..range.end)
                .map_err(|err| parquet::errors::ParquetError::External(Box::new(err))),
        )
    }

    // TODO: currently we don't respect `ArrowReaderOptions` cause it don't expose any method to access the option field
    // we will fix it after `v55.1.0` is released in https://github.com/apache/arrow-rs/issues/7393
    fn get_metadata(
        &mut self,
        _options: Option<&'_ ArrowReaderOptions>,
    ) -> BoxFuture<'_, parquet::errors::Result<Arc<ParquetMetaData>>> {
        async move {
            let reader = ParquetMetaDataReader::new()
                .with_prefetch_hint(self.metadata_size_hint)
                // Set the page policy first because it updates both column and offset policies.
                .with_page_index_policy(PageIndexPolicy::from(self.preload_page_index))
                .with_column_index_policy(PageIndexPolicy::from(self.preload_column_index))
                .with_offset_index_policy(PageIndexPolicy::from(self.preload_offset_index));
            let size = self.meta.size;
            let meta = reader.load_and_finish(self, size).await?;

            Ok(Arc::new(meta))
        }
        .boxed()
    }
}

/// The Arrow type of an array that the Parquet reader reads may not match the exact Arrow type
/// that Iceberg uses for literals - but they are effectively the same logical type,
/// i.e. LargeUtf8 and Utf8 or Utf8View and Utf8 or Utf8View and LargeUtf8.
///
/// The Arrow compute kernels that we use must match the type exactly, so first cast the literal
/// into the type of the batch we read from Parquet before sending it to the compute kernel.
fn try_cast_literal(
    literal: &Arc<dyn ArrowDatum + Send + Sync>,
    column_type: &DataType,
) -> std::result::Result<Arc<dyn ArrowDatum + Send + Sync>, ArrowError> {
    let literal_array = literal.get().0;

    // No cast required
    if literal_array.data_type() == column_type {
        return Ok(Arc::clone(literal));
    }

    let literal_array = cast(literal_array, column_type)?;
    Ok(Arc::new(Scalar::new(literal_array)))
}

#[cfg(test)]
mod tests {
    use std::collections::{HashMap, HashSet};
    use std::fs::File;
    use std::sync::Arc;

    use arrow_array::cast::AsArray;
    use arrow_array::{ArrayRef, LargeStringArray, RecordBatch, StringArray};
    use arrow_schema::{DataType, Field, Schema as ArrowSchema, TimeUnit};
    use futures::TryStreamExt;
    use parquet::arrow::arrow_reader::{RowSelection, RowSelector};
    use parquet::arrow::{ArrowWriter, ProjectionMask};
    use parquet::basic::Compression;
    use parquet::file::metadata::{ColumnChunkMetaData, RowGroupMetaData};
    use parquet::file::properties::WriterProperties;
    use parquet::schema::parser::parse_message_type;
    use parquet::schema::types::{SchemaDescPtr, SchemaDescriptor};
    use roaring::RoaringTreemap;
    use tempfile::TempDir;

    use crate::ErrorKind;
    use crate::arrow::reader::{CollectFieldIdVisitor, PARQUET_FIELD_ID_META_KEY};
    use crate::arrow::{ArrowReader, ArrowReaderBuilder};
    use crate::delete_vector::DeleteVector;
    use crate::expr::visitors::bound_predicate_visitor::visit;
    use crate::expr::{Bind, Predicate, Reference};
    use crate::io::FileIO;
    use crate::scan::{FileScanTask, FileScanTaskDeleteFile, FileScanTaskStream};
    use crate::spec::{
        DataContentType, DataFileFormat, Datum, NestedField, PrimitiveType, Schema, SchemaRef, Type,
    };

    fn table_schema_simple() -> SchemaRef {
        Arc::new(
            Schema::builder()
                .with_schema_id(1)
                .with_identifier_field_ids(vec![2])
                .with_fields(vec![
                    NestedField::optional(1, "foo", Type::Primitive(PrimitiveType::String)).into(),
                    NestedField::required(2, "bar", Type::Primitive(PrimitiveType::Int)).into(),
                    NestedField::optional(3, "baz", Type::Primitive(PrimitiveType::Boolean)).into(),
                    NestedField::optional(4, "qux", Type::Primitive(PrimitiveType::Float)).into(),
                ])
                .build()
                .unwrap(),
        )
    }

    #[test]
    fn test_collect_field_id() {
        let schema = table_schema_simple();
        let expr = Reference::new("qux").is_null();
        let bound_expr = expr.bind(schema, true).unwrap();

        let mut visitor = CollectFieldIdVisitor {
            field_ids: HashSet::default(),
        };
        visit(&mut visitor, &bound_expr).unwrap();

        let mut expected = HashSet::default();
        expected.insert(4_i32);

        assert_eq!(visitor.field_ids, expected);
    }

    #[test]
    fn test_collect_field_id_with_and() {
        let schema = table_schema_simple();
        let expr = Reference::new("qux")
            .is_null()
            .and(Reference::new("baz").is_null());
        let bound_expr = expr.bind(schema, true).unwrap();

        let mut visitor = CollectFieldIdVisitor {
            field_ids: HashSet::default(),
        };
        visit(&mut visitor, &bound_expr).unwrap();

        let mut expected = HashSet::default();
        expected.insert(4_i32);
        expected.insert(3);

        assert_eq!(visitor.field_ids, expected);
    }

    #[test]
    fn test_collect_field_id_with_or() {
        let schema = table_schema_simple();
        let expr = Reference::new("qux")
            .is_null()
            .or(Reference::new("baz").is_null());
        let bound_expr = expr.bind(schema, true).unwrap();

        let mut visitor = CollectFieldIdVisitor {
            field_ids: HashSet::default(),
        };
        visit(&mut visitor, &bound_expr).unwrap();

        let mut expected = HashSet::default();
        expected.insert(4_i32);
        expected.insert(3);

        assert_eq!(visitor.field_ids, expected);
    }

    #[test]
    fn test_arrow_projection_mask() {
        let schema = Arc::new(
            Schema::builder()
                .with_schema_id(1)
                .with_identifier_field_ids(vec![1])
                .with_fields(vec![
                    NestedField::required(1, "c1", Type::Primitive(PrimitiveType::String)).into(),
                    NestedField::optional(2, "c2", Type::Primitive(PrimitiveType::Int)).into(),
                    NestedField::optional(
                        3,
                        "c3",
                        Type::Primitive(PrimitiveType::Decimal {
                            precision: 38,
                            scale: 3,
                        }),
                    )
                    .into(),
                ])
                .build()
                .unwrap(),
        );
        let arrow_schema = Arc::new(ArrowSchema::new(vec![
            Field::new("c1", DataType::Utf8, false).with_metadata(HashMap::from([(
                PARQUET_FIELD_ID_META_KEY.to_string(),
                "1".to_string(),
            )])),
            // Type not supported
            Field::new("c2", DataType::Duration(TimeUnit::Microsecond), true).with_metadata(
                HashMap::from([(PARQUET_FIELD_ID_META_KEY.to_string(), "2".to_string())]),
            ),
            // Precision is beyond the supported range
            Field::new("c3", DataType::Decimal128(39, 3), true).with_metadata(HashMap::from([(
                PARQUET_FIELD_ID_META_KEY.to_string(),
                "3".to_string(),
            )])),
        ]));

        let message_type = "
message schema {
  required binary c1 (STRING) = 1;
  optional int32 c2 (INTEGER(8,true)) = 2;
  optional fixed_len_byte_array(17) c3 (DECIMAL(39,3)) = 3;
}
    ";
        let parquet_type = parse_message_type(message_type).expect("should parse schema");
        let parquet_schema = SchemaDescriptor::new(Arc::new(parquet_type));

        // Try projecting the fields c2 and c3 with the unsupported data types
        let err = ArrowReader::get_arrow_projection_mask(
            &[1, 2, 3],
            &schema,
            &parquet_schema,
            &arrow_schema,
            false,
        )
        .unwrap_err();

        assert_eq!(err.kind(), ErrorKind::DataInvalid);
        assert_eq!(
            err.to_string(),
            "DataInvalid => Unsupported Arrow data type: Duration(µs)".to_string()
        );

        // Omitting field c2, we still get an error due to c3 being selected
        let err = ArrowReader::get_arrow_projection_mask(
            &[1, 3],
            &schema,
            &parquet_schema,
            &arrow_schema,
            false,
        )
        .unwrap_err();

        assert_eq!(err.kind(), ErrorKind::DataInvalid);
        assert_eq!(
            err.to_string(),
            "DataInvalid => Failed to create decimal type, source: DataInvalid => Decimals with precision larger than 38 are not supported: 39".to_string()
        );

        // Finally avoid selecting fields with unsupported data types
        let mask = ArrowReader::get_arrow_projection_mask(
            &[1],
            &schema,
            &parquet_schema,
            &arrow_schema,
            false,
        )
        .expect("Some ProjectionMask");
        assert_eq!(mask, ProjectionMask::leaves(&parquet_schema, vec![0]));
    }

    #[tokio::test]
    async fn test_kleene_logic_or_behaviour() {
        // a IS NULL OR a = 'foo'
        let predicate = Reference::new("a")
            .is_null()
            .or(Reference::new("a").equal_to(Datum::string("foo")));

        // Table data: [NULL, "foo", "bar"]
        let data_for_col_a = vec![None, Some("foo".to_string()), Some("bar".to_string())];

        // Expected: [NULL, "foo"].
        let expected = vec![None, Some("foo".to_string())];

        let (file_io, schema, table_location, _temp_dir) =
            setup_kleene_logic(data_for_col_a, DataType::Utf8);
        let reader = ArrowReaderBuilder::new(file_io).build();

        let result_data = test_perform_read(predicate, schema, table_location, reader).await;

        assert_eq!(result_data, expected);
    }

    #[tokio::test]
    async fn test_kleene_logic_and_behaviour() {
        // a IS NOT NULL AND a != 'foo'
        let predicate = Reference::new("a")
            .is_not_null()
            .and(Reference::new("a").not_equal_to(Datum::string("foo")));

        // Table data: [NULL, "foo", "bar"]
        let data_for_col_a = vec![None, Some("foo".to_string()), Some("bar".to_string())];

        // Expected: ["bar"].
        let expected = vec![Some("bar".to_string())];

        let (file_io, schema, table_location, _temp_dir) =
            setup_kleene_logic(data_for_col_a, DataType::Utf8);
        let reader = ArrowReaderBuilder::new(file_io).build();

        let result_data = test_perform_read(predicate, schema, table_location, reader).await;

        assert_eq!(result_data, expected);
    }

    #[tokio::test]
    async fn test_predicate_cast_literal() {
        let predicates = vec![
            // a == 'foo'
            (Reference::new("a").equal_to(Datum::string("foo")), vec![
                Some("foo".to_string()),
            ]),
            // a != 'foo'
            (
                Reference::new("a").not_equal_to(Datum::string("foo")),
                vec![Some("bar".to_string())],
            ),
            // STARTS_WITH(a, 'foo')
            (Reference::new("a").starts_with(Datum::string("f")), vec![
                Some("foo".to_string()),
            ]),
            // NOT STARTS_WITH(a, 'foo')
            (
                Reference::new("a").not_starts_with(Datum::string("f")),
                vec![Some("bar".to_string())],
            ),
            // a < 'foo'
            (Reference::new("a").less_than(Datum::string("foo")), vec![
                Some("bar".to_string()),
            ]),
            // a <= 'foo'
            (
                Reference::new("a").less_than_or_equal_to(Datum::string("foo")),
                vec![Some("foo".to_string()), Some("bar".to_string())],
            ),
            // a > 'foo'
            (
                Reference::new("a").greater_than(Datum::string("bar")),
                vec![Some("foo".to_string())],
            ),
            // a >= 'foo'
            (
                Reference::new("a").greater_than_or_equal_to(Datum::string("foo")),
                vec![Some("foo".to_string())],
            ),
            // a IN ('foo', 'bar')
            (
                Reference::new("a").is_in([Datum::string("foo"), Datum::string("baz")]),
                vec![Some("foo".to_string())],
            ),
            // a NOT IN ('foo', 'bar')
            (
                Reference::new("a").is_not_in([Datum::string("foo"), Datum::string("baz")]),
                vec![Some("bar".to_string())],
            ),
        ];

        // Table data: ["foo", "bar"]
        let data_for_col_a = vec![Some("foo".to_string()), Some("bar".to_string())];

        let (file_io, schema, table_location, _temp_dir) =
            setup_kleene_logic(data_for_col_a, DataType::LargeUtf8);
        let reader = ArrowReaderBuilder::new(file_io).build();

        for (predicate, expected) in predicates {
            println!("testing predicate {predicate}");
            let result_data = test_perform_read(
                predicate.clone(),
                schema.clone(),
                table_location.clone(),
                reader.clone(),
            )
            .await;

            assert_eq!(result_data, expected, "predicate={predicate}");
        }
    }

    async fn test_perform_read(
        predicate: Predicate,
        schema: SchemaRef,
        table_location: String,
        reader: ArrowReader,
    ) -> Vec<Option<String>> {
        let tasks = Box::pin(futures::stream::iter(
            vec![Ok(FileScanTask {
                start: 0,
                length: 0,
                record_count: None,
                data_file_path: format!("{table_location}/1.parquet"),
                data_file_format: DataFileFormat::Parquet,
                schema: schema.clone(),
                project_field_ids: vec![1],
                predicate: Some(predicate.bind(schema, true).unwrap()),
                deletes: vec![],
                partition: None,
                partition_spec: None,
                name_mapping: None,
                case_sensitive: false,
            })]
            .into_iter(),
        )) as FileScanTaskStream;

        let result = reader
            .read(tasks)
            .unwrap()
            .try_collect::<Vec<RecordBatch>>()
            .await
            .unwrap();

        result[0].columns()[0]
            .as_string_opt::<i32>()
            .unwrap()
            .iter()
            .map(|v| v.map(ToOwned::to_owned))
            .collect::<Vec<_>>()
    }

    fn setup_kleene_logic(
        data_for_col_a: Vec<Option<String>>,
        col_a_type: DataType,
    ) -> (FileIO, SchemaRef, String, TempDir) {
        let schema = Arc::new(
            Schema::builder()
                .with_schema_id(1)
                .with_fields(vec![
                    NestedField::optional(1, "a", Type::Primitive(PrimitiveType::String)).into(),
                ])
                .build()
                .unwrap(),
        );

        let arrow_schema = Arc::new(ArrowSchema::new(vec![
            Field::new("a", col_a_type.clone(), true).with_metadata(HashMap::from([(
                PARQUET_FIELD_ID_META_KEY.to_string(),
                "1".to_string(),
            )])),
        ]));

        let tmp_dir = TempDir::new().unwrap();
        let table_location = tmp_dir.path().to_str().unwrap().to_string();

        let file_io = FileIO::from_path(&table_location).unwrap().build().unwrap();

        let col = match col_a_type {
            DataType::Utf8 => Arc::new(StringArray::from(data_for_col_a)) as ArrayRef,
            DataType::LargeUtf8 => Arc::new(LargeStringArray::from(data_for_col_a)) as ArrayRef,
            _ => panic!("unexpected col_a_type"),
        };

        let to_write = RecordBatch::try_new(arrow_schema.clone(), vec![col]).unwrap();

        // Write the Parquet files
        let props = WriterProperties::builder()
            .set_compression(Compression::SNAPPY)
            .build();

        let file = File::create(format!("{table_location}/1.parquet")).unwrap();
        let mut writer =
            ArrowWriter::try_new(file, to_write.schema(), Some(props.clone())).unwrap();

        writer.write(&to_write).expect("Writing batch");

        // writer must be closed to write footer
        writer.close().unwrap();

        (file_io, schema, table_location, tmp_dir)
    }

    #[test]
    fn test_build_deletes_row_selection() {
        let schema_descr = get_test_schema_descr();

        let mut columns = vec![];
        for ptr in schema_descr.columns() {
            let column = ColumnChunkMetaData::builder(ptr.clone()).build().unwrap();
            columns.push(column);
        }

        let row_groups_metadata = vec![
            build_test_row_group_meta(schema_descr.clone(), columns.clone(), 1000, 0),
            build_test_row_group_meta(schema_descr.clone(), columns.clone(), 500, 1),
            build_test_row_group_meta(schema_descr.clone(), columns.clone(), 500, 2),
            build_test_row_group_meta(schema_descr.clone(), columns.clone(), 1000, 3),
            build_test_row_group_meta(schema_descr.clone(), columns.clone(), 500, 4),
        ];

        let selected_row_groups = Some(vec![1, 3]);

        /* cases to cover:
           * {skip|select} {first|intermediate|last} {one row|multiple rows} in
             {first|intermediate|last} {skipped|selected} row group
           * row group selection disabled
        */

        let positional_deletes = RoaringTreemap::from_iter(&[
            1, // in skipped rg 0, should be ignored
            3, // run of three consecutive items in skipped rg0
            4, 5, 998, // two consecutive items at end of skipped rg0
            999, 1000, // solitary row at start of selected rg1 (1, 9)
            1010, // run of 3 rows in selected rg1
            1011, 1012, // (3, 485)
            1498, // run of two items at end of selected rg1
            1499, 1500, // run of two items at start of skipped rg2
            1501, 1600, // should ignore, in skipped rg2
            1999, // single row at end of skipped rg2
            2000, // run of two items at start of selected rg3
            2001, // (4, 98)
            2100, // single row in selected row group 3 (1, 99)
            2200, // run of 3 consecutive rows in selected row group 3
            2201, 2202, // (3, 796)
            2999, // single item at end of selected rg3 (1)
            3000, // single item at start of skipped rg4
        ]);

        let positional_deletes = DeleteVector::new(positional_deletes);

        // using selected row groups 1 and 3
        let result = ArrowReader::build_deletes_row_selection(
            &row_groups_metadata,
            &selected_row_groups,
            &positional_deletes,
        )
        .unwrap();

        let expected = RowSelection::from(vec![
            RowSelector::skip(1),
            RowSelector::select(9),
            RowSelector::skip(3),
            RowSelector::select(485),
            RowSelector::skip(4),
            RowSelector::select(98),
            RowSelector::skip(1),
            RowSelector::select(99),
            RowSelector::skip(3),
            RowSelector::select(796),
            RowSelector::skip(1),
        ]);

        assert_eq!(result, expected);

        // selecting all row groups
        let result = ArrowReader::build_deletes_row_selection(
            &row_groups_metadata,
            &None,
            &positional_deletes,
        )
        .unwrap();

        let expected = RowSelection::from(vec![
            RowSelector::select(1),
            RowSelector::skip(1),
            RowSelector::select(1),
            RowSelector::skip(3),
            RowSelector::select(992),
            RowSelector::skip(3),
            RowSelector::select(9),
            RowSelector::skip(3),
            RowSelector::select(485),
            RowSelector::skip(4),
            RowSelector::select(98),
            RowSelector::skip(1),
            RowSelector::select(398),
            RowSelector::skip(3),
            RowSelector::select(98),
            RowSelector::skip(1),
            RowSelector::select(99),
            RowSelector::skip(3),
            RowSelector::select(796),
            RowSelector::skip(2),
            RowSelector::select(499),
        ]);

        assert_eq!(result, expected);
    }

    fn build_test_row_group_meta(
        schema_descr: SchemaDescPtr,
        columns: Vec<ColumnChunkMetaData>,
        num_rows: i64,
        ordinal: i16,
    ) -> RowGroupMetaData {
        RowGroupMetaData::builder(schema_descr.clone())
            .set_num_rows(num_rows)
            .set_total_byte_size(2000)
            .set_column_metadata(columns)
            .set_ordinal(ordinal)
            .build()
            .unwrap()
    }

    fn get_test_schema_descr() -> SchemaDescPtr {
        use parquet::schema::types::Type as SchemaType;

        let schema = SchemaType::group_type_builder("schema")
            .with_fields(vec![
                Arc::new(
                    SchemaType::primitive_type_builder("a", parquet::basic::Type::INT32)
                        .build()
                        .unwrap(),
                ),
                Arc::new(
                    SchemaType::primitive_type_builder("b", parquet::basic::Type::INT32)
                        .build()
                        .unwrap(),
                ),
            ])
            .build()
            .unwrap();

        Arc::new(SchemaDescriptor::new(Arc::new(schema)))
    }

    /// Verifies that file splits respect byte ranges and only read specific row groups.
    #[tokio::test]
    async fn test_file_splits_respect_byte_ranges() {
        use arrow_array::Int32Array;
        use parquet::file::reader::{FileReader, SerializedFileReader};

        let schema = Arc::new(
            Schema::builder()
                .with_schema_id(1)
                .with_fields(vec![
                    NestedField::required(1, "id", Type::Primitive(PrimitiveType::Int)).into(),
                ])
                .build()
                .unwrap(),
        );

        let arrow_schema = Arc::new(ArrowSchema::new(vec![
            Field::new("id", DataType::Int32, false).with_metadata(HashMap::from([(
                PARQUET_FIELD_ID_META_KEY.to_string(),
                "1".to_string(),
            )])),
        ]));

        let tmp_dir = TempDir::new().unwrap();
        let table_location = tmp_dir.path().to_str().unwrap().to_string();
        let file_path = format!("{table_location}/multi_row_group.parquet");

        // Force each batch into its own row group for testing byte range filtering.
        let batch1 = RecordBatch::try_new(arrow_schema.clone(), vec![Arc::new(Int32Array::from(
            (0..100).collect::<Vec<i32>>(),
        ))])
        .unwrap();
        let batch2 = RecordBatch::try_new(arrow_schema.clone(), vec![Arc::new(Int32Array::from(
            (100..200).collect::<Vec<i32>>(),
        ))])
        .unwrap();
        let batch3 = RecordBatch::try_new(arrow_schema.clone(), vec![Arc::new(Int32Array::from(
            (200..300).collect::<Vec<i32>>(),
        ))])
        .unwrap();

        let props = WriterProperties::builder()
            .set_compression(Compression::SNAPPY)
            .set_max_row_group_size(100)
            .build();

        let file = File::create(&file_path).unwrap();
        let mut writer = ArrowWriter::try_new(file, arrow_schema.clone(), Some(props)).unwrap();
        writer.write(&batch1).expect("Writing batch 1");
        writer.write(&batch2).expect("Writing batch 2");
        writer.write(&batch3).expect("Writing batch 3");
        writer.close().unwrap();

        // Read the file metadata to get row group byte positions
        let file = File::open(&file_path).unwrap();
        let reader = SerializedFileReader::new(file).unwrap();
        let metadata = reader.metadata();

        println!("File has {} row groups", metadata.num_row_groups());
        assert_eq!(metadata.num_row_groups(), 3, "Expected 3 row groups");

        // Get byte positions for each row group
        let row_group_0 = metadata.row_group(0);
        let row_group_1 = metadata.row_group(1);
        let row_group_2 = metadata.row_group(2);

        let rg0_start = 4u64; // Parquet files start with 4-byte magic "PAR1"
        let rg1_start = rg0_start + row_group_0.compressed_size() as u64;
        let rg2_start = rg1_start + row_group_1.compressed_size() as u64;
        let file_end = rg2_start + row_group_2.compressed_size() as u64;

        println!(
            "Row group 0: {} rows, starts at byte {}, {} bytes compressed",
            row_group_0.num_rows(),
            rg0_start,
            row_group_0.compressed_size()
        );
        println!(
            "Row group 1: {} rows, starts at byte {}, {} bytes compressed",
            row_group_1.num_rows(),
            rg1_start,
            row_group_1.compressed_size()
        );
        println!(
            "Row group 2: {} rows, starts at byte {}, {} bytes compressed",
            row_group_2.num_rows(),
            rg2_start,
            row_group_2.compressed_size()
        );

        let file_io = FileIO::from_path(&table_location).unwrap().build().unwrap();
        let reader = ArrowReaderBuilder::new(file_io).build();

        // Task 1: read only the first row group
        let task1 = FileScanTask {
            start: rg0_start,
            length: row_group_0.compressed_size() as u64,
            record_count: Some(100),
            data_file_path: file_path.clone(),
            data_file_format: DataFileFormat::Parquet,
            schema: schema.clone(),
            project_field_ids: vec![1],
            predicate: None,
            deletes: vec![],
            partition: None,
            partition_spec: None,
            name_mapping: None,
            case_sensitive: false,
        };

        // Task 2: read the second and third row groups
        let task2 = FileScanTask {
            start: rg1_start,
            length: file_end - rg1_start,
            record_count: Some(200),
            data_file_path: file_path.clone(),
            data_file_format: DataFileFormat::Parquet,
            schema: schema.clone(),
            project_field_ids: vec![1],
            predicate: None,
            deletes: vec![],
            partition: None,
            partition_spec: None,
            name_mapping: None,
            case_sensitive: false,
        };

        let tasks1 = Box::pin(futures::stream::iter(vec![Ok(task1)])) as FileScanTaskStream;
        let result1 = reader
            .clone()
            .read(tasks1)
            .unwrap()
            .try_collect::<Vec<RecordBatch>>()
            .await
            .unwrap();

        let total_rows_task1: usize = result1.iter().map(|b| b.num_rows()).sum();
        println!(
            "Task 1 (bytes {}-{}) returned {} rows",
            rg0_start,
            rg0_start + row_group_0.compressed_size() as u64,
            total_rows_task1
        );

        let tasks2 = Box::pin(futures::stream::iter(vec![Ok(task2)])) as FileScanTaskStream;
        let result2 = reader
            .read(tasks2)
            .unwrap()
            .try_collect::<Vec<RecordBatch>>()
            .await
            .unwrap();

        let total_rows_task2: usize = result2.iter().map(|b| b.num_rows()).sum();
        println!("Task 2 (bytes {rg1_start}-{file_end}) returned {total_rows_task2} rows");

        assert_eq!(
            total_rows_task1, 100,
            "Task 1 should read only the first row group (100 rows), but got {total_rows_task1} rows"
        );

        assert_eq!(
            total_rows_task2, 200,
            "Task 2 should read only the second+third row groups (200 rows), but got {total_rows_task2} rows"
        );

        // Verify the actual data values are correct (not just the row count)
        if total_rows_task1 > 0 {
            let first_batch = &result1[0];
            let id_col = first_batch
                .column(0)
                .as_primitive::<arrow_array::types::Int32Type>();
            let first_val = id_col.value(0);
            let last_val = id_col.value(id_col.len() - 1);
            println!("Task 1 data range: {first_val} to {last_val}");

            assert_eq!(first_val, 0, "Task 1 should start with id=0");
            assert_eq!(last_val, 99, "Task 1 should end with id=99");
        }

        if total_rows_task2 > 0 {
            let first_batch = &result2[0];
            let id_col = first_batch
                .column(0)
                .as_primitive::<arrow_array::types::Int32Type>();
            let first_val = id_col.value(0);
            println!("Task 2 first value: {first_val}");

            assert_eq!(first_val, 100, "Task 2 should start with id=100, not id=0");
        }
    }

    /// Test schema evolution: reading old Parquet file (with only column 'a')
    /// using a newer table schema (with columns 'a' and 'b').
    /// This tests that:
    /// 1. get_arrow_projection_mask allows missing columns
    /// 2. RecordBatchTransformer adds missing column 'b' with NULL values
    #[tokio::test]
    async fn test_schema_evolution_add_column() {
        use arrow_array::{Array, Int32Array};

        // New table schema: columns 'a' and 'b' (b was added later, file only has 'a')
        let new_schema = Arc::new(
            Schema::builder()
                .with_schema_id(2)
                .with_fields(vec![
                    NestedField::required(1, "a", Type::Primitive(PrimitiveType::Int)).into(),
                    NestedField::optional(2, "b", Type::Primitive(PrimitiveType::Int)).into(),
                ])
                .build()
                .unwrap(),
        );

        // Create Arrow schema for old Parquet file (only has column 'a')
        let arrow_schema_old = Arc::new(ArrowSchema::new(vec![
            Field::new("a", DataType::Int32, false).with_metadata(HashMap::from([(
                PARQUET_FIELD_ID_META_KEY.to_string(),
                "1".to_string(),
            )])),
        ]));

        // Write old Parquet file with only column 'a'
        let tmp_dir = TempDir::new().unwrap();
        let table_location = tmp_dir.path().to_str().unwrap().to_string();
        let file_io = FileIO::from_path(&table_location).unwrap().build().unwrap();

        let data_a = Arc::new(Int32Array::from(vec![1, 2, 3])) as ArrayRef;
        let to_write = RecordBatch::try_new(arrow_schema_old.clone(), vec![data_a]).unwrap();

        let props = WriterProperties::builder()
            .set_compression(Compression::SNAPPY)
            .build();
        let file = File::create(format!("{table_location}/old_file.parquet")).unwrap();
        let mut writer = ArrowWriter::try_new(file, to_write.schema(), Some(props)).unwrap();
        writer.write(&to_write).expect("Writing batch");
        writer.close().unwrap();

        // Read the old Parquet file using the NEW schema (with column 'b')
        let reader = ArrowReaderBuilder::new(file_io).build();
        let tasks = Box::pin(futures::stream::iter(
            vec![Ok(FileScanTask {
                start: 0,
                length: 0,
                record_count: None,
                data_file_path: format!("{table_location}/old_file.parquet"),
                data_file_format: DataFileFormat::Parquet,
                schema: new_schema.clone(),
                project_field_ids: vec![1, 2], // Request both columns 'a' and 'b'
                predicate: None,
                deletes: vec![],
                partition: None,
                partition_spec: None,
                name_mapping: None,
                case_sensitive: false,
            })]
            .into_iter(),
        )) as FileScanTaskStream;

        let result = reader
            .read(tasks)
            .unwrap()
            .try_collect::<Vec<RecordBatch>>()
            .await
            .unwrap();

        // Verify we got the correct data
        assert_eq!(result.len(), 1);
        let batch = &result[0];

        // Should have 2 columns now
        assert_eq!(batch.num_columns(), 2);
        assert_eq!(batch.num_rows(), 3);

        // Column 'a' should have the original data
        let col_a = batch
            .column(0)
            .as_primitive::<arrow_array::types::Int32Type>();
        assert_eq!(col_a.values(), &[1, 2, 3]);

        // Column 'b' should be all NULLs (it didn't exist in the old file)
        let col_b = batch
            .column(1)
            .as_primitive::<arrow_array::types::Int32Type>();
        assert_eq!(col_b.null_count(), 3);
        assert!(col_b.is_null(0));
        assert!(col_b.is_null(1));
        assert!(col_b.is_null(2));
    }

    /// Test for bug where position deletes in later row groups are not applied correctly.
    ///
    /// When a file has multiple row groups and a position delete targets a row in a later
    /// row group, the `build_deletes_row_selection` function had a bug where it would
    /// fail to increment `current_row_group_base_idx` when skipping row groups.
    ///
    /// This test creates:
    /// - A data file with 200 rows split into 2 row groups (0-99, 100-199)
    /// - A position delete file that deletes row 199 (last row in second row group)
    ///
    /// Expected behavior: Should return 199 rows (with id=200 deleted)
    /// Bug behavior: Returns 200 rows (delete is not applied)
    ///
    /// This bug was discovered while running Apache Spark + Apache Iceberg integration tests
    /// through DataFusion Comet. The following Iceberg Java tests failed due to this bug:
    /// - `org.apache.iceberg.spark.extensions.TestMergeOnReadDelete::testDeleteWithMultipleRowGroupsParquet`
    /// - `org.apache.iceberg.spark.extensions.TestMergeOnReadUpdate::testUpdateWithMultipleRowGroupsParquet`
    #[tokio::test]
    async fn test_position_delete_across_multiple_row_groups() {
        use arrow_array::{Int32Array, Int64Array};
        use parquet::file::reader::{FileReader, SerializedFileReader};

        // Field IDs for positional delete schema
        const FIELD_ID_POSITIONAL_DELETE_FILE_PATH: u64 = 2147483546;
        const FIELD_ID_POSITIONAL_DELETE_POS: u64 = 2147483545;

        let tmp_dir = TempDir::new().unwrap();
        let table_location = tmp_dir.path().to_str().unwrap().to_string();

        // Create table schema with a single 'id' column
        let table_schema = Arc::new(
            Schema::builder()
                .with_schema_id(1)
                .with_fields(vec![
                    NestedField::required(1, "id", Type::Primitive(PrimitiveType::Int)).into(),
                ])
                .build()
                .unwrap(),
        );

        let arrow_schema = Arc::new(ArrowSchema::new(vec![
            Field::new("id", DataType::Int32, false).with_metadata(HashMap::from([(
                PARQUET_FIELD_ID_META_KEY.to_string(),
                "1".to_string(),
            )])),
        ]));

        // Step 1: Create data file with 200 rows in 2 row groups
        // Row group 0: rows 0-99 (ids 1-100)
        // Row group 1: rows 100-199 (ids 101-200)
        let data_file_path = format!("{table_location}/data.parquet");

        let batch1 = RecordBatch::try_new(arrow_schema.clone(), vec![Arc::new(
            Int32Array::from_iter_values(1..=100),
        )])
        .unwrap();

        let batch2 = RecordBatch::try_new(arrow_schema.clone(), vec![Arc::new(
            Int32Array::from_iter_values(101..=200),
        )])
        .unwrap();

        // Force each batch into its own row group
        let props = WriterProperties::builder()
            .set_compression(Compression::SNAPPY)
            .set_max_row_group_size(100)
            .build();

        let file = File::create(&data_file_path).unwrap();
        let mut writer = ArrowWriter::try_new(file, arrow_schema.clone(), Some(props)).unwrap();
        writer.write(&batch1).expect("Writing batch 1");
        writer.write(&batch2).expect("Writing batch 2");
        writer.close().unwrap();

        // Verify we created 2 row groups
        let verify_file = File::open(&data_file_path).unwrap();
        let verify_reader = SerializedFileReader::new(verify_file).unwrap();
        assert_eq!(
            verify_reader.metadata().num_row_groups(),
            2,
            "Should have 2 row groups"
        );

        // Step 2: Create position delete file that deletes row 199 (id=200, last row in row group 1)
        let delete_file_path = format!("{table_location}/deletes.parquet");

        let delete_schema = Arc::new(ArrowSchema::new(vec![
            Field::new("file_path", DataType::Utf8, false).with_metadata(HashMap::from([(
                PARQUET_FIELD_ID_META_KEY.to_string(),
                FIELD_ID_POSITIONAL_DELETE_FILE_PATH.to_string(),
            )])),
            Field::new("pos", DataType::Int64, false).with_metadata(HashMap::from([(
                PARQUET_FIELD_ID_META_KEY.to_string(),
                FIELD_ID_POSITIONAL_DELETE_POS.to_string(),
            )])),
        ]));

        // Delete row at position 199 (0-indexed, so it's the last row: id=200)
        let delete_batch = RecordBatch::try_new(delete_schema.clone(), vec![
            Arc::new(StringArray::from_iter_values(vec![data_file_path.clone()])),
            Arc::new(Int64Array::from_iter_values(vec![199i64])),
        ])
        .unwrap();

        let delete_props = WriterProperties::builder()
            .set_compression(Compression::SNAPPY)
            .build();

        let delete_file = File::create(&delete_file_path).unwrap();
        let mut delete_writer =
            ArrowWriter::try_new(delete_file, delete_schema, Some(delete_props)).unwrap();
        delete_writer.write(&delete_batch).unwrap();
        delete_writer.close().unwrap();

        // Step 3: Read the data file with the delete applied
        let file_io = FileIO::from_path(&table_location).unwrap().build().unwrap();
        let reader = ArrowReaderBuilder::new(file_io).build();

        let task = FileScanTask {
            start: 0,
            length: 0,
            record_count: Some(200),
            data_file_path: data_file_path.clone(),
            data_file_format: DataFileFormat::Parquet,
            schema: table_schema.clone(),
            project_field_ids: vec![1],
            predicate: None,
            deletes: vec![FileScanTaskDeleteFile {
                file_path: delete_file_path,
                file_type: DataContentType::PositionDeletes,
                partition_spec_id: 0,
                equality_ids: None,
            }],
            partition: None,
            partition_spec: None,
            name_mapping: None,
            case_sensitive: false,
        };

        let tasks = Box::pin(futures::stream::iter(vec![Ok(task)])) as FileScanTaskStream;
        let result = reader
            .read(tasks)
            .unwrap()
            .try_collect::<Vec<RecordBatch>>()
            .await
            .unwrap();

        // Step 4: Verify we got 199 rows (not 200)
        let total_rows: usize = result.iter().map(|b| b.num_rows()).sum();

        println!("Total rows read: {total_rows}");
        println!("Expected: 199 rows (deleted row 199 which had id=200)");

        // This assertion will FAIL before the fix and PASS after the fix
        assert_eq!(
            total_rows, 199,
            "Expected 199 rows after deleting row 199, but got {total_rows} rows. \
             The bug causes position deletes in later row groups to be ignored."
        );

        // Verify the deleted row (id=200) is not present
        let all_ids: Vec<i32> = result
            .iter()
            .flat_map(|batch| {
                batch
                    .column(0)
                    .as_primitive::<arrow_array::types::Int32Type>()
                    .values()
                    .iter()
                    .copied()
            })
            .collect();

        assert!(
            !all_ids.contains(&200),
            "Row with id=200 should be deleted but was found in results"
        );

        // Verify we have all other ids (1-199)
        let expected_ids: Vec<i32> = (1..=199).collect();
        assert_eq!(
            all_ids, expected_ids,
            "Should have ids 1-199 but got different values"
        );
    }

    /// Test for bug where position deletes are lost when skipping unselected row groups.
    ///
    /// This is a variant of `test_position_delete_across_multiple_row_groups` that exercises
    /// the row group selection code path (`selected_row_groups: Some([...])`).
    ///
    /// When a file has multiple row groups and only some are selected for reading,
    /// the `build_deletes_row_selection` function must correctly skip over deletes in
    /// unselected row groups WITHOUT consuming deletes that belong to selected row groups.
    ///
    /// This test creates:
    /// - A data file with 200 rows split into 2 row groups (0-99, 100-199)
    /// - A position delete file that deletes row 199 (last row in second row group)
    /// - Row group selection that reads ONLY row group 1 (rows 100-199)
    ///
    /// Expected behavior: Should return 99 rows (with row 199 deleted)
    /// Bug behavior: Returns 100 rows (delete is lost when skipping row group 0)
    ///
    /// The bug occurs when processing row group 0 (unselected):
    /// ```rust
    /// delete_vector_iter.advance_to(next_row_group_base_idx); // Position at first delete >= 100
    /// next_deleted_row_idx_opt = delete_vector_iter.next(); // BUG: Consumes delete at 199!
    /// ```
    ///
    /// The fix is to NOT call `next()` after `advance_to()` when skipping unselected row groups,
    /// because `advance_to()` already positions the iterator correctly without consuming elements.
    #[tokio::test]
    async fn test_position_delete_with_row_group_selection() {
        use arrow_array::{Int32Array, Int64Array};
        use parquet::file::reader::{FileReader, SerializedFileReader};

        // Field IDs for positional delete schema
        const FIELD_ID_POSITIONAL_DELETE_FILE_PATH: u64 = 2147483546;
        const FIELD_ID_POSITIONAL_DELETE_POS: u64 = 2147483545;

        let tmp_dir = TempDir::new().unwrap();
        let table_location = tmp_dir.path().to_str().unwrap().to_string();

        // Create table schema with a single 'id' column
        let table_schema = Arc::new(
            Schema::builder()
                .with_schema_id(1)
                .with_fields(vec![
                    NestedField::required(1, "id", Type::Primitive(PrimitiveType::Int)).into(),
                ])
                .build()
                .unwrap(),
        );

        let arrow_schema = Arc::new(ArrowSchema::new(vec![
            Field::new("id", DataType::Int32, false).with_metadata(HashMap::from([(
                PARQUET_FIELD_ID_META_KEY.to_string(),
                "1".to_string(),
            )])),
        ]));

        // Step 1: Create data file with 200 rows in 2 row groups
        // Row group 0: rows 0-99 (ids 1-100)
        // Row group 1: rows 100-199 (ids 101-200)
        let data_file_path = format!("{table_location}/data.parquet");

        let batch1 = RecordBatch::try_new(arrow_schema.clone(), vec![Arc::new(
            Int32Array::from_iter_values(1..=100),
        )])
        .unwrap();

        let batch2 = RecordBatch::try_new(arrow_schema.clone(), vec![Arc::new(
            Int32Array::from_iter_values(101..=200),
        )])
        .unwrap();

        // Force each batch into its own row group
        let props = WriterProperties::builder()
            .set_compression(Compression::SNAPPY)
            .set_max_row_group_size(100)
            .build();

        let file = File::create(&data_file_path).unwrap();
        let mut writer = ArrowWriter::try_new(file, arrow_schema.clone(), Some(props)).unwrap();
        writer.write(&batch1).expect("Writing batch 1");
        writer.write(&batch2).expect("Writing batch 2");
        writer.close().unwrap();

        // Verify we created 2 row groups
        let verify_file = File::open(&data_file_path).unwrap();
        let verify_reader = SerializedFileReader::new(verify_file).unwrap();
        assert_eq!(
            verify_reader.metadata().num_row_groups(),
            2,
            "Should have 2 row groups"
        );

        // Step 2: Create position delete file that deletes row 199 (id=200, last row in row group 1)
        let delete_file_path = format!("{table_location}/deletes.parquet");

        let delete_schema = Arc::new(ArrowSchema::new(vec![
            Field::new("file_path", DataType::Utf8, false).with_metadata(HashMap::from([(
                PARQUET_FIELD_ID_META_KEY.to_string(),
                FIELD_ID_POSITIONAL_DELETE_FILE_PATH.to_string(),
            )])),
            Field::new("pos", DataType::Int64, false).with_metadata(HashMap::from([(
                PARQUET_FIELD_ID_META_KEY.to_string(),
                FIELD_ID_POSITIONAL_DELETE_POS.to_string(),
            )])),
        ]));

        // Delete row at position 199 (0-indexed, so it's the last row: id=200)
        let delete_batch = RecordBatch::try_new(delete_schema.clone(), vec![
            Arc::new(StringArray::from_iter_values(vec![data_file_path.clone()])),
            Arc::new(Int64Array::from_iter_values(vec![199i64])),
        ])
        .unwrap();

        let delete_props = WriterProperties::builder()
            .set_compression(Compression::SNAPPY)
            .build();

        let delete_file = File::create(&delete_file_path).unwrap();
        let mut delete_writer =
            ArrowWriter::try_new(delete_file, delete_schema, Some(delete_props)).unwrap();
        delete_writer.write(&delete_batch).unwrap();
        delete_writer.close().unwrap();

        // Step 3: Get byte ranges to read ONLY row group 1 (rows 100-199)
        // This exercises the row group selection code path where row group 0 is skipped
        let metadata_file = File::open(&data_file_path).unwrap();
        let metadata_reader = SerializedFileReader::new(metadata_file).unwrap();
        let metadata = metadata_reader.metadata();

        let row_group_0 = metadata.row_group(0);
        let row_group_1 = metadata.row_group(1);

        let rg0_start = 4u64; // Parquet files start with 4-byte magic "PAR1"
        let rg1_start = rg0_start + row_group_0.compressed_size() as u64;
        let rg1_length = row_group_1.compressed_size() as u64;

        println!(
            "Row group 0: starts at byte {}, {} bytes compressed",
            rg0_start,
            row_group_0.compressed_size()
        );
        println!(
            "Row group 1: starts at byte {}, {} bytes compressed",
            rg1_start,
            row_group_1.compressed_size()
        );

        let file_io = FileIO::from_path(&table_location).unwrap().build().unwrap();
        let reader = ArrowReaderBuilder::new(file_io).build();

        // Create FileScanTask that reads ONLY row group 1 via byte range filtering
        let task = FileScanTask {
            start: rg1_start,
            length: rg1_length,
            record_count: Some(100), // Row group 1 has 100 rows
            data_file_path: data_file_path.clone(),
            data_file_format: DataFileFormat::Parquet,
            schema: table_schema.clone(),
            project_field_ids: vec![1],
            predicate: None,
            deletes: vec![FileScanTaskDeleteFile {
                file_path: delete_file_path,
                file_type: DataContentType::PositionDeletes,
                partition_spec_id: 0,
                equality_ids: None,
            }],
            partition: None,
            partition_spec: None,
            name_mapping: None,
            case_sensitive: false,
        };

        let tasks = Box::pin(futures::stream::iter(vec![Ok(task)])) as FileScanTaskStream;
        let result = reader
            .read(tasks)
            .unwrap()
            .try_collect::<Vec<RecordBatch>>()
            .await
            .unwrap();

        // Step 4: Verify we got 99 rows (not 100)
        // Row group 1 has 100 rows (ids 101-200), minus 1 delete (id=200) = 99 rows
        let total_rows: usize = result.iter().map(|b| b.num_rows()).sum();

        println!("Total rows read from row group 1: {total_rows}");
        println!("Expected: 99 rows (row group 1 has 100 rows, 1 delete at position 199)");

        // This assertion will FAIL before the fix and PASS after the fix
        assert_eq!(
            total_rows, 99,
            "Expected 99 rows from row group 1 after deleting position 199, but got {total_rows} rows. \
             The bug causes position deletes to be lost when advance_to() is followed by next() \
             when skipping unselected row groups."
        );

        // Verify the deleted row (id=200) is not present
        let all_ids: Vec<i32> = result
            .iter()
            .flat_map(|batch| {
                batch
                    .column(0)
                    .as_primitive::<arrow_array::types::Int32Type>()
                    .values()
                    .iter()
                    .copied()
            })
            .collect();

        assert!(
            !all_ids.contains(&200),
            "Row with id=200 should be deleted but was found in results"
        );

        // Verify we have ids 101-199 (not 101-200)
        let expected_ids: Vec<i32> = (101..=199).collect();
        assert_eq!(
            all_ids, expected_ids,
            "Should have ids 101-199 but got different values"
        );
    }
    /// Test for bug where stale cached delete causes infinite loop when skipping row groups.
    ///
    /// This test exposes the inverse scenario of `test_position_delete_with_row_group_selection`:
    /// - Position delete targets a row in the SKIPPED row group (not the selected one)
    /// - After calling advance_to(), the cached delete index is stale
    /// - Without updating the cache, the code enters an infinite loop
    ///
    /// This test creates:
    /// - A data file with 200 rows split into 2 row groups (0-99, 100-199)
    /// - A position delete file that deletes row 0 (first row in SKIPPED row group 0)
    /// - Row group selection that reads ONLY row group 1 (rows 100-199)
    ///
    /// The bug occurs when skipping row group 0:
    /// ```rust
    /// let mut next_deleted_row_idx_opt = delete_vector_iter.next(); // Some(0)
    /// // ... skip to row group 1 ...
    /// delete_vector_iter.advance_to(100); // Iterator advances past delete at 0
    /// // BUG: next_deleted_row_idx_opt is still Some(0) - STALE!
    /// // When processing row group 1:
    /// //   current_idx = 100, next_deleted_row_idx = 0, next_row_group_base_idx = 200
    /// //   Loop condition: 0 < 200 (true)
    /// //   But: current_idx (100) > next_deleted_row_idx (0)
    /// //   And: current_idx (100) != next_deleted_row_idx (0)
    /// //   Neither branch executes -> INFINITE LOOP!
    /// ```
    ///
    /// Expected behavior: Should return 100 rows (delete at 0 doesn't affect row group 1)
    /// Bug behavior: Infinite loop in build_deletes_row_selection
    #[tokio::test]
    async fn test_position_delete_in_skipped_row_group() {
        use arrow_array::{Int32Array, Int64Array};
        use parquet::file::reader::{FileReader, SerializedFileReader};

        // Field IDs for positional delete schema
        const FIELD_ID_POSITIONAL_DELETE_FILE_PATH: u64 = 2147483546;
        const FIELD_ID_POSITIONAL_DELETE_POS: u64 = 2147483545;

        let tmp_dir = TempDir::new().unwrap();
        let table_location = tmp_dir.path().to_str().unwrap().to_string();

        // Create table schema with a single 'id' column
        let table_schema = Arc::new(
            Schema::builder()
                .with_schema_id(1)
                .with_fields(vec![
                    NestedField::required(1, "id", Type::Primitive(PrimitiveType::Int)).into(),
                ])
                .build()
                .unwrap(),
        );

        let arrow_schema = Arc::new(ArrowSchema::new(vec![
            Field::new("id", DataType::Int32, false).with_metadata(HashMap::from([(
                PARQUET_FIELD_ID_META_KEY.to_string(),
                "1".to_string(),
            )])),
        ]));

        // Step 1: Create data file with 200 rows in 2 row groups
        // Row group 0: rows 0-99 (ids 1-100)
        // Row group 1: rows 100-199 (ids 101-200)
        let data_file_path = format!("{table_location}/data.parquet");

        let batch1 = RecordBatch::try_new(arrow_schema.clone(), vec![Arc::new(
            Int32Array::from_iter_values(1..=100),
        )])
        .unwrap();

        let batch2 = RecordBatch::try_new(arrow_schema.clone(), vec![Arc::new(
            Int32Array::from_iter_values(101..=200),
        )])
        .unwrap();

        // Force each batch into its own row group
        let props = WriterProperties::builder()
            .set_compression(Compression::SNAPPY)
            .set_max_row_group_size(100)
            .build();

        let file = File::create(&data_file_path).unwrap();
        let mut writer = ArrowWriter::try_new(file, arrow_schema.clone(), Some(props)).unwrap();
        writer.write(&batch1).expect("Writing batch 1");
        writer.write(&batch2).expect("Writing batch 2");
        writer.close().unwrap();

        // Verify we created 2 row groups
        let verify_file = File::open(&data_file_path).unwrap();
        let verify_reader = SerializedFileReader::new(verify_file).unwrap();
        assert_eq!(
            verify_reader.metadata().num_row_groups(),
            2,
            "Should have 2 row groups"
        );

        // Step 2: Create position delete file that deletes row 0 (id=1, first row in row group 0)
        let delete_file_path = format!("{table_location}/deletes.parquet");

        let delete_schema = Arc::new(ArrowSchema::new(vec![
            Field::new("file_path", DataType::Utf8, false).with_metadata(HashMap::from([(
                PARQUET_FIELD_ID_META_KEY.to_string(),
                FIELD_ID_POSITIONAL_DELETE_FILE_PATH.to_string(),
            )])),
            Field::new("pos", DataType::Int64, false).with_metadata(HashMap::from([(
                PARQUET_FIELD_ID_META_KEY.to_string(),
                FIELD_ID_POSITIONAL_DELETE_POS.to_string(),
            )])),
        ]));

        // Delete row at position 0 (0-indexed, so it's the first row: id=1)
        let delete_batch = RecordBatch::try_new(delete_schema.clone(), vec![
            Arc::new(StringArray::from_iter_values(vec![data_file_path.clone()])),
            Arc::new(Int64Array::from_iter_values(vec![0i64])),
        ])
        .unwrap();

        let delete_props = WriterProperties::builder()
            .set_compression(Compression::SNAPPY)
            .build();

        let delete_file = File::create(&delete_file_path).unwrap();
        let mut delete_writer =
            ArrowWriter::try_new(delete_file, delete_schema, Some(delete_props)).unwrap();
        delete_writer.write(&delete_batch).unwrap();
        delete_writer.close().unwrap();

        // Step 3: Get byte ranges to read ONLY row group 1 (rows 100-199)
        // This exercises the row group selection code path where row group 0 is skipped
        let metadata_file = File::open(&data_file_path).unwrap();
        let metadata_reader = SerializedFileReader::new(metadata_file).unwrap();
        let metadata = metadata_reader.metadata();

        let row_group_0 = metadata.row_group(0);
        let row_group_1 = metadata.row_group(1);

        let rg0_start = 4u64; // Parquet files start with 4-byte magic "PAR1"
        let rg1_start = rg0_start + row_group_0.compressed_size() as u64;
        let rg1_length = row_group_1.compressed_size() as u64;

        let file_io = FileIO::from_path(&table_location).unwrap().build().unwrap();
        let reader = ArrowReaderBuilder::new(file_io).build();

        // Create FileScanTask that reads ONLY row group 1 via byte range filtering
        let task = FileScanTask {
            start: rg1_start,
            length: rg1_length,
            record_count: Some(100), // Row group 1 has 100 rows
            data_file_path: data_file_path.clone(),
            data_file_format: DataFileFormat::Parquet,
            schema: table_schema.clone(),
            project_field_ids: vec![1],
            predicate: None,
            deletes: vec![FileScanTaskDeleteFile {
                file_path: delete_file_path,
                file_type: DataContentType::PositionDeletes,
                partition_spec_id: 0,
                equality_ids: None,
            }],
            partition: None,
            partition_spec: None,
            name_mapping: None,
            case_sensitive: false,
        };

        let tasks = Box::pin(futures::stream::iter(vec![Ok(task)])) as FileScanTaskStream;
        let result = reader
            .read(tasks)
            .unwrap()
            .try_collect::<Vec<RecordBatch>>()
            .await
            .unwrap();

        // Step 4: Verify we got 100 rows (all of row group 1)
        // The delete at position 0 is in row group 0, which is skipped, so it doesn't affect us
        let total_rows: usize = result.iter().map(|b| b.num_rows()).sum();

        assert_eq!(
            total_rows, 100,
            "Expected 100 rows from row group 1 (delete at position 0 is in skipped row group 0). \
             If this hangs or fails, it indicates the cached delete index was not updated after advance_to()."
        );

        // Verify we have all ids from row group 1 (101-200)
        let all_ids: Vec<i32> = result
            .iter()
            .flat_map(|batch| {
                batch
                    .column(0)
                    .as_primitive::<arrow_array::types::Int32Type>()
                    .values()
                    .iter()
                    .copied()
            })
            .collect();

        let expected_ids: Vec<i32> = (101..=200).collect();
        assert_eq!(
            all_ids, expected_ids,
            "Should have ids 101-200 (all of row group 1)"
        );
    }

    /// Test reading Parquet files without field ID metadata (e.g., migrated tables).
    /// This exercises the position-based fallback path.
    ///
    /// Corresponds to Java's ParquetSchemaUtil.addFallbackIds() + pruneColumnsFallback()
    /// in /parquet/src/main/java/org/apache/iceberg/parquet/ParquetSchemaUtil.java
    #[tokio::test]
    async fn test_read_parquet_file_without_field_ids() {
        let schema = Arc::new(
            Schema::builder()
                .with_schema_id(1)
                .with_fields(vec![
                    NestedField::required(1, "name", Type::Primitive(PrimitiveType::String)).into(),
                    NestedField::required(2, "age", Type::Primitive(PrimitiveType::Int)).into(),
                ])
                .build()
                .unwrap(),
        );

        // Parquet file from a migrated table - no field ID metadata
        let arrow_schema = Arc::new(ArrowSchema::new(vec![
            Field::new("name", DataType::Utf8, false),
            Field::new("age", DataType::Int32, false),
        ]));

        let tmp_dir = TempDir::new().unwrap();
        let table_location = tmp_dir.path().to_str().unwrap().to_string();
        let file_io = FileIO::from_path(&table_location).unwrap().build().unwrap();

        let name_data = vec!["Alice", "Bob", "Charlie"];
        let age_data = vec![30, 25, 35];

        use arrow_array::Int32Array;
        let name_col = Arc::new(StringArray::from(name_data.clone())) as ArrayRef;
        let age_col = Arc::new(Int32Array::from(age_data.clone())) as ArrayRef;

        let to_write = RecordBatch::try_new(arrow_schema.clone(), vec![name_col, age_col]).unwrap();

        let props = WriterProperties::builder()
            .set_compression(Compression::SNAPPY)
            .build();

        let file = File::create(format!("{table_location}/1.parquet")).unwrap();
        let mut writer = ArrowWriter::try_new(file, to_write.schema(), Some(props)).unwrap();

        writer.write(&to_write).expect("Writing batch");
        writer.close().unwrap();

        let reader = ArrowReaderBuilder::new(file_io).build();

        let tasks = Box::pin(futures::stream::iter(
            vec![Ok(FileScanTask {
                start: 0,
                length: 0,
                record_count: None,
                data_file_path: format!("{table_location}/1.parquet"),
                data_file_format: DataFileFormat::Parquet,
                schema: schema.clone(),
                project_field_ids: vec![1, 2],
                predicate: None,
                deletes: vec![],
                partition: None,
                partition_spec: None,
                name_mapping: None,
                case_sensitive: false,
            })]
            .into_iter(),
        )) as FileScanTaskStream;

        let result = reader
            .read(tasks)
            .unwrap()
            .try_collect::<Vec<RecordBatch>>()
            .await
            .unwrap();

        assert_eq!(result.len(), 1);
        let batch = &result[0];
        assert_eq!(batch.num_rows(), 3);
        assert_eq!(batch.num_columns(), 2);

        // Verify position-based mapping: field_id 1 → position 0, field_id 2 → position 1
        let name_array = batch.column(0).as_string::<i32>();
        assert_eq!(name_array.value(0), "Alice");
        assert_eq!(name_array.value(1), "Bob");
        assert_eq!(name_array.value(2), "Charlie");

        let age_array = batch
            .column(1)
            .as_primitive::<arrow_array::types::Int32Type>();
        assert_eq!(age_array.value(0), 30);
        assert_eq!(age_array.value(1), 25);
        assert_eq!(age_array.value(2), 35);
    }

    /// Test reading Parquet files without field IDs with partial projection.
    /// Only a subset of columns are requested, verifying position-based fallback
    /// handles column selection correctly.
    #[tokio::test]
    async fn test_read_parquet_without_field_ids_partial_projection() {
        use arrow_array::Int32Array;

        let schema = Arc::new(
            Schema::builder()
                .with_schema_id(1)
                .with_fields(vec![
                    NestedField::required(1, "col1", Type::Primitive(PrimitiveType::String)).into(),
                    NestedField::required(2, "col2", Type::Primitive(PrimitiveType::Int)).into(),
                    NestedField::required(3, "col3", Type::Primitive(PrimitiveType::String)).into(),
                    NestedField::required(4, "col4", Type::Primitive(PrimitiveType::Int)).into(),
                ])
                .build()
                .unwrap(),
        );

        let arrow_schema = Arc::new(ArrowSchema::new(vec![
            Field::new("col1", DataType::Utf8, false),
            Field::new("col2", DataType::Int32, false),
            Field::new("col3", DataType::Utf8, false),
            Field::new("col4", DataType::Int32, false),
        ]));

        let tmp_dir = TempDir::new().unwrap();
        let table_location = tmp_dir.path().to_str().unwrap().to_string();
        let file_io = FileIO::from_path(&table_location).unwrap().build().unwrap();

        let col1_data = Arc::new(StringArray::from(vec!["a", "b"])) as ArrayRef;
        let col2_data = Arc::new(Int32Array::from(vec![10, 20])) as ArrayRef;
        let col3_data = Arc::new(StringArray::from(vec!["c", "d"])) as ArrayRef;
        let col4_data = Arc::new(Int32Array::from(vec![30, 40])) as ArrayRef;

        let to_write = RecordBatch::try_new(arrow_schema.clone(), vec![
            col1_data, col2_data, col3_data, col4_data,
        ])
        .unwrap();

        let props = WriterProperties::builder()
            .set_compression(Compression::SNAPPY)
            .build();

        let file = File::create(format!("{table_location}/1.parquet")).unwrap();
        let mut writer = ArrowWriter::try_new(file, to_write.schema(), Some(props)).unwrap();

        writer.write(&to_write).expect("Writing batch");
        writer.close().unwrap();

        let reader = ArrowReaderBuilder::new(file_io).build();

        let tasks = Box::pin(futures::stream::iter(
            vec![Ok(FileScanTask {
                start: 0,
                length: 0,
                record_count: None,
                data_file_path: format!("{table_location}/1.parquet"),
                data_file_format: DataFileFormat::Parquet,
                schema: schema.clone(),
                project_field_ids: vec![1, 3],
                predicate: None,
                deletes: vec![],
                partition: None,
                partition_spec: None,
                name_mapping: None,
                case_sensitive: false,
            })]
            .into_iter(),
        )) as FileScanTaskStream;

        let result = reader
            .read(tasks)
            .unwrap()
            .try_collect::<Vec<RecordBatch>>()
            .await
            .unwrap();

        assert_eq!(result.len(), 1);
        let batch = &result[0];
        assert_eq!(batch.num_rows(), 2);
        assert_eq!(batch.num_columns(), 2);

        let col1_array = batch.column(0).as_string::<i32>();
        assert_eq!(col1_array.value(0), "a");
        assert_eq!(col1_array.value(1), "b");

        let col3_array = batch.column(1).as_string::<i32>();
        assert_eq!(col3_array.value(0), "c");
        assert_eq!(col3_array.value(1), "d");
    }

    /// Test reading Parquet files without field IDs with schema evolution.
    /// The Iceberg schema has more fields than the Parquet file, testing that
    /// missing columns are filled with NULLs.
    #[tokio::test]
    async fn test_read_parquet_without_field_ids_schema_evolution() {
        use arrow_array::{Array, Int32Array};

        // Schema with field 3 added after the file was written
        let schema = Arc::new(
            Schema::builder()
                .with_schema_id(1)
                .with_fields(vec![
                    NestedField::required(1, "name", Type::Primitive(PrimitiveType::String)).into(),
                    NestedField::required(2, "age", Type::Primitive(PrimitiveType::Int)).into(),
                    NestedField::optional(3, "city", Type::Primitive(PrimitiveType::String)).into(),
                ])
                .build()
                .unwrap(),
        );

        let arrow_schema = Arc::new(ArrowSchema::new(vec![
            Field::new("name", DataType::Utf8, false),
            Field::new("age", DataType::Int32, false),
        ]));

        let tmp_dir = TempDir::new().unwrap();
        let table_location = tmp_dir.path().to_str().unwrap().to_string();
        let file_io = FileIO::from_path(&table_location).unwrap().build().unwrap();

        let name_data = Arc::new(StringArray::from(vec!["Alice", "Bob"])) as ArrayRef;
        let age_data = Arc::new(Int32Array::from(vec![30, 25])) as ArrayRef;

        let to_write =
            RecordBatch::try_new(arrow_schema.clone(), vec![name_data, age_data]).unwrap();

        let props = WriterProperties::builder()
            .set_compression(Compression::SNAPPY)
            .build();

        let file = File::create(format!("{table_location}/1.parquet")).unwrap();
        let mut writer = ArrowWriter::try_new(file, to_write.schema(), Some(props)).unwrap();

        writer.write(&to_write).expect("Writing batch");
        writer.close().unwrap();

        let reader = ArrowReaderBuilder::new(file_io).build();

        let tasks = Box::pin(futures::stream::iter(
            vec![Ok(FileScanTask {
                start: 0,
                length: 0,
                record_count: None,
                data_file_path: format!("{table_location}/1.parquet"),
                data_file_format: DataFileFormat::Parquet,
                schema: schema.clone(),
                project_field_ids: vec![1, 2, 3],
                predicate: None,
                deletes: vec![],
                partition: None,
                partition_spec: None,
                name_mapping: None,
                case_sensitive: false,
            })]
            .into_iter(),
        )) as FileScanTaskStream;

        let result = reader
            .read(tasks)
            .unwrap()
            .try_collect::<Vec<RecordBatch>>()
            .await
            .unwrap();

        assert_eq!(result.len(), 1);
        let batch = &result[0];
        assert_eq!(batch.num_rows(), 2);
        assert_eq!(batch.num_columns(), 3);

        let name_array = batch.column(0).as_string::<i32>();
        assert_eq!(name_array.value(0), "Alice");
        assert_eq!(name_array.value(1), "Bob");

        let age_array = batch
            .column(1)
            .as_primitive::<arrow_array::types::Int32Type>();
        assert_eq!(age_array.value(0), 30);
        assert_eq!(age_array.value(1), 25);

        // Verify missing column filled with NULLs
        let city_array = batch.column(2).as_string::<i32>();
        assert_eq!(city_array.null_count(), 2);
        assert!(city_array.is_null(0));
        assert!(city_array.is_null(1));
    }

    /// Test reading Parquet files without field IDs that have multiple row groups.
    /// This ensures the position-based fallback works correctly across row group boundaries.
    #[tokio::test]
    async fn test_read_parquet_without_field_ids_multiple_row_groups() {
        use arrow_array::Int32Array;

        let schema = Arc::new(
            Schema::builder()
                .with_schema_id(1)
                .with_fields(vec![
                    NestedField::required(1, "name", Type::Primitive(PrimitiveType::String)).into(),
                    NestedField::required(2, "value", Type::Primitive(PrimitiveType::Int)).into(),
                ])
                .build()
                .unwrap(),
        );

        let arrow_schema = Arc::new(ArrowSchema::new(vec![
            Field::new("name", DataType::Utf8, false),
            Field::new("value", DataType::Int32, false),
        ]));

        let tmp_dir = TempDir::new().unwrap();
        let table_location = tmp_dir.path().to_str().unwrap().to_string();
        let file_io = FileIO::from_path(&table_location).unwrap().build().unwrap();

        // Small row group size to create multiple row groups
        let props = WriterProperties::builder()
            .set_compression(Compression::SNAPPY)
            .set_write_batch_size(2)
            .set_max_row_group_size(2)
            .build();

        let file = File::create(format!("{table_location}/1.parquet")).unwrap();
        let mut writer = ArrowWriter::try_new(file, arrow_schema.clone(), Some(props)).unwrap();

        // Write 6 rows in 3 batches (will create 3 row groups)
        for batch_num in 0..3 {
            let name_data = Arc::new(StringArray::from(vec![
                format!("name_{}", batch_num * 2),
                format!("name_{}", batch_num * 2 + 1),
            ])) as ArrayRef;
            let value_data =
                Arc::new(Int32Array::from(vec![batch_num * 2, batch_num * 2 + 1])) as ArrayRef;

            let batch =
                RecordBatch::try_new(arrow_schema.clone(), vec![name_data, value_data]).unwrap();
            writer.write(&batch).expect("Writing batch");
        }
        writer.close().unwrap();

        let reader = ArrowReaderBuilder::new(file_io).build();

        let tasks = Box::pin(futures::stream::iter(
            vec![Ok(FileScanTask {
                start: 0,
                length: 0,
                record_count: None,
                data_file_path: format!("{table_location}/1.parquet"),
                data_file_format: DataFileFormat::Parquet,
                schema: schema.clone(),
                project_field_ids: vec![1, 2],
                predicate: None,
                deletes: vec![],
                partition: None,
                partition_spec: None,
                name_mapping: None,
                case_sensitive: false,
            })]
            .into_iter(),
        )) as FileScanTaskStream;

        let result = reader
            .read(tasks)
            .unwrap()
            .try_collect::<Vec<RecordBatch>>()
            .await
            .unwrap();

        assert!(!result.is_empty());

        let mut all_names = Vec::new();
        let mut all_values = Vec::new();

        for batch in &result {
            let name_array = batch.column(0).as_string::<i32>();
            let value_array = batch
                .column(1)
                .as_primitive::<arrow_array::types::Int32Type>();

            for i in 0..batch.num_rows() {
                all_names.push(name_array.value(i).to_string());
                all_values.push(value_array.value(i));
            }
        }

        assert_eq!(all_names.len(), 6);
        assert_eq!(all_values.len(), 6);

        for i in 0..6 {
            assert_eq!(all_names[i], format!("name_{i}"));
            assert_eq!(all_values[i], i as i32);
        }
    }

    /// Test reading Parquet files without field IDs with nested types (struct).
    /// Java's pruneColumnsFallback() projects entire top-level columns including nested content.
    /// This test verifies that a top-level struct field is projected correctly with all its nested fields.
    #[tokio::test]
    async fn test_read_parquet_without_field_ids_with_struct() {
        use arrow_array::{Int32Array, StructArray};
        use arrow_schema::Fields;

        let schema = Arc::new(
            Schema::builder()
                .with_schema_id(1)
                .with_fields(vec![
                    NestedField::required(1, "id", Type::Primitive(PrimitiveType::Int)).into(),
                    NestedField::required(
                        2,
                        "person",
                        Type::Struct(crate::spec::StructType::new(vec![
                            NestedField::required(
                                3,
                                "name",
                                Type::Primitive(PrimitiveType::String),
                            )
                            .into(),
                            NestedField::required(4, "age", Type::Primitive(PrimitiveType::Int))
                                .into(),
                        ])),
                    )
                    .into(),
                ])
                .build()
                .unwrap(),
        );

        let arrow_schema = Arc::new(ArrowSchema::new(vec![
            Field::new("id", DataType::Int32, false),
            Field::new(
                "person",
                DataType::Struct(Fields::from(vec![
                    Field::new("name", DataType::Utf8, false),
                    Field::new("age", DataType::Int32, false),
                ])),
                false,
            ),
        ]));

        let tmp_dir = TempDir::new().unwrap();
        let table_location = tmp_dir.path().to_str().unwrap().to_string();
        let file_io = FileIO::from_path(&table_location).unwrap().build().unwrap();

        let id_data = Arc::new(Int32Array::from(vec![1, 2])) as ArrayRef;
        let name_data = Arc::new(StringArray::from(vec!["Alice", "Bob"])) as ArrayRef;
        let age_data = Arc::new(Int32Array::from(vec![30, 25])) as ArrayRef;
        let person_data = Arc::new(StructArray::from(vec![
            (
                Arc::new(Field::new("name", DataType::Utf8, false)),
                name_data,
            ),
            (
                Arc::new(Field::new("age", DataType::Int32, false)),
                age_data,
            ),
        ])) as ArrayRef;

        let to_write =
            RecordBatch::try_new(arrow_schema.clone(), vec![id_data, person_data]).unwrap();

        let props = WriterProperties::builder()
            .set_compression(Compression::SNAPPY)
            .build();

        let file = File::create(format!("{table_location}/1.parquet")).unwrap();
        let mut writer = ArrowWriter::try_new(file, to_write.schema(), Some(props)).unwrap();

        writer.write(&to_write).expect("Writing batch");
        writer.close().unwrap();

        let reader = ArrowReaderBuilder::new(file_io).build();

        let tasks = Box::pin(futures::stream::iter(
            vec![Ok(FileScanTask {
                start: 0,
                length: 0,
                record_count: None,
                data_file_path: format!("{table_location}/1.parquet"),
                data_file_format: DataFileFormat::Parquet,
                schema: schema.clone(),
                project_field_ids: vec![1, 2],
                predicate: None,
                deletes: vec![],
                partition: None,
                partition_spec: None,
                name_mapping: None,
                case_sensitive: false,
            })]
            .into_iter(),
        )) as FileScanTaskStream;

        let result = reader
            .read(tasks)
            .unwrap()
            .try_collect::<Vec<RecordBatch>>()
            .await
            .unwrap();

        assert_eq!(result.len(), 1);
        let batch = &result[0];
        assert_eq!(batch.num_rows(), 2);
        assert_eq!(batch.num_columns(), 2);

        let id_array = batch
            .column(0)
            .as_primitive::<arrow_array::types::Int32Type>();
        assert_eq!(id_array.value(0), 1);
        assert_eq!(id_array.value(1), 2);

        let person_array = batch.column(1).as_struct();
        assert_eq!(person_array.num_columns(), 2);

        let name_array = person_array.column(0).as_string::<i32>();
        assert_eq!(name_array.value(0), "Alice");
        assert_eq!(name_array.value(1), "Bob");

        let age_array = person_array
            .column(1)
            .as_primitive::<arrow_array::types::Int32Type>();
        assert_eq!(age_array.value(0), 30);
        assert_eq!(age_array.value(1), 25);
    }

    /// Test reading Parquet files without field IDs with schema evolution - column added in the middle.
    /// When a new column is inserted between existing columns in the schema order,
    /// the fallback projection must correctly map field IDs to output positions.
    #[tokio::test]
    async fn test_read_parquet_without_field_ids_schema_evolution_add_column_in_middle() {
        use arrow_array::{Array, Int32Array};

        let arrow_schema_old = Arc::new(ArrowSchema::new(vec![
            Field::new("col0", DataType::Int32, true),
            Field::new("col1", DataType::Int32, true),
        ]));

        // New column added between existing columns: col0 (id=1), newCol (id=5), col1 (id=2)
        let schema = Arc::new(
            Schema::builder()
                .with_schema_id(1)
                .with_fields(vec![
                    NestedField::optional(1, "col0", Type::Primitive(PrimitiveType::Int)).into(),
                    NestedField::optional(5, "newCol", Type::Primitive(PrimitiveType::Int)).into(),
                    NestedField::optional(2, "col1", Type::Primitive(PrimitiveType::Int)).into(),
                ])
                .build()
                .unwrap(),
        );

        let tmp_dir = TempDir::new().unwrap();
        let table_location = tmp_dir.path().to_str().unwrap().to_string();
        let file_io = FileIO::from_path(&table_location).unwrap().build().unwrap();

        let col0_data = Arc::new(Int32Array::from(vec![1, 2])) as ArrayRef;
        let col1_data = Arc::new(Int32Array::from(vec![10, 20])) as ArrayRef;

        let to_write =
            RecordBatch::try_new(arrow_schema_old.clone(), vec![col0_data, col1_data]).unwrap();

        let props = WriterProperties::builder()
            .set_compression(Compression::SNAPPY)
            .build();

        let file = File::create(format!("{table_location}/1.parquet")).unwrap();
        let mut writer = ArrowWriter::try_new(file, to_write.schema(), Some(props)).unwrap();
        writer.write(&to_write).expect("Writing batch");
        writer.close().unwrap();

        let reader = ArrowReaderBuilder::new(file_io).build();

        let tasks = Box::pin(futures::stream::iter(
            vec![Ok(FileScanTask {
                start: 0,
                length: 0,
                record_count: None,
                data_file_path: format!("{table_location}/1.parquet"),
                data_file_format: DataFileFormat::Parquet,
                schema: schema.clone(),
                project_field_ids: vec![1, 5, 2],
                predicate: None,
                deletes: vec![],
                partition: None,
                partition_spec: None,
                name_mapping: None,
                case_sensitive: false,
            })]
            .into_iter(),
        )) as FileScanTaskStream;

        let result = reader
            .read(tasks)
            .unwrap()
            .try_collect::<Vec<RecordBatch>>()
            .await
            .unwrap();

        assert_eq!(result.len(), 1);
        let batch = &result[0];
        assert_eq!(batch.num_rows(), 2);
        assert_eq!(batch.num_columns(), 3);

        let result_col0 = batch
            .column(0)
            .as_primitive::<arrow_array::types::Int32Type>();
        assert_eq!(result_col0.value(0), 1);
        assert_eq!(result_col0.value(1), 2);

        // New column should be NULL (doesn't exist in old file)
        let result_newcol = batch
            .column(1)
            .as_primitive::<arrow_array::types::Int32Type>();
        assert_eq!(result_newcol.null_count(), 2);
        assert!(result_newcol.is_null(0));
        assert!(result_newcol.is_null(1));

        let result_col1 = batch
            .column(2)
            .as_primitive::<arrow_array::types::Int32Type>();
        assert_eq!(result_col1.value(0), 10);
        assert_eq!(result_col1.value(1), 20);
    }

    /// Test reading Parquet files without field IDs with a filter that eliminates all row groups.
    /// During development of field ID mapping, we saw a panic when row_selection_enabled=true and
    /// all row groups are filtered out.
    #[tokio::test]
    async fn test_read_parquet_without_field_ids_filter_eliminates_all_rows() {
        use arrow_array::{Float64Array, Int32Array};

        // Schema with fields that will use fallback IDs 1, 2, 3
        let schema = Arc::new(
            Schema::builder()
                .with_schema_id(1)
                .with_fields(vec![
                    NestedField::required(1, "id", Type::Primitive(PrimitiveType::Int)).into(),
                    NestedField::required(2, "name", Type::Primitive(PrimitiveType::String)).into(),
                    NestedField::required(3, "value", Type::Primitive(PrimitiveType::Double))
                        .into(),
                ])
                .build()
                .unwrap(),
        );

        let arrow_schema = Arc::new(ArrowSchema::new(vec![
            Field::new("id", DataType::Int32, false),
            Field::new("name", DataType::Utf8, false),
            Field::new("value", DataType::Float64, false),
        ]));

        let tmp_dir = TempDir::new().unwrap();
        let table_location = tmp_dir.path().to_str().unwrap().to_string();
        let file_io = FileIO::from_path(&table_location).unwrap().build().unwrap();

        // Write data where all ids are >= 10
        let id_data = Arc::new(Int32Array::from(vec![10, 11, 12])) as ArrayRef;
        let name_data = Arc::new(StringArray::from(vec!["a", "b", "c"])) as ArrayRef;
        let value_data = Arc::new(Float64Array::from(vec![100.0, 200.0, 300.0])) as ArrayRef;

        let to_write =
            RecordBatch::try_new(arrow_schema.clone(), vec![id_data, name_data, value_data])
                .unwrap();

        let props = WriterProperties::builder()
            .set_compression(Compression::SNAPPY)
            .build();

        let file = File::create(format!("{table_location}/1.parquet")).unwrap();
        let mut writer = ArrowWriter::try_new(file, to_write.schema(), Some(props)).unwrap();
        writer.write(&to_write).expect("Writing batch");
        writer.close().unwrap();

        // Filter that eliminates all row groups: id < 5
        let predicate = Reference::new("id").less_than(Datum::int(5));

        // Enable both row_group_filtering and row_selection - triggered the panic
        let reader = ArrowReaderBuilder::new(file_io)
            .with_row_group_filtering_enabled(true)
            .with_row_selection_enabled(true)
            .build();

        let tasks = Box::pin(futures::stream::iter(
            vec![Ok(FileScanTask {
                start: 0,
                length: 0,
                record_count: None,
                data_file_path: format!("{table_location}/1.parquet"),
                data_file_format: DataFileFormat::Parquet,
                schema: schema.clone(),
                project_field_ids: vec![1, 2, 3],
                predicate: Some(predicate.bind(schema, true).unwrap()),
                deletes: vec![],
                partition: None,
                partition_spec: None,
                name_mapping: None,
                case_sensitive: false,
            })]
            .into_iter(),
        )) as FileScanTaskStream;

        // Should no longer panic
        let result = reader
            .read(tasks)
            .unwrap()
            .try_collect::<Vec<RecordBatch>>()
            .await
            .unwrap();

        // Should return empty results
        assert!(result.is_empty() || result.iter().all(|batch| batch.num_rows() == 0));
    }

    /// Test that concurrency=1 reads all files correctly and in deterministic order.
    /// This verifies the fast-path optimization for single concurrency.
    #[tokio::test]
    async fn test_read_with_concurrency_one() {
        use arrow_array::Int32Array;

        let schema = Arc::new(
            Schema::builder()
                .with_schema_id(1)
                .with_fields(vec![
                    NestedField::required(1, "id", Type::Primitive(PrimitiveType::Int)).into(),
                    NestedField::required(2, "file_num", Type::Primitive(PrimitiveType::Int))
                        .into(),
                ])
                .build()
                .unwrap(),
        );

        let arrow_schema = Arc::new(ArrowSchema::new(vec![
            Field::new("id", DataType::Int32, false).with_metadata(HashMap::from([(
                PARQUET_FIELD_ID_META_KEY.to_string(),
                "1".to_string(),
            )])),
            Field::new("file_num", DataType::Int32, false).with_metadata(HashMap::from([(
                PARQUET_FIELD_ID_META_KEY.to_string(),
                "2".to_string(),
            )])),
        ]));

        let tmp_dir = TempDir::new().unwrap();
        let table_location = tmp_dir.path().to_str().unwrap().to_string();
        let file_io = FileIO::from_path(&table_location).unwrap().build().unwrap();

        // Create 3 parquet files with different data
        let props = WriterProperties::builder()
            .set_compression(Compression::SNAPPY)
            .build();

        for file_num in 0..3 {
            let id_data = Arc::new(Int32Array::from_iter_values(
                file_num * 10..(file_num + 1) * 10,
            )) as ArrayRef;
            let file_num_data = Arc::new(Int32Array::from(vec![file_num; 10])) as ArrayRef;

            let to_write =
                RecordBatch::try_new(arrow_schema.clone(), vec![id_data, file_num_data]).unwrap();

            let file = File::create(format!("{table_location}/file_{file_num}.parquet")).unwrap();
            let mut writer =
                ArrowWriter::try_new(file, to_write.schema(), Some(props.clone())).unwrap();
            writer.write(&to_write).expect("Writing batch");
            writer.close().unwrap();
        }

        // Read with concurrency=1 (fast-path)
        let reader = ArrowReaderBuilder::new(file_io)
            .with_data_file_concurrency_limit(1)
            .build();

        // Create tasks in a specific order: file_0, file_1, file_2
        let tasks = vec![
            Ok(FileScanTask {
                start: 0,
                length: 0,
                record_count: None,
                data_file_path: format!("{table_location}/file_0.parquet"),
                data_file_format: DataFileFormat::Parquet,
                schema: schema.clone(),
                project_field_ids: vec![1, 2],
                predicate: None,
                deletes: vec![],
                partition: None,
                partition_spec: None,
                name_mapping: None,
                case_sensitive: false,
            }),
            Ok(FileScanTask {
                start: 0,
                length: 0,
                record_count: None,
                data_file_path: format!("{table_location}/file_1.parquet"),
                data_file_format: DataFileFormat::Parquet,
                schema: schema.clone(),
                project_field_ids: vec![1, 2],
                predicate: None,
                deletes: vec![],
                partition: None,
                partition_spec: None,
                name_mapping: None,
                case_sensitive: false,
            }),
            Ok(FileScanTask {
                start: 0,
                length: 0,
                record_count: None,
                data_file_path: format!("{table_location}/file_2.parquet"),
                data_file_format: DataFileFormat::Parquet,
                schema: schema.clone(),
                project_field_ids: vec![1, 2],
                predicate: None,
                deletes: vec![],
                partition: None,
                partition_spec: None,
                name_mapping: None,
                case_sensitive: false,
            }),
        ];

        let tasks_stream = Box::pin(futures::stream::iter(tasks)) as FileScanTaskStream;

        let result = reader
            .read(tasks_stream)
            .unwrap()
            .try_collect::<Vec<RecordBatch>>()
            .await
            .unwrap();

        // Verify we got all 30 rows (10 from each file)
        let total_rows: usize = result.iter().map(|b| b.num_rows()).sum();
        assert_eq!(total_rows, 30, "Should have 30 total rows");

        // Collect all ids and file_nums to verify data
        let mut all_ids = Vec::new();
        let mut all_file_nums = Vec::new();

        for batch in &result {
            let id_col = batch
                .column(0)
                .as_primitive::<arrow_array::types::Int32Type>();
            let file_num_col = batch
                .column(1)
                .as_primitive::<arrow_array::types::Int32Type>();

            for i in 0..batch.num_rows() {
                all_ids.push(id_col.value(i));
                all_file_nums.push(file_num_col.value(i));
            }
        }

        assert_eq!(all_ids.len(), 30);
        assert_eq!(all_file_nums.len(), 30);

        // With concurrency=1 and sequential processing, files should be processed in order
        // file_0: ids 0-9, file_num=0
        // file_1: ids 10-19, file_num=1
        // file_2: ids 20-29, file_num=2
        for i in 0..10 {
            assert_eq!(all_file_nums[i], 0, "First 10 rows should be from file_0");
            assert_eq!(all_ids[i], i as i32, "IDs should be 0-9");
        }
        for i in 10..20 {
            assert_eq!(all_file_nums[i], 1, "Next 10 rows should be from file_1");
            assert_eq!(all_ids[i], i as i32, "IDs should be 10-19");
        }
        for i in 20..30 {
            assert_eq!(all_file_nums[i], 2, "Last 10 rows should be from file_2");
            assert_eq!(all_ids[i], i as i32, "IDs should be 20-29");
        }
    }

    /// Test bucket partitioning reads source column from data file (not partition metadata).
    ///
    /// This is an integration test verifying the complete ArrowReader pipeline with bucket partitioning.
    /// It corresponds to TestRuntimeFiltering tests in Iceberg Java (e.g., testRenamedSourceColumnTable).
    ///
    /// # Iceberg Spec Requirements
    ///
    /// Per the Iceberg spec "Column Projection" section:
    /// > "Return the value from partition metadata if an **Identity Transform** exists for the field"
    ///
    /// This means:
    /// - Identity transforms (e.g., `identity(dept)`) use constants from partition metadata
    /// - Non-identity transforms (e.g., `bucket(4, id)`) must read source columns from data files
    /// - Partition metadata for bucket transforms stores bucket numbers (0-3), NOT source values
    ///
    /// Java's PartitionUtil.constantsMap() implements this via:
    /// ```java
    /// if (field.transform().isIdentity()) {
    ///     idToConstant.put(field.sourceId(), converted);
    /// }
    /// ```
    ///
    /// # What This Test Verifies
    ///
    /// This test ensures the full ArrowReader → RecordBatchTransformer pipeline correctly handles
    /// bucket partitioning when FileScanTask provides partition_spec and partition_data:
    ///
    /// - Parquet file has field_id=1 named "id" with actual data [1, 5, 9, 13]
    /// - FileScanTask specifies partition_spec with bucket(4, id) and partition_data with bucket=1
    /// - RecordBatchTransformer.constants_map() excludes bucket-partitioned field from constants
    /// - ArrowReader correctly reads [1, 5, 9, 13] from the data file
    /// - Values are NOT replaced with constant 1 from partition metadata
    ///
    /// # Why This Matters
    ///
    /// Without correct handling:
    /// - Runtime filtering would break (e.g., `WHERE id = 5` would fail)
    /// - Query results would be incorrect (all rows would have id=1)
    /// - Bucket partitioning would be unusable for query optimization
    ///
    /// # References
    /// - Iceberg spec: format/spec.md "Column Projection" + "Partition Transforms"
    /// - Java test: spark/src/test/java/.../TestRuntimeFiltering.java
    /// - Java impl: core/src/main/java/org/apache/iceberg/util/PartitionUtil.java
    #[tokio::test]
    async fn test_bucket_partitioning_reads_source_column_from_file() {
        use arrow_array::Int32Array;

        use crate::spec::{Literal, PartitionSpec, Struct, Transform};

        // Iceberg schema with id and name columns
        let schema = Arc::new(
            Schema::builder()
                .with_schema_id(0)
                .with_fields(vec![
                    NestedField::required(1, "id", Type::Primitive(PrimitiveType::Int)).into(),
                    NestedField::optional(2, "name", Type::Primitive(PrimitiveType::String)).into(),
                ])
                .build()
                .unwrap(),
        );

        // Partition spec: bucket(4, id)
        let partition_spec = Arc::new(
            PartitionSpec::builder(schema.clone())
                .with_spec_id(0)
                .add_partition_field("id", "id_bucket", Transform::Bucket(4))
                .unwrap()
                .build()
                .unwrap(),
        );

        // Partition data: bucket value is 1
        let partition_data = Struct::from_iter(vec![Some(Literal::int(1))]);

        // Create Arrow schema with field IDs for Parquet file
        let arrow_schema = Arc::new(ArrowSchema::new(vec![
            Field::new("id", DataType::Int32, false).with_metadata(HashMap::from([(
                PARQUET_FIELD_ID_META_KEY.to_string(),
                "1".to_string(),
            )])),
            Field::new("name", DataType::Utf8, true).with_metadata(HashMap::from([(
                PARQUET_FIELD_ID_META_KEY.to_string(),
                "2".to_string(),
            )])),
        ]));

        // Write Parquet file with data
        let tmp_dir = TempDir::new().unwrap();
        let table_location = tmp_dir.path().to_str().unwrap().to_string();
        let file_io = FileIO::from_path(&table_location).unwrap().build().unwrap();

        let id_data = Arc::new(Int32Array::from(vec![1, 5, 9, 13])) as ArrayRef;
        let name_data =
            Arc::new(StringArray::from(vec!["Alice", "Bob", "Charlie", "Dave"])) as ArrayRef;

        let to_write =
            RecordBatch::try_new(arrow_schema.clone(), vec![id_data, name_data]).unwrap();

        let props = WriterProperties::builder()
            .set_compression(Compression::SNAPPY)
            .build();
        let file = File::create(format!("{}/data.parquet", &table_location)).unwrap();
        let mut writer = ArrowWriter::try_new(file, to_write.schema(), Some(props)).unwrap();
        writer.write(&to_write).expect("Writing batch");
        writer.close().unwrap();

        // Read the Parquet file with partition spec and data
        let reader = ArrowReaderBuilder::new(file_io).build();
        let tasks = Box::pin(futures::stream::iter(
            vec![Ok(FileScanTask {
                start: 0,
                length: 0,
                record_count: None,
                data_file_path: format!("{table_location}/data.parquet"),
                data_file_format: DataFileFormat::Parquet,
                schema: schema.clone(),
                project_field_ids: vec![1, 2],
                predicate: None,
                deletes: vec![],
                partition: Some(partition_data),
                partition_spec: Some(partition_spec),
                name_mapping: None,
                case_sensitive: false,
            })]
            .into_iter(),
        )) as FileScanTaskStream;

        let result = reader
            .read(tasks)
            .unwrap()
            .try_collect::<Vec<RecordBatch>>()
            .await
            .unwrap();

        // Verify we got the correct data
        assert_eq!(result.len(), 1);
        let batch = &result[0];

        assert_eq!(batch.num_columns(), 2);
        assert_eq!(batch.num_rows(), 4);

        // The id column MUST contain actual values from the Parquet file [1, 5, 9, 13],
        // NOT the constant partition value 1
        let id_col = batch
            .column(0)
            .as_primitive::<arrow_array::types::Int32Type>();
        assert_eq!(id_col.value(0), 1);
        assert_eq!(id_col.value(1), 5);
        assert_eq!(id_col.value(2), 9);
        assert_eq!(id_col.value(3), 13);

        let name_col = batch.column(1).as_string::<i32>();
        assert_eq!(name_col.value(0), "Alice");
        assert_eq!(name_col.value(1), "Bob");
        assert_eq!(name_col.value(2), "Charlie");
        assert_eq!(name_col.value(3), "Dave");
    }
}