mito2/sst/parquet/

reader.rs

// Copyright 2023 Greptime Team
//
// Licensed 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 reader.

use std::collections::{HashMap, HashSet};
use std::sync::Arc;
use std::time::{Duration, Instant};

use api::v1::SemanticType;
use arrow_schema::extension::{
    EXTENSION_TYPE_METADATA_KEY, EXTENSION_TYPE_NAME_KEY, ExtensionType,
};
use common_recordbatch::filter::{SimpleFilterEvaluator, TimestampUnitCast};
use common_telemetry::{debug, error, tracing, warn};
use datafusion::physical_plan::PhysicalExpr;
use datafusion_common::tree_node::{TreeNode, TreeNodeRecursion};
use datafusion_expr::utils::expr_to_columns;
use datafusion_expr::{Expr, Volatility};
use datatypes::arrow::array::ArrayRef;
use datatypes::arrow::datatypes::{Field, Schema as ArrowSchema, SchemaRef};
use datatypes::arrow::record_batch::RecordBatch;
use datatypes::data_type::ConcreteDataType;
use datatypes::extension::json::{Json2ExtensionType, is_json2_extension_type};
use datatypes::prelude::DataType;
use datatypes::vectors::json::json2_physical_data_type;
use futures::StreamExt;
use mito_codec::row_converter::build_primary_key_codec;
use object_store::ObjectStore;
use parquet::arrow::arrow_reader::{ArrowReaderMetadata, ArrowReaderOptions, RowSelection};
use parquet::arrow::{ProjectionMask, parquet_to_arrow_schema};
use parquet::file::metadata::{PageIndexPolicy, ParquetMetaData};
use parquet::file::properties::DEFAULT_DICTIONARY_PAGE_SIZE_LIMIT;
use partition::expr::PartitionExpr;
use snafu::{OptionExt, ResultExt};
use store_api::codec::PrimaryKeyEncoding;
use store_api::metadata::{ColumnMetadata, RegionMetadata, RegionMetadataRef};
use store_api::region_request::PathType;
use store_api::storage::consts::PRIMARY_KEY_COLUMN_NAME;
use store_api::storage::{ColumnId, FileId};
use table::predicate::Predicate;

use crate::cache::index::result_cache::PredicateKey;
use crate::cache::{CacheStrategy, CachedSstMeta, SstMetaPreparation, prepare_sst_meta};
use crate::error::{
    ParquetToArrowSchemaSnafu, ReadDataPartSnafu, Result, SerializePartitionExprSnafu,
    UnexpectedSnafu,
};
use crate::metrics::{
    PRECISE_FILTER_ROWS_TOTAL, READ_ROW_GROUPS_TOTAL, READ_ROWS_IN_ROW_GROUP_TOTAL,
    READ_ROWS_TOTAL, READ_STAGE_ELAPSED,
};
use crate::read::flat_projection::CompactionProjectionMapper;
use crate::read::prune::FlatPruneReader;
use crate::read::read_columns::ReadColumns;
use crate::series_index::SeriesIndexReadContext;
use crate::sst::file::FileHandle;
use crate::sst::index::bloom_filter::applier::{
    BloomFilterIndexApplierRef, BloomFilterIndexApplyMetrics,
};
use crate::sst::index::fulltext_index::applier::{
    FulltextIndexApplierRef, FulltextIndexApplyMetrics,
};
use crate::sst::index::inverted_index::applier::{
    InvertedIndexApplierRef, InvertedIndexApplyMetrics,
};
use crate::sst::parquet::file_range::{
    FileRangeContext, FileRangeContextRef, PartitionFilterContext, PreFilterMode, RangeBase,
};
use crate::sst::parquet::flat_format::{FlatReadFormat, primary_key_column_index};
use crate::sst::parquet::format::{INTERNAL_COLUMN_NUM, need_override_sequence};
use crate::sst::parquet::json_align::{AlignMode, JsonSchemaAligner, ProjectedRecordBatchStream};
use crate::sst::parquet::metadata::MetadataLoader;
use crate::sst::parquet::prefilter::{
    PrefilterContextBuilder, build_reader_filter_plan, execute_prefilter,
};
use crate::sst::parquet::push_decoder::{
    SstParquetRangeFetcher, build_sst_parquet_record_batch_stream,
};
use crate::sst::parquet::read_columns::{ProjectionMaskPlan, build_projection_plan};
use crate::sst::parquet::row_group::ParquetFetchMetrics;
use crate::sst::parquet::row_selection::RowGroupSelection;
use crate::sst::parquet::stats::RowGroupPruningStats;
use crate::sst::parquet::{DEFAULT_READ_BATCH_SIZE, Json2RewriteTargets, Json2TargetLayout};
use crate::sst::{override_pk_field_to_binary, tag_maybe_to_dictionary_field};

const INDEX_TYPE_FULLTEXT: &str = "fulltext";

/// Number of leading row groups sampled by [`should_read_pk_as_binary`].
const MAX_ROW_GROUPS_TO_CHECK_PK: usize = 4;

/// Returns `true` if the `__primary_key` chunk in any of the first
/// [`MAX_ROW_GROUPS_TO_CHECK_PK`] row groups exceeds the dictionary page size
/// limit, signalling the writer likely fell back to plain encoding.
fn should_read_pk_as_binary(parquet_meta: &ParquetMetaData) -> bool {
    should_read_pk_as_binary_with_limit(parquet_meta, DEFAULT_DICTIONARY_PAGE_SIZE_LIMIT)
}

fn apply_json2_rewrite_targets(
    read_format: &FlatReadFormat,
    targets: &Json2RewriteTargets,
) -> Result<SchemaRef> {
    let schema = read_format.output_arrow_schema()?;
    if targets.is_empty() {
        return Ok(schema);
    }

    let mut schema = schema.as_ref().clone();
    let mut fields = schema.fields().iter().cloned().collect::<Vec<_>>();
    for (column_id, layout) in targets.iter() {
        let Some(index) = read_format.parquet_projected_index_by_id(*column_id) else {
            continue;
        };
        let Some(field) = fields.get(index) else {
            continue;
        };
        let mut field = field.as_ref().clone();
        field.set_data_type(json2_physical_data_type(&layout.target_layout));

        let mut metadata = field.metadata().clone();
        metadata.insert(
            EXTENSION_TYPE_NAME_KEY.to_string(),
            Json2ExtensionType::NAME.to_string(),
        );
        metadata.insert(
            EXTENSION_TYPE_METADATA_KEY.to_string(),
            layout.extension_metadata.clone(),
        );
        field.set_metadata(metadata);
        fields[index] = Arc::new(field);
    }
    schema.fields = fields.into();
    Ok(Arc::new(schema))
}

fn should_read_pk_as_binary_with_limit(
    parquet_meta: &ParquetMetaData,
    dict_page_size_limit: usize,
) -> bool {
    let num_columns = parquet_meta.file_metadata().schema_descr().num_columns();
    if num_columns < INTERNAL_COLUMN_NUM {
        return false;
    }
    let pk_idx = primary_key_column_index(num_columns);
    parquet_meta
        .row_groups()
        .iter()
        .take(MAX_ROW_GROUPS_TO_CHECK_PK)
        .any(|rg| rg.column(pk_idx).uncompressed_size() as usize > dict_page_size_limit)
}
const INDEX_TYPE_INVERTED: &str = "inverted";
const INDEX_TYPE_BLOOM: &str = "bloom filter";

macro_rules! handle_index_error {
    ($err:expr, $file_handle:expr, $index_type:expr) => {
        if cfg!(any(test, feature = "test")) {
            panic!(
                "Failed to apply {} index, region_id: {}, file_id: {}, err: {:?}",
                $index_type,
                $file_handle.region_id(),
                $file_handle.file_id(),
                $err
            );
        } else {
            warn!(
                $err; "Failed to apply {} index, region_id: {}, file_id: {}",
                $index_type,
                $file_handle.region_id(),
                $file_handle.file_id()
            );
        }
    };
}

/// Parquet SST reader builder.
pub struct ParquetReaderBuilder {
    /// Index snapshot and store used to determine range-index availability.
    series_index: Option<SeriesIndexReadContext>,
    /// SST directory.
    table_dir: String,
    /// Path type for generating file paths.
    path_type: PathType,
    file_handle: FileHandle,
    object_store: ObjectStore,
    /// Predicate to push down.
    predicate: Option<Predicate>,
    /// The columns to read.
    ///
    /// `None` reads all columns. Due to schema change, the projection
    /// can contain columns not in the parquet file.
    read_cols: Option<ReadColumns>,
    /// Compaction-only JSON2 physical rewrite targets.
    json2_rewrite_targets: Json2RewriteTargets,
    /// Strategy to cache SST data.
    cache_strategy: CacheStrategy,
    /// Index appliers.
    inverted_index_appliers: [Option<InvertedIndexApplierRef>; 2],
    bloom_filter_index_appliers: [Option<BloomFilterIndexApplierRef>; 2],
    fulltext_index_appliers: [Option<FulltextIndexApplierRef>; 2],
    /// Expected metadata of the region while reading the SST.
    /// This is usually the latest metadata of the region. The reader use
    /// it get the correct column id of a column by name.
    expected_metadata: Option<RegionMetadataRef>,
    /// Whether this reader is for compaction.
    compaction: bool,
    /// Mode to pre-filter columns.
    pre_filter_mode: PreFilterMode,
    /// Whether to run the reduced-column predicate prefilter pass.
    enable_predicate_prefilter: bool,
    /// Whether to apply simple time index filters during the normal precise-filter pass.
    postpone_time_index_filter: bool,
    /// Whether to decode primary key values eagerly when reading primary key format SSTs.
    decode_primary_key_values: bool,
    page_index_policy: PageIndexPolicy,
    defer_optional_page_index: bool,
    /// Scan-wide hint for rows in a decoded batch.
    batch_size: usize,
}

impl ParquetReaderBuilder {
    /// Returns a new [ParquetReaderBuilder] to read specific SST.
    pub fn new(
        table_dir: String,
        path_type: PathType,
        file_handle: FileHandle,
        object_store: ObjectStore,
    ) -> ParquetReaderBuilder {
        ParquetReaderBuilder {
            series_index: None,
            table_dir,
            path_type,
            file_handle,
            object_store,
            predicate: None,
            read_cols: None,
            json2_rewrite_targets: Arc::default(),
            cache_strategy: CacheStrategy::Disabled,
            inverted_index_appliers: [None, None],
            bloom_filter_index_appliers: [None, None],
            fulltext_index_appliers: [None, None],
            expected_metadata: None,
            compaction: false,
            pre_filter_mode: PreFilterMode::All,
            enable_predicate_prefilter: true,
            postpone_time_index_filter: false,
            decode_primary_key_values: false,
            page_index_policy: Default::default(),
            defer_optional_page_index: false,
            batch_size: DEFAULT_READ_BATCH_SIZE,
        }
    }

    /// Sets the captured series-index context for range-index reads.
    #[must_use]
    pub(crate) fn series_index(mut self, context: Option<SeriesIndexReadContext>) -> Self {
        self.series_index = context;
        self
    }

    /// Sets the scan-wide hint for rows in a decoded batch.
    #[must_use]
    pub(crate) fn batch_size(mut self, batch_size: usize) -> Self {
        self.batch_size = batch_size.clamp(1, DEFAULT_READ_BATCH_SIZE);
        self
    }

    /// Attaches the predicate to the builder.
    #[must_use]
    pub fn predicate(mut self, predicate: Option<Predicate>) -> ParquetReaderBuilder {
        self.predicate = predicate;
        self
    }

    /// Attaches the projection to the builder.
    ///
    /// The reader only applies the projection to fields.
    #[must_use]
    pub fn projection(mut self, read_cols: Option<ReadColumns>) -> ParquetReaderBuilder {
        self.read_cols = read_cols;
        self
    }

    /// Attaches fixed JSON2 physical layouts used by compaction readers.
    #[must_use]
    pub(crate) fn json2_rewrite_targets(mut self, targets: Json2RewriteTargets) -> Self {
        self.json2_rewrite_targets = targets;
        self
    }

    /// Attaches the cache to the builder.
    #[must_use]
    pub fn cache(mut self, cache: CacheStrategy) -> ParquetReaderBuilder {
        self.cache_strategy = cache;
        self
    }

    /// Attaches the inverted index appliers to the builder.
    #[must_use]
    pub(crate) fn inverted_index_appliers(
        mut self,
        index_appliers: [Option<InvertedIndexApplierRef>; 2],
    ) -> Self {
        self.inverted_index_appliers = index_appliers;
        self
    }

    /// Attaches the bloom filter index appliers to the builder.
    #[must_use]
    pub(crate) fn bloom_filter_index_appliers(
        mut self,
        index_appliers: [Option<BloomFilterIndexApplierRef>; 2],
    ) -> Self {
        self.bloom_filter_index_appliers = index_appliers;
        self
    }

    /// Attaches the fulltext index appliers to the builder.
    #[must_use]
    pub(crate) fn fulltext_index_appliers(
        mut self,
        index_appliers: [Option<FulltextIndexApplierRef>; 2],
    ) -> Self {
        self.fulltext_index_appliers = index_appliers;
        self
    }

    /// Attaches the expected metadata to the builder.
    #[must_use]
    pub fn expected_metadata(mut self, expected_metadata: Option<RegionMetadataRef>) -> Self {
        self.expected_metadata = expected_metadata;
        self
    }

    /// Sets the compaction flag.
    #[must_use]
    pub fn compaction(mut self, compaction: bool) -> Self {
        self.compaction = compaction;
        self
    }

    /// Sets the pre-filter mode.
    #[must_use]
    pub(crate) fn pre_filter_mode(mut self, pre_filter_mode: PreFilterMode) -> Self {
        self.pre_filter_mode = pre_filter_mode;
        self
    }

    /// Sets whether to run the reduced-column predicate prefilter pass.
    #[must_use]
    pub(crate) fn enable_predicate_prefilter(mut self, enable: bool) -> Self {
        self.enable_predicate_prefilter = enable;
        self
    }

    /// Sets whether to postpone simple time index filters to precise filtering.
    #[must_use]
    pub(crate) fn postpone_time_index_filter(mut self, postpone: bool) -> Self {
        self.postpone_time_index_filter = postpone;
        self
    }

    /// Decodes primary key values eagerly when reading primary key format SSTs.
    #[must_use]
    pub(crate) fn decode_primary_key_values(mut self, decode: bool) -> Self {
        self.decode_primary_key_values = decode;
        self
    }

    #[must_use]
    pub fn page_index_policy(mut self, page_index_policy: PageIndexPolicy) -> Self {
        self.page_index_policy = page_index_policy;
        self
    }

    /// Defers loading optional page indexes until row-level selections can use them.
    #[must_use]
    pub(crate) fn deferred_optional_page_index(mut self) -> Self {
        self.page_index_policy = PageIndexPolicy::Optional;
        self.defer_optional_page_index = true;
        self
    }

    /// Builds a [ParquetReader].
    ///
    /// This needs to perform IO operation.
    #[tracing::instrument(
        skip_all,
        fields(
            region_id = %self.file_handle.region_id(),
            file_id = %self.file_handle.file_id()
        )
    )]
    pub async fn build(&self) -> Result<Option<ParquetReader>> {
        let mut metrics = ReaderMetrics::default();

        let Some((context, selection)) = self.build_reader_input_inner(&mut metrics).await? else {
            return Ok(None);
        };
        ParquetReader::new(Arc::new(context), selection)
            .await
            .map(Some)
    }

    /// Builds a [FileRangeContext] and collects row groups to read.
    ///
    /// This needs to perform IO operation.
    #[tracing::instrument(
        skip_all,
        fields(
            region_id = %self.file_handle.region_id(),
            file_id = %self.file_handle.file_id()
        )
    )]
    pub async fn build_reader_input(
        &self,
        metrics: &mut ReaderMetrics,
    ) -> Result<Option<(FileRangeContext, RowGroupSelection)>> {
        self.build_reader_input_inner(metrics).await
    }

    async fn build_reader_input_inner(
        &self,
        metrics: &mut ReaderMetrics,
    ) -> Result<Option<(FileRangeContext, RowGroupSelection)>> {
        let start = Instant::now();

        let file_path = self.file_handle.file_path(&self.table_dir, self.path_type);
        let file_size = self.file_handle.meta_ref().file_size;

        // Loads parquet metadata of the file.
        let initial_page_index_policy = if self.defer_optional_page_index
            && self.page_index_policy == PageIndexPolicy::Optional
        {
            PageIndexPolicy::Skip
        } else {
            self.page_index_policy
        };
        let (sst_meta, mut cache_miss) = self
            .read_parquet_metadata(
                &file_path,
                file_size,
                &mut metrics.metadata_cache_metrics,
                initial_page_index_policy,
            )
            .await?;
        let mut parquet_meta = sst_meta.parquet_metadata();
        let mut parquet_metadata_size = sst_meta.parquet_metadata_size();
        let region_meta = sst_meta.region_metadata();
        let region_partition_expr_str = self
            .expected_metadata
            .as_ref()
            .and_then(|meta| meta.partition_expr.as_ref())
            .map(|expr| expr.as_str());
        let (_, is_same_region_partition) = Self::is_same_region_partition(
            region_partition_expr_str,
            self.file_handle.meta_ref().partition_expr.as_ref(),
        )?;
        // Skip auto convert when:
        // - compaction is enabled
        // - region partition expr is same with file partition expr (no need to auto convert)
        let skip_auto_convert = self.compaction && is_same_region_partition;

        // Build a compaction projection helper when:
        // - compaction is enabled
        // - region partition expr differs from file partition expr
        // - flat format is enabled
        // - primary key encoding is sparse
        //
        // This is applied after row-group filtering to align batches with flat output schema
        // before compat handling.
        let compaction_projection_mapper = if self.compaction
            && !is_same_region_partition
            && region_meta.primary_key_encoding == PrimaryKeyEncoding::Sparse
        {
            Some(CompactionProjectionMapper::try_new(&region_meta)?)
        } else {
            None
        };

        let read_cols = if let Some(read_cols) = &self.read_cols {
            read_cols.clone()
        } else {
            let expected_meta = self.expected_metadata.as_ref().unwrap_or(&region_meta);
            // Lists all column ids to read, we always use the expected metadata if possible.
            ReadColumns::new(
                expected_meta
                    .column_metadatas
                    .iter()
                    .map(|col| col.column_id),
            )
        };

        let file_metadata = parquet_meta.file_metadata();
        let parquet_schema_desc = file_metadata.schema_descr();
        let file_schema = Arc::new(
            parquet_to_arrow_schema(parquet_schema_desc, file_metadata.key_value_metadata())
                .context(ParquetToArrowSchemaSnafu { file: &file_path })?,
        );
        let mut read_format = FlatReadFormat::new(
            region_meta.clone(),
            read_cols,
            Some(file_schema.clone()),
            &file_path,
            skip_auto_convert,
        )?;
        // `region_meta` comes from the Parquet/source file and must not be used as the
        // target identity. When the caller has no current metadata, the handle is the
        // only local identity available and therefore denotes a local read.
        let expected_region_id = self
            .expected_metadata
            .as_ref()
            .map(|metadata| metadata.region_id)
            .unwrap_or(self.file_handle.region_id());
        let is_foreign = self.file_handle.region_id() != expected_region_id;
        if is_foreign {
            debug!(
                "Reading foreign SST, file_id: {}, source_region_id: {}, expected_region_id: {}",
                self.file_handle.file_id().file_id(),
                self.file_handle.region_id(),
                expected_region_id,
            );
        }
        let file_meta = self.file_handle.meta_ref();
        let override_sequence = if is_foreign {
            file_meta.sequence.map(|sequence| sequence.get())
        } else if file_meta.preserve_row_sequence {
            None
        } else if need_override_sequence(&parquet_meta) {
            file_meta.sequence.map(|sequence| sequence.get())
        } else {
            None
        };
        if let Some(sequence) = override_sequence {
            read_format.set_override_sequence(Some(sequence));
        }

        // Computes the projection mask.
        let parquet_read_cols = read_format.parquet_read_columns();
        let projection_plan =
            build_projection_plan(parquet_read_cols, parquet_schema_desc, &file_schema)?;
        let has_nested_projection = parquet_read_cols.has_nested();
        let selection = self
            .row_groups_to_read(&read_format, &parquet_meta, &mut metrics.filter_metrics)
            .await;

        if selection.is_empty() {
            metrics.build_cost += start.elapsed();
            return Ok(None);
        }

        let prune_schema = self
            .expected_metadata
            .as_ref()
            .map(|meta| meta.schema.clone())
            .unwrap_or_else(|| region_meta.schema.clone());

        let dyn_filters = if let Some(predicate) = &self.predicate {
            predicate.dyn_filters().as_ref().clone()
        } else {
            vec![]
        };

        let codec = build_primary_key_codec(read_format.metadata());

        let filter_plan = build_reader_filter_plan(
            self.predicate.as_ref(),
            self.expected_metadata.as_deref(),
            self.pre_filter_mode,
            self.enable_predicate_prefilter,
            self.postpone_time_index_filter,
            &read_format,
            &codec,
            &parquet_meta,
        );

        if self.defer_optional_page_index
            && self.page_index_policy == PageIndexPolicy::Optional
            && (filter_plan.prefilter_builder.is_some()
                || has_row_level_selection(&selection, &parquet_meta))
        {
            let (sst_meta, page_index_cache_miss) = self
                .read_parquet_metadata(
                    &file_path,
                    file_size,
                    &mut metrics.metadata_cache_metrics,
                    PageIndexPolicy::Optional,
                )
                .await?;
            parquet_meta = sst_meta.parquet_metadata();
            parquet_metadata_size = sst_meta.parquet_metadata_size();
            cache_miss |= page_index_cache_miss;
        }

        // Trigger background download if metadata had a cache miss and selection is not empty
        if cache_miss && !selection.is_empty() {
            use crate::cache::file_cache::{FileType, IndexKey};
            let index_key = IndexKey::new(
                self.file_handle.region_id(),
                self.file_handle.file_id().file_id(),
                FileType::Parquet,
            );
            self.cache_strategy.maybe_download_background(
                index_key,
                file_path.clone(),
                self.object_store.clone(),
                file_size,
            );
        }

        let output_schema = apply_json2_rewrite_targets(&read_format, &self.json2_rewrite_targets)?;

        // Create ArrowReaderMetadata for async stream building.
        let mut arrow_reader_options = ArrowReaderOptions::new();
        if !read_format
            .arrow_schema()
            .fields()
            .iter()
            .any(is_json2_extension_type)
        {
            // Read `__primary_key` as Binary when it's too large for dictionary
            // encoding; convert_batch wraps it back to a DictionaryArray.
            let schema_for_reader = if should_read_pk_as_binary(&parquet_meta) {
                read_format.set_pk_as_binary(output_schema.clone());
                override_pk_field_to_binary(read_format.arrow_schema())
            } else {
                read_format.arrow_schema().clone()
            };
            arrow_reader_options = arrow_reader_options.with_schema(schema_for_reader);
        }
        let arrow_metadata =
            ArrowReaderMetadata::try_new(parquet_meta.clone(), arrow_reader_options)
                .context(ReadDataPartSnafu)?;

        let json2_rewrite_targets = self
            .json2_rewrite_targets
            .iter()
            .map(|(column_id, layout)| {
                let column = region_meta
                    .column_by_id(*column_id)
                    .context(UnexpectedSnafu {
                        reason: format!("JSON2 target column by id {column_id} does not exist"),
                    })?;
                Ok((column.column_schema.name.clone(), layout.clone()))
            })
            .collect::<Result<HashMap<_, _>>>()?;

        let reader_builder = RowGroupReaderBuilder {
            file_handle: self.file_handle.clone(),
            file_path,
            parquet_meta,
            parquet_metadata_size,
            arrow_metadata,
            output_schema,
            json2_rewrite_targets,
            object_store: self.object_store.clone(),
            projection: projection_plan,
            has_nested_projection,
            cache_strategy: self.cache_strategy.clone(),
            prefilter_builder: filter_plan.prefilter_builder,
            batch_size: self.batch_size,
        };

        let partition_filter = self.build_partition_filter(&read_format, &prune_schema)?;

        let range_index_store = self.series_index.as_ref().and_then(|context| {
            let region_id = self
                .expected_metadata
                .as_ref()
                .unwrap_or(&region_meta)
                .region_id;
            (self.file_handle.region_id() == region_id
                && context
                    .version
                    .range_indexes
                    .contains_key(&self.file_handle.file_id().file_id()))
            .then(|| context.store.clone())
        });
        let context = FileRangeContext::new(
            reader_builder,
            RangeBase {
                filters: filter_plan.remaining_simple_filters,
                dyn_filters,
                read_format,
                expected_metadata: self.expected_metadata.clone(),
                prune_schema,
                codec,
                compat_batch: None,
                compaction_projection_mapper,
                pre_filter_mode: self.pre_filter_mode,
                partition_filter,
            },
            range_index_store,
        );

        metrics.build_cost += start.elapsed();

        Ok(Some((context, selection)))
    }

    fn is_same_region_partition(
        region_partition_expr_str: Option<&str>,
        file_partition_expr: Option<&PartitionExpr>,
    ) -> Result<(Option<PartitionExpr>, bool)> {
        let region_partition_expr = match region_partition_expr_str {
            Some(expr_str) => crate::region::parse_partition_expr(Some(expr_str))?,
            None => None,
        };

        let is_same = region_partition_expr.as_ref() == file_partition_expr;
        Ok((region_partition_expr, is_same))
    }

    /// Compare partition expressions from expected metadata and file metadata,
    /// and build a partition filter if they differ.
    fn build_partition_filter(
        &self,
        read_format: &FlatReadFormat,
        prune_schema: &Arc<datatypes::schema::Schema>,
    ) -> Result<Option<PartitionFilterContext>> {
        let region_partition_expr_str = self
            .expected_metadata
            .as_ref()
            .and_then(|meta| meta.partition_expr.as_ref());
        let file_partition_expr_ref = self.file_handle.meta_ref().partition_expr.as_ref();

        let (region_partition_expr, is_same_region_partition) = Self::is_same_region_partition(
            region_partition_expr_str.map(|s| s.as_str()),
            file_partition_expr_ref,
        )?;

        if is_same_region_partition {
            return Ok(None);
        }

        let Some(region_partition_expr) = region_partition_expr else {
            return Ok(None);
        };

        // Collect columns referenced by the partition expression.
        let mut referenced_columns = HashSet::new();
        region_partition_expr.collect_column_names(&mut referenced_columns);

        // Build a partition_schema containing only referenced columns.
        let partition_schema = Arc::new(datatypes::schema::Schema::new(
            prune_schema
                .column_schemas()
                .iter()
                .filter(|col| referenced_columns.contains(&col.name))
                .map(|col| {
                    if let Some(column_meta) = read_format.metadata().column_by_name(&col.name)
                        && column_meta.semantic_type == SemanticType::Tag
                        && col.data_type.is_string()
                    {
                        let field = Arc::new(Field::new(
                            &col.name,
                            col.data_type.as_arrow_type(),
                            col.is_nullable(),
                        ));
                        let dict_field = tag_maybe_to_dictionary_field(&col.data_type, &field);
                        let mut column = col.clone();
                        column.data_type =
                            ConcreteDataType::from_arrow_type(dict_field.data_type());
                        return column;
                    }

                    col.clone()
                })
                .collect::<Vec<_>>(),
        ));

        let region_partition_physical_expr = region_partition_expr
            .try_as_physical_expr(partition_schema.arrow_schema())
            .context(SerializePartitionExprSnafu)?;

        Ok(Some(PartitionFilterContext {
            region_partition_physical_expr,
            partition_schema,
        }))
    }

    /// Reads parquet metadata of specific file.
    /// Returns (fused metadata, cache_miss_flag).
    pub(crate) async fn read_parquet_metadata(
        &self,
        file_path: &str,
        file_size: u64,
        cache_metrics: &mut MetadataCacheMetrics,
        page_index_policy: PageIndexPolicy,
    ) -> Result<(Arc<CachedSstMeta>, bool)> {
        let start = Instant::now();
        let _t = READ_STAGE_ELAPSED
            .with_label_values(&["read_parquet_metadata"])
            .start_timer();

        let file_id = self.file_handle.file_id();
        // Tries to get from cache with metrics tracking.
        if let Some(metadata) = self
            .cache_strategy
            .get_sst_meta_data(file_id, cache_metrics, page_index_policy)
            .await
        {
            cache_metrics.metadata_load_cost += start.elapsed();
            return Ok((metadata, false));
        }

        // Cache miss, load metadata directly.
        let mut metadata_loader =
            MetadataLoader::new(self.object_store.clone(), file_path, file_size);
        metadata_loader.with_page_index_policy(page_index_policy);
        let metadata = metadata_loader.load(cache_metrics).await?;

        let decoded = if self.cache_strategy.sst_meta_cache_enabled() {
            let metadata = prepare_sst_meta(
                file_path,
                metadata,
                None,
                page_index_policy,
                &self.cache_strategy.sst_meta_runtime(),
            )
            .await?;
            let decoded = metadata.decoded();
            match metadata {
                SstMetaPreparation::Prepared(metadata) => {
                    self.cache_strategy
                        .put_prepared_sst_meta(file_id, metadata, true);
                }
                SstMetaPreparation::DecodedOnly { encoding_error, .. } => {
                    warn!(
                        encoding_error;
                        "Failed to encode SST metadata for cache, using decoded metadata for {}",
                        file_path
                    );
                }
            }
            decoded
        } else {
            Arc::new(CachedSstMeta::try_new_with_page_index_policy(
                file_path,
                metadata,
                None,
                page_index_policy,
            )?)
        };

        cache_metrics.metadata_load_cost += start.elapsed();
        Ok((decoded, true))
    }

    /// Computes row groups to read, along with their respective row selections.
    #[tracing::instrument(
        skip_all,
        fields(
            region_id = %self.file_handle.region_id(),
            file_id = %self.file_handle.file_id()
        )
    )]
    async fn row_groups_to_read(
        &self,
        read_format: &FlatReadFormat,
        parquet_meta: &ParquetMetaData,
        metrics: &mut ReaderFilterMetrics,
    ) -> RowGroupSelection {
        let num_row_groups = parquet_meta.num_row_groups();
        let num_rows = parquet_meta.file_metadata().num_rows();
        if num_row_groups == 0 || num_rows == 0 {
            return RowGroupSelection::default();
        }

        // Let's assume that the number of rows in the first row group
        // can represent the `row_group_size` of the Parquet file.
        let row_group_size = parquet_meta.row_group(0).num_rows() as usize;
        if row_group_size == 0 {
            return RowGroupSelection::default();
        }

        metrics.rg_total += num_row_groups;
        metrics.rows_total += num_rows as usize;

        // Compute skip_fields once for all pruning operations
        let skip_fields = self.pre_filter_mode.skip_fields();

        let mut output = self.row_groups_by_minmax(
            read_format,
            parquet_meta,
            row_group_size,
            num_rows as usize,
            metrics,
            skip_fields,
        );
        if output.is_empty() {
            return output;
        }

        let fulltext_filtered = self
            .prune_row_groups_by_fulltext_index(
                row_group_size,
                num_row_groups,
                &mut output,
                metrics,
                skip_fields,
            )
            .await;
        if output.is_empty() {
            return output;
        }

        self.prune_row_groups_by_inverted_index(
            read_format.metadata(),
            row_group_size,
            parquet_meta,
            &mut output,
            metrics,
            skip_fields,
        )
        .await;
        if output.is_empty() {
            return output;
        }

        self.prune_row_groups_by_bloom_filter(
            read_format.metadata(),
            row_group_size,
            parquet_meta,
            &mut output,
            metrics,
            skip_fields,
        )
        .await;
        if output.is_empty() {
            return output;
        }

        if !fulltext_filtered {
            self.prune_row_groups_by_fulltext_bloom(
                row_group_size,
                parquet_meta,
                &mut output,
                metrics,
                skip_fields,
            )
            .await;
        }
        output
    }

    /// Prunes row groups by fulltext index. Returns `true` if the row groups are pruned.
    async fn prune_row_groups_by_fulltext_index(
        &self,
        row_group_size: usize,
        num_row_groups: usize,
        output: &mut RowGroupSelection,
        metrics: &mut ReaderFilterMetrics,
        skip_fields: bool,
    ) -> bool {
        if !self.file_handle.meta_ref().fulltext_index_available() {
            return false;
        }

        let mut pruned = false;
        // If skip_fields is true, only apply the first applier (for tags).
        let appliers = if skip_fields {
            &self.fulltext_index_appliers[..1]
        } else {
            &self.fulltext_index_appliers[..]
        };
        for index_applier in appliers.iter().flatten() {
            let predicate_key = index_applier.predicate_key();
            // Fast path: return early if the result is in the cache.
            let cached = self
                .cache_strategy
                .index_result_cache()
                .and_then(|cache| cache.get(predicate_key, self.file_handle.file_id().file_id()));
            if let Some(result) = cached.as_ref()
                && all_required_row_groups_searched(output, result)
            {
                apply_selection_and_update_metrics(output, result, metrics, INDEX_TYPE_FULLTEXT);
                metrics.fulltext_index_cache_hit += 1;
                pruned = true;
                continue;
            }

            // Slow path: apply the index from the file.
            metrics.fulltext_index_cache_miss += 1;
            let file_size_hint = self.file_handle.meta_ref().index_file_size();
            let apply_res = index_applier
                .apply_fine(
                    self.file_handle.index_id(),
                    Some(file_size_hint),
                    metrics.fulltext_index_apply_metrics.as_mut(),
                )
                .await;
            let selection = match apply_res {
                Ok(Some(res)) => {
                    RowGroupSelection::from_row_ids(res, row_group_size, num_row_groups)
                }
                Ok(None) => continue,
                Err(err) => {
                    handle_index_error!(err, self.file_handle, INDEX_TYPE_FULLTEXT);
                    continue;
                }
            };

            self.apply_index_result_and_update_cache(
                predicate_key,
                self.file_handle.file_id().file_id(),
                selection,
                output,
                metrics,
                INDEX_TYPE_FULLTEXT,
            );
            pruned = true;
        }
        pruned
    }

    /// Applies index to prune row groups.
    ///
    /// TODO(zhongzc): Devise a mechanism to enforce the non-use of indices
    /// as an escape route in case of index issues, and it can be used to test
    /// the correctness of the index.
    async fn prune_row_groups_by_inverted_index(
        &self,
        sst_metadata: &RegionMetadataRef,
        row_group_size: usize,
        parquet_meta: &ParquetMetaData,
        output: &mut RowGroupSelection,
        metrics: &mut ReaderFilterMetrics,
        skip_fields: bool,
    ) -> bool {
        if !self.file_handle.meta_ref().inverted_index_available() {
            return false;
        }

        let num_row_groups = parquet_meta.num_row_groups();
        let total_row_count = parquet_meta.file_metadata().num_rows() as usize;
        let mut pruned = false;
        // If skip_fields is true, only apply the first applier (for tags).
        let appliers = if skip_fields {
            &self.inverted_index_appliers[..1]
        } else {
            &self.inverted_index_appliers[..]
        };
        for index_applier in appliers.iter().flatten() {
            let Ok(Some(plan)) = index_applier
                .plan_for_sst(sst_metadata)
                .inspect_err(|e| warn!(e; "failed to build compatible plan for sst"))
            else {
                continue;
            };

            // Fast path: return early if the result is in the cache.
            let cached = self.cache_strategy.index_result_cache().and_then(|cache| {
                let file_id = self.file_handle.file_id().file_id();
                cache.get(&plan.predicate_key, file_id)
            });

            if let Some(result) = cached.as_ref()
                && all_required_row_groups_searched(output, result)
            {
                apply_selection_and_update_metrics(output, result, metrics, INDEX_TYPE_INVERTED);
                metrics.inverted_index_cache_hit += 1;
                pruned = true;
                continue;
            }

            // Slow path: apply the index from the file.
            metrics.inverted_index_cache_miss += 1;
            let file_size_hint = self.file_handle.meta_ref().index_file_size();
            let apply_res = index_applier
                .apply(
                    self.file_handle.index_id(),
                    Some(file_size_hint),
                    &plan.index_applier,
                    metrics.inverted_index_apply_metrics.as_mut(),
                )
                .await;

            let selection = match apply_res {
                Ok(apply_output) => {
                    let index_row_count = apply_output.total_row_count;
                    let Some(selection) = RowGroupSelection::from_inverted_index_apply_output(
                        row_group_size,
                        num_row_groups,
                        total_row_count,
                        apply_output,
                    ) else {
                        warn!(
                            "Ignore inverted index with mismatched row count, file_id: {:?}, index_row_count: {}, parquet_row_count: {}",
                            self.file_handle.file_id(),
                            index_row_count,
                            total_row_count,
                        );
                        continue;
                    };
                    selection
                }
                Err(err) => {
                    handle_index_error!(err, self.file_handle, INDEX_TYPE_INVERTED);
                    continue;
                }
            };

            self.apply_index_result_and_update_cache(
                &plan.predicate_key,
                self.file_handle.file_id().file_id(),
                selection,
                output,
                metrics,
                INDEX_TYPE_INVERTED,
            );
            pruned = true;
        }
        pruned
    }

    async fn prune_row_groups_by_bloom_filter(
        &self,
        sst_metadata: &RegionMetadataRef,
        row_group_size: usize,
        parquet_meta: &ParquetMetaData,
        output: &mut RowGroupSelection,
        metrics: &mut ReaderFilterMetrics,
        skip_fields: bool,
    ) -> bool {
        if !self.file_handle.meta_ref().bloom_filter_index_available() {
            return false;
        }

        let mut pruned = false;
        // If skip_fields is true, only apply the first applier (for tags).
        let appliers = if skip_fields {
            &self.bloom_filter_index_appliers[..1]
        } else {
            &self.bloom_filter_index_appliers[..]
        };
        for index_applier in appliers.iter().flatten() {
            let Some(compatible_predicates) =
                index_applier.compatible_predicate_for_sst(sst_metadata)
            else {
                continue;
            };
            let predicate_key = PredicateKey::new_bloom(compatible_predicates.clone());
            // Fast path: return early if the result is in the cache.
            let cached = self.cache_strategy.index_result_cache().and_then(|cache| {
                let file_id = self.file_handle.file_id().file_id();
                cache.get(&predicate_key, file_id)
            });
            if let Some(result) = cached.as_ref()
                && all_required_row_groups_searched(output, result)
            {
                apply_selection_and_update_metrics(output, result, metrics, INDEX_TYPE_BLOOM);
                metrics.bloom_filter_cache_hit += 1;
                pruned = true;
                continue;
            }

            // Slow path: apply the index from the file.
            metrics.bloom_filter_cache_miss += 1;
            let file_size_hint = self.file_handle.meta_ref().index_file_size();
            let rgs = parquet_meta.row_groups().iter().enumerate().map(|(i, rg)| {
                (
                    rg.num_rows() as usize,
                    // Optimize: only search the row group that required by `output` and not stored in `cached`.
                    output.contains_non_empty_row_group(i)
                        && cached
                            .as_ref()
                            .map(|c| !c.contains_row_group(i))
                            .unwrap_or(true),
                )
            });
            let apply_res = index_applier
                .apply(
                    self.file_handle.index_id(),
                    Some(file_size_hint),
                    &compatible_predicates,
                    rgs,
                    metrics.bloom_filter_apply_metrics.as_mut(),
                )
                .await;
            let mut selection = match apply_res {
                Ok(apply_output) => {
                    RowGroupSelection::from_row_ranges(apply_output, row_group_size)
                }
                Err(err) => {
                    handle_index_error!(err, self.file_handle, INDEX_TYPE_BLOOM);
                    continue;
                }
            };

            // New searched row groups are added to `selection`, concat them with `cached`.
            if let Some(cached) = cached.as_ref() {
                selection.concat(cached);
            }

            self.apply_index_result_and_update_cache(
                &predicate_key,
                self.file_handle.file_id().file_id(),
                selection,
                output,
                metrics,
                INDEX_TYPE_BLOOM,
            );
            pruned = true;
        }
        pruned
    }

    async fn prune_row_groups_by_fulltext_bloom(
        &self,
        row_group_size: usize,
        parquet_meta: &ParquetMetaData,
        output: &mut RowGroupSelection,
        metrics: &mut ReaderFilterMetrics,
        skip_fields: bool,
    ) -> bool {
        if !self.file_handle.meta_ref().fulltext_index_available() {
            return false;
        }

        let mut pruned = false;
        // If skip_fields is true, only apply the first applier (for tags).
        let appliers = if skip_fields {
            &self.fulltext_index_appliers[..1]
        } else {
            &self.fulltext_index_appliers[..]
        };
        for index_applier in appliers.iter().flatten() {
            let predicate_key = index_applier.predicate_key();
            // Fast path: return early if the result is in the cache.
            let cached = self
                .cache_strategy
                .index_result_cache()
                .and_then(|cache| cache.get(predicate_key, self.file_handle.file_id().file_id()));
            if let Some(result) = cached.as_ref()
                && all_required_row_groups_searched(output, result)
            {
                apply_selection_and_update_metrics(output, result, metrics, INDEX_TYPE_FULLTEXT);
                metrics.fulltext_index_cache_hit += 1;
                pruned = true;
                continue;
            }

            // Slow path: apply the index from the file.
            metrics.fulltext_index_cache_miss += 1;
            let file_size_hint = self.file_handle.meta_ref().index_file_size();
            let rgs = parquet_meta.row_groups().iter().enumerate().map(|(i, rg)| {
                (
                    rg.num_rows() as usize,
                    // Optimize: only search the row group that required by `output` and not stored in `cached`.
                    output.contains_non_empty_row_group(i)
                        && cached
                            .as_ref()
                            .map(|c| !c.contains_row_group(i))
                            .unwrap_or(true),
                )
            });
            let apply_res = index_applier
                .apply_coarse(
                    self.file_handle.index_id(),
                    Some(file_size_hint),
                    rgs,
                    metrics.fulltext_index_apply_metrics.as_mut(),
                )
                .await;
            let mut selection = match apply_res {
                Ok(Some(apply_output)) => {
                    RowGroupSelection::from_row_ranges(apply_output, row_group_size)
                }
                Ok(None) => continue,
                Err(err) => {
                    handle_index_error!(err, self.file_handle, INDEX_TYPE_FULLTEXT);
                    continue;
                }
            };

            // New searched row groups are added to `selection`, concat them with `cached`.
            if let Some(cached) = cached.as_ref() {
                selection.concat(cached);
            }

            self.apply_index_result_and_update_cache(
                predicate_key,
                self.file_handle.file_id().file_id(),
                selection,
                output,
                metrics,
                INDEX_TYPE_FULLTEXT,
            );
            pruned = true;
        }
        pruned
    }

    /// Computes row groups selection after min-max pruning.
    fn row_groups_by_minmax(
        &self,
        read_format: &FlatReadFormat,
        parquet_meta: &ParquetMetaData,
        row_group_size: usize,
        total_row_count: usize,
        metrics: &mut ReaderFilterMetrics,
        skip_fields: bool,
    ) -> RowGroupSelection {
        let Some(predicate) = &self.predicate else {
            return RowGroupSelection::new(row_group_size, total_row_count);
        };

        let file_id = self.file_handle.file_id().file_id();
        let index_result_cache = self.cache_strategy.index_result_cache();
        let cached_minmax_key =
            if index_result_cache.is_some() && predicate.dyn_filters().is_empty() {
                // Cache min-max pruning results keyed by predicate expressions. This avoids repeatedly
                // building row-group pruning stats for identical predicates across queries.
                let mut exprs = predicate
                    .exprs()
                    .iter()
                    .map(|expr| format!("{expr:?}"))
                    .collect::<Vec<_>>();
                exprs.sort();
                let schema_version = self
                    .expected_metadata
                    .as_ref()
                    .map(|meta| meta.schema_version)
                    .unwrap_or_else(|| read_format.metadata().schema_version);
                Some(PredicateKey::new_minmax(
                    Arc::new(exprs),
                    schema_version,
                    skip_fields,
                ))
            } else {
                None
            };

        if let Some(index_result_cache) = index_result_cache
            && let Some(predicate_key) = cached_minmax_key.as_ref()
        {
            if let Some(result) = index_result_cache.get(predicate_key, file_id) {
                metrics.minmax_cache_hit += 1;
                let num_row_groups = parquet_meta.num_row_groups();
                metrics.rg_minmax_filtered +=
                    num_row_groups.saturating_sub(result.row_group_count());
                return (*result).clone();
            }

            metrics.minmax_cache_miss += 1;
        }

        let region_meta = read_format.metadata();
        let row_groups = parquet_meta.row_groups();
        let stats = RowGroupPruningStats::new(
            row_groups,
            read_format,
            self.expected_metadata.clone(),
            skip_fields,
        );
        let prune_schema = self
            .expected_metadata
            .as_ref()
            .map(|meta| meta.schema.arrow_schema())
            .unwrap_or_else(|| region_meta.schema.arrow_schema());

        // Here we use the schema of the SST to build the physical expression. If the column
        // in the SST doesn't have the same column id as the column in the expected metadata,
        // we will get a None statistics for that column.
        let mask = predicate.prune_with_stats(&stats, prune_schema);
        let output = RowGroupSelection::from_full_row_group_ids(
            mask.iter()
                .enumerate()
                .filter_map(|(row_group, keep)| keep.then_some(row_group)),
            row_group_size,
            total_row_count,
        );

        metrics.rg_minmax_filtered += parquet_meta
            .num_row_groups()
            .saturating_sub(output.row_group_count());

        if let Some(index_result_cache) = index_result_cache
            && let Some(predicate_key) = cached_minmax_key
        {
            index_result_cache.put(predicate_key, file_id, Arc::new(output.clone()));
        }

        output
    }

    fn apply_index_result_and_update_cache(
        &self,
        predicate_key: &PredicateKey,
        file_id: FileId,
        result: RowGroupSelection,
        output: &mut RowGroupSelection,
        metrics: &mut ReaderFilterMetrics,
        index_type: &str,
    ) {
        apply_selection_and_update_metrics(output, &result, metrics, index_type);

        if let Some(index_result_cache) = &self.cache_strategy.index_result_cache() {
            index_result_cache.put(predicate_key.clone(), file_id, Arc::new(result));
        }
    }
}

fn has_row_level_selection(selection: &RowGroupSelection, parquet_meta: &ParquetMetaData) -> bool {
    selection.iter().any(|(row_group_idx, row_selection)| {
        let Some(row_group) = parquet_meta.row_groups().get(*row_group_idx) else {
            return false;
        };

        row_selection.row_count() != row_group.num_rows() as usize
            || row_selection.iter().any(|selector| selector.skip)
    })
}

fn apply_selection_and_update_metrics(
    output: &mut RowGroupSelection,
    result: &RowGroupSelection,
    metrics: &mut ReaderFilterMetrics,
    index_type: &str,
) {
    let intersection = output.intersect(result);

    let row_group_count = output.row_group_count() - intersection.row_group_count();
    let row_count = output.row_count() - intersection.row_count();

    metrics.update_index_metrics(index_type, row_group_count, row_count);

    *output = intersection;
}

fn all_required_row_groups_searched(
    required_row_groups: &RowGroupSelection,
    cached_row_groups: &RowGroupSelection,
) -> bool {
    required_row_groups.iter().all(|(rg_id, _)| {
        // Row group with no rows is not required to search.
        !required_row_groups.contains_non_empty_row_group(*rg_id)
            // The row group is already searched.
            || cached_row_groups.contains_row_group(*rg_id)
    })
}

/// Metrics of filtering rows groups and rows.
#[derive(Debug, Default, Clone)]
pub(crate) struct ReaderFilterMetrics {
    /// Number of row groups before filtering.
    pub(crate) rg_total: usize,
    /// Number of row groups filtered by fulltext index.
    pub(crate) rg_fulltext_filtered: usize,
    /// Number of row groups filtered by inverted index.
    pub(crate) rg_inverted_filtered: usize,
    /// Number of row groups filtered by min-max index.
    pub(crate) rg_minmax_filtered: usize,
    /// Number of row groups filtered by bloom filter index.
    pub(crate) rg_bloom_filtered: usize,

    /// Number of rows in row group before filtering.
    pub(crate) rows_total: usize,
    /// Number of rows in row group filtered by fulltext index.
    pub(crate) rows_fulltext_filtered: usize,
    /// Number of rows in row group filtered by inverted index.
    pub(crate) rows_inverted_filtered: usize,
    /// Number of rows in row group filtered by bloom filter index.
    pub(crate) rows_bloom_filtered: usize,
    /// Number of rows filtered by precise filter.
    pub(crate) rows_precise_filtered: usize,

    /// Number of index result cache hits for fulltext index.
    pub(crate) fulltext_index_cache_hit: usize,
    /// Number of index result cache misses for fulltext index.
    pub(crate) fulltext_index_cache_miss: usize,
    /// Number of index result cache hits for inverted index.
    pub(crate) inverted_index_cache_hit: usize,
    /// Number of index result cache misses for inverted index.
    pub(crate) inverted_index_cache_miss: usize,
    /// Number of index result cache hits for bloom filter index.
    pub(crate) bloom_filter_cache_hit: usize,
    /// Number of index result cache misses for bloom filter index.
    pub(crate) bloom_filter_cache_miss: usize,
    /// Number of index result cache hits for minmax pruning.
    pub(crate) minmax_cache_hit: usize,
    /// Number of index result cache misses for minmax pruning.
    pub(crate) minmax_cache_miss: usize,

    /// Optional metrics for inverted index applier.
    pub(crate) inverted_index_apply_metrics: Option<InvertedIndexApplyMetrics>,
    /// Optional metrics for bloom filter index applier.
    pub(crate) bloom_filter_apply_metrics: Option<BloomFilterIndexApplyMetrics>,
    /// Optional metrics for fulltext index applier.
    pub(crate) fulltext_index_apply_metrics: Option<FulltextIndexApplyMetrics>,

    /// Number of pruner builder cache hits.
    pub(crate) pruner_cache_hit: usize,
    /// Number of pruner builder cache misses.
    pub(crate) pruner_cache_miss: usize,
    /// Duration spent waiting for pruner to build file ranges.
    pub(crate) pruner_prune_cost: Duration,
    /// Number of files filtered by manifest time-range pruning.
    pub(crate) files_time_range_pruned: usize,
}

impl ReaderFilterMetrics {
    /// Adds `other` metrics to this metrics.
    pub(crate) fn merge_from(&mut self, other: &ReaderFilterMetrics) {
        self.rg_total += other.rg_total;
        self.rg_fulltext_filtered += other.rg_fulltext_filtered;
        self.rg_inverted_filtered += other.rg_inverted_filtered;
        self.rg_minmax_filtered += other.rg_minmax_filtered;
        self.rg_bloom_filtered += other.rg_bloom_filtered;

        self.rows_total += other.rows_total;
        self.rows_fulltext_filtered += other.rows_fulltext_filtered;
        self.rows_inverted_filtered += other.rows_inverted_filtered;
        self.rows_bloom_filtered += other.rows_bloom_filtered;
        self.rows_precise_filtered += other.rows_precise_filtered;

        self.fulltext_index_cache_hit += other.fulltext_index_cache_hit;
        self.fulltext_index_cache_miss += other.fulltext_index_cache_miss;
        self.inverted_index_cache_hit += other.inverted_index_cache_hit;
        self.inverted_index_cache_miss += other.inverted_index_cache_miss;
        self.bloom_filter_cache_hit += other.bloom_filter_cache_hit;
        self.bloom_filter_cache_miss += other.bloom_filter_cache_miss;
        self.minmax_cache_hit += other.minmax_cache_hit;
        self.minmax_cache_miss += other.minmax_cache_miss;

        self.pruner_cache_hit += other.pruner_cache_hit;
        self.pruner_cache_miss += other.pruner_cache_miss;
        self.pruner_prune_cost += other.pruner_prune_cost;
        self.files_time_range_pruned += other.files_time_range_pruned;

        // Merge optional applier metrics
        if let Some(other_metrics) = &other.inverted_index_apply_metrics {
            self.inverted_index_apply_metrics
                .get_or_insert_with(Default::default)
                .merge_from(other_metrics);
        }
        if let Some(other_metrics) = &other.bloom_filter_apply_metrics {
            self.bloom_filter_apply_metrics
                .get_or_insert_with(Default::default)
                .merge_from(other_metrics);
        }
        if let Some(other_metrics) = &other.fulltext_index_apply_metrics {
            self.fulltext_index_apply_metrics
                .get_or_insert_with(Default::default)
                .merge_from(other_metrics);
        }
    }

    /// Reports metrics.
    pub(crate) fn observe(&self) {
        READ_ROW_GROUPS_TOTAL
            .with_label_values(&["before_filtering"])
            .inc_by(self.rg_total as u64);
        READ_ROW_GROUPS_TOTAL
            .with_label_values(&["fulltext_index_filtered"])
            .inc_by(self.rg_fulltext_filtered as u64);
        READ_ROW_GROUPS_TOTAL
            .with_label_values(&["inverted_index_filtered"])
            .inc_by(self.rg_inverted_filtered as u64);
        READ_ROW_GROUPS_TOTAL
            .with_label_values(&["minmax_index_filtered"])
            .inc_by(self.rg_minmax_filtered as u64);
        READ_ROW_GROUPS_TOTAL
            .with_label_values(&["bloom_filter_index_filtered"])
            .inc_by(self.rg_bloom_filtered as u64);

        PRECISE_FILTER_ROWS_TOTAL
            .with_label_values(&["parquet"])
            .inc_by(self.rows_precise_filtered as u64);
        READ_ROWS_IN_ROW_GROUP_TOTAL
            .with_label_values(&["before_filtering"])
            .inc_by(self.rows_total as u64);
        READ_ROWS_IN_ROW_GROUP_TOTAL
            .with_label_values(&["fulltext_index_filtered"])
            .inc_by(self.rows_fulltext_filtered as u64);
        READ_ROWS_IN_ROW_GROUP_TOTAL
            .with_label_values(&["inverted_index_filtered"])
            .inc_by(self.rows_inverted_filtered as u64);
        READ_ROWS_IN_ROW_GROUP_TOTAL
            .with_label_values(&["bloom_filter_index_filtered"])
            .inc_by(self.rows_bloom_filtered as u64);
    }

    fn update_index_metrics(&mut self, index_type: &str, row_group_count: usize, row_count: usize) {
        match index_type {
            INDEX_TYPE_FULLTEXT => {
                self.rg_fulltext_filtered += row_group_count;
                self.rows_fulltext_filtered += row_count;
            }
            INDEX_TYPE_INVERTED => {
                self.rg_inverted_filtered += row_group_count;
                self.rows_inverted_filtered += row_count;
            }
            INDEX_TYPE_BLOOM => {
                self.rg_bloom_filtered += row_group_count;
                self.rows_bloom_filtered += row_count;
            }
            _ => {}
        }
    }
}

/// Metrics for parquet metadata cache operations.
#[derive(Default, Clone, Copy)]
pub struct MetadataCacheMetrics {
    /// Number of memory cache hits for parquet metadata.
    pub mem_cache_hit: usize,
    /// Number of file cache hits for parquet metadata.
    pub file_cache_hit: usize,
    /// Number of cache misses for parquet metadata.
    pub cache_miss: usize,
    /// Duration to load parquet metadata.
    pub metadata_load_cost: Duration,
    /// Number of read operations performed.
    pub num_reads: usize,
    /// Total bytes read from storage.
    pub bytes_read: u64,
}

impl std::fmt::Debug for MetadataCacheMetrics {
    fn fmt(&self, f: &mut std::fmt::Formatter<'_>) -> std::fmt::Result {
        let Self {
            mem_cache_hit,
            file_cache_hit,
            cache_miss,
            metadata_load_cost,
            num_reads,
            bytes_read,
        } = self;

        if self.is_empty() {
            return write!(f, "{{}}");
        }
        write!(f, "{{")?;

        write!(f, "\"metadata_load_cost\":\"{:?}\"", metadata_load_cost)?;

        if *mem_cache_hit > 0 {
            write!(f, ", \"mem_cache_hit\":{}", mem_cache_hit)?;
        }
        if *file_cache_hit > 0 {
            write!(f, ", \"file_cache_hit\":{}", file_cache_hit)?;
        }
        if *cache_miss > 0 {
            write!(f, ", \"cache_miss\":{}", cache_miss)?;
        }
        if *num_reads > 0 {
            write!(f, ", \"num_reads\":{}", num_reads)?;
        }
        if *bytes_read > 0 {
            write!(f, ", \"bytes_read\":{}", bytes_read)?;
        }

        write!(f, "}}")
    }
}

impl MetadataCacheMetrics {
    /// Returns true if the metrics are empty (contain no meaningful data).
    pub(crate) fn is_empty(&self) -> bool {
        self.metadata_load_cost.is_zero()
    }

    /// Adds `other` metrics to this metrics.
    pub(crate) fn merge_from(&mut self, other: &MetadataCacheMetrics) {
        self.mem_cache_hit += other.mem_cache_hit;
        self.file_cache_hit += other.file_cache_hit;
        self.cache_miss += other.cache_miss;
        self.metadata_load_cost += other.metadata_load_cost;
        self.num_reads += other.num_reads;
        self.bytes_read += other.bytes_read;
    }
}

/// Parquet reader metrics.
#[derive(Debug, Default, Clone)]
pub struct ReaderMetrics {
    /// Filtered row groups and rows metrics.
    pub(crate) filter_metrics: ReaderFilterMetrics,
    /// Duration to build the parquet reader.
    pub(crate) build_cost: Duration,
    /// Duration to scan the reader.
    pub(crate) scan_cost: Duration,
    /// Number of record batches read.
    pub(crate) num_record_batches: usize,
    /// Number of batches decoded.
    pub(crate) num_batches: usize,
    /// Number of rows read.
    pub(crate) num_rows: usize,
    /// Metrics for parquet metadata cache.
    pub(crate) metadata_cache_metrics: MetadataCacheMetrics,
    /// Optional metrics for page/row group fetch operations.
    pub(crate) fetch_metrics: Option<Arc<ParquetFetchMetrics>>,
    /// Memory size of metadata loaded for building file ranges.
    pub(crate) metadata_mem_size: isize,
    /// Number of file range builders created.
    pub(crate) num_range_builders: isize,
}

impl ReaderMetrics {
    /// Adds `other` metrics to this metrics.
    pub(crate) fn merge_from(&mut self, other: &ReaderMetrics) {
        self.filter_metrics.merge_from(&other.filter_metrics);
        self.build_cost += other.build_cost;
        self.scan_cost += other.scan_cost;
        self.num_record_batches += other.num_record_batches;
        self.num_batches += other.num_batches;
        self.num_rows += other.num_rows;
        self.metadata_cache_metrics
            .merge_from(&other.metadata_cache_metrics);
        if let Some(other_fetch) = &other.fetch_metrics {
            if let Some(self_fetch) = &self.fetch_metrics {
                self_fetch.merge_from(other_fetch);
            } else {
                self.fetch_metrics = Some(other_fetch.clone());
            }
        }
        self.metadata_mem_size += other.metadata_mem_size;
        self.num_range_builders += other.num_range_builders;
    }

    /// Reports total rows.
    pub(crate) fn observe_rows(&self, read_type: &str) {
        READ_ROWS_TOTAL
            .with_label_values(&[read_type])
            .inc_by(self.num_rows as u64);
    }
}

/// Builder to build a parquet record batch stream for a row group.
pub(crate) struct RowGroupReaderBuilder {
    /// SST file to read.
    ///
    /// Holds the file handle to avoid the file purge it.
    file_handle: FileHandle,
    /// Path of the file.
    file_path: String,
    /// Metadata of the parquet file.
    parquet_meta: Arc<ParquetMetaData>,
    /// Immutable metadata size, computed once when the footer is decoded.
    parquet_metadata_size: usize,
    /// Arrow reader metadata for building async stream.
    arrow_metadata: ArrowReaderMetadata,
    /// Projected output schema aligned with `projection.projected_root_presence`.
    output_schema: SchemaRef,
    /// JSON2 columns that must be semantically rewritten into the projected target layout.
    json2_rewrite_targets: HashMap<String, Json2TargetLayout>,
    /// Object store as an Operator.
    object_store: ObjectStore,
    /// Projection mask.
    projection: ProjectionMaskPlan,
    /// Whether projected read columns include nested paths.
    has_nested_projection: bool,
    /// Cache.
    cache_strategy: CacheStrategy,
    /// Pre-built prefilter state. `None` if prefiltering is not applicable.
    prefilter_builder: Option<PrefilterContextBuilder>,
    /// Hint for rows in a decoded batch.
    batch_size: usize,
}

/// Context passed to [RowGroupReaderBuilder::build()] carrying all information
/// needed for prefiltering decisions.
pub(crate) struct RowGroupBuildContext<'a> {
    /// Index of the row group to read.
    pub(crate) row_group_idx: usize,
    /// Row selection for the row group. `None` means all rows.
    pub(crate) row_selection: Option<RowSelection>,
    /// Metrics for tracking fetch operations.
    pub(crate) fetch_metrics: Option<&'a ParquetFetchMetrics>,
}

impl RowGroupReaderBuilder {
    /// Path of the file to read.
    pub(crate) fn file_path(&self) -> &str {
        &self.file_path
    }

    /// Handle of the file to read.
    pub(crate) fn file_handle(&self) -> &FileHandle {
        &self.file_handle
    }

    pub(crate) fn parquet_metadata(&self) -> &Arc<ParquetMetaData> {
        &self.parquet_meta
    }

    pub(crate) fn parquet_metadata_size(&self) -> usize {
        self.parquet_metadata_size
    }

    pub(crate) fn cache_strategy(&self) -> &CacheStrategy {
        &self.cache_strategy
    }

    pub(crate) fn has_predicate_prefilter(&self) -> bool {
        self.prefilter_builder.is_some()
    }

    /// Builds a parquet record batch stream to read the row group at `row_group_idx`.
    ///
    /// If prefiltering is applicable (based on `build_ctx`), this performs a two-phase read:
    /// 1. Reads only the prefilter columns (e.g. PK column), applies filters to get a refined row selection
    /// 2. Reads the full projection with the refined row selection
    ///
    /// The prefilter pass is *best-effort pruning*, not the precise filter for the query.
    /// Predicates that cannot be lowered to prefilter columns (column not projected,
    /// expression not supported, etc.) are silently skipped. Correctness rests on the
    /// DataFusion `FilterExec` above this reader, which always re-applies the original
    /// predicate. With predicate prefiltering enabled, tag and timestamp predicates that
    /// flow through [`SimpleFilterEvaluator`] are enforced precisely in this pass. See
    /// [`build_reader_filter_plan`] for the bucketing rules and disabled mode.
    ///
    /// When the prefilter result selects no rows, the second read still issues but
    /// parquet-rs short-circuits before any column-chunk IO: the row-group state machine
    /// jumps to `Finished` once it sees `num_rows_selected() == 0`, so no fast path is
    /// added here.
    pub(crate) async fn build(
        &self,
        build_ctx: RowGroupBuildContext<'_>,
    ) -> Result<ProjectedRecordBatchStream> {
        let prefilter_ctx = self
            .prefilter_builder
            .as_ref()
            .map(|b| b.build(build_ctx.row_group_idx));

        let Some(mut prefilter_ctx) = prefilter_ctx else {
            // No prefilter applicable, build stream with full projection.
            let stream = self
                .build_with_projection(
                    build_ctx.row_group_idx,
                    build_ctx.row_selection,
                    self.projection.mask.clone(),
                    build_ctx.fetch_metrics,
                )
                .await?;
            return self.make_projected_stream(stream);
        };

        let prefilter_start = Instant::now();
        let prefilter_result = execute_prefilter(&mut prefilter_ctx, self, &build_ctx).await?;
        if let Some(metrics) = build_ctx.fetch_metrics {
            let mut data = metrics.data.lock().unwrap();
            data.prefilter_cost += prefilter_start.elapsed();
            data.prefilter_filtered_rows += prefilter_result.filtered_rows;
        }

        let refined_selection = Some(prefilter_result.refined_selection);

        let stream = self
            .build_with_projection(
                build_ctx.row_group_idx,
                refined_selection,
                self.projection.mask.clone(),
                build_ctx.fetch_metrics,
            )
            .await?;
        self.make_projected_stream(stream)
    }

    /// Builds the normal projection without running the generic predicate prefilter.
    ///
    /// The series reader uses this after computing its own primary-key-only row
    /// selection.
    pub(crate) async fn build_without_prefilter(
        &self,
        build_ctx: RowGroupBuildContext<'_>,
    ) -> Result<ProjectedRecordBatchStream> {
        let stream = self
            .build_with_projection(
                build_ctx.row_group_idx,
                build_ctx.row_selection,
                self.projection.mask.clone(),
                build_ctx.fetch_metrics,
            )
            .await?;
        self.make_projected_stream(stream)
    }

    /// Builds a stream that reads only the encoded primary-key column.
    ///
    /// It preserves the normal reader's binary-or-dictionary decision. This path deliberately
    /// skips the normal prefilter pass: the caller reads `__primary_key` once and applies all
    /// encoded-primary-key filters to the returned batches.
    pub(crate) async fn build_primary_key(
        &self,
        build_ctx: RowGroupBuildContext<'_>,
    ) -> Result<ProjectedRecordBatchStream> {
        let parquet_schema = self.parquet_meta.file_metadata().schema_descr();
        let primary_key_index = parquet_schema
            .columns()
            .iter()
            .position(|column| column.name() == PRIMARY_KEY_COLUMN_NAME)
            .context(UnexpectedSnafu {
                reason: "SST does not contain __primary_key",
            })?;
        let projection = ProjectionMask::leaves(parquet_schema, [primary_key_index]);

        self.build_with_projection(
            build_ctx.row_group_idx,
            build_ctx.row_selection,
            projection,
            build_ctx.fetch_metrics,
        )
        .await
    }

    fn make_projected_stream(
        &self,
        stream: ProjectedRecordBatchStream,
    ) -> Result<ProjectedRecordBatchStream> {
        if !self.has_nested_projection && self.json2_rewrite_targets.is_empty() {
            return Ok(stream);
        }

        let mode = if self.json2_rewrite_targets.is_empty() {
            AlignMode::AlignToSchema
        } else {
            AlignMode::Rewrite {
                columns: self.json2_rewrite_targets.clone(),
            }
        };

        Ok(JsonSchemaAligner::new(
            stream,
            self.projection.projected_root_presence.clone(),
            self.output_schema.clone(),
            mode,
        )?
        .boxed())
    }

    /// Builds a parquet record batch stream with a custom projection mask.
    pub(crate) async fn build_with_projection(
        &self,
        row_group_idx: usize,
        row_selection: Option<RowSelection>,
        projection: ProjectionMask,
        fetch_metrics: Option<&ParquetFetchMetrics>,
    ) -> Result<ProjectedRecordBatchStream> {
        let range_fetcher = SstParquetRangeFetcher::new(
            self.file_handle.file_id(),
            self.file_path.clone(),
            self.object_store.clone(),
            self.cache_strategy.clone(),
            row_group_idx,
            fetch_metrics.cloned(),
        );

        build_sst_parquet_record_batch_stream(
            self.arrow_metadata.clone(),
            row_group_idx,
            row_selection,
            projection,
            range_fetcher,
            self.file_path.clone(),
            self.batch_size,
        )
    }
}

#[derive(Clone)]
/// The filter to evaluate or the prune result of the default value.
pub(crate) enum MaybeFilter {
    /// The filter to evaluate.
    Filter(SimpleFilterEvaluator),
    /// The filter matches the default value.
    Matched,
    /// The filter is pruned.
    Pruned,
}

impl MaybeFilter {
    /// Returns the inner filter when it is available.
    pub(crate) fn as_filter(&self) -> Option<&SimpleFilterEvaluator> {
        match self {
            MaybeFilter::Filter(filter) => Some(filter),
            MaybeFilter::Matched | MaybeFilter::Pruned => None,
        }
    }
}

#[derive(Clone)]
/// Context to evaluate the column filter for a parquet file.
pub(crate) struct SimpleFilterContext {
    /// Filter to evaluate.
    filter: MaybeFilter,
    /// Debug string of the original logical expression.
    expr_str: String,
    /// Id of the column to evaluate.
    column_id: ColumnId,
    /// Semantic type of the column.
    semantic_type: SemanticType,
}

impl SimpleFilterContext {
    /// Creates a context for the `expr`.
    ///
    /// Returns None if the column to filter doesn't exist in the SST metadata or the
    /// expected metadata.
    pub(crate) fn new_opt(
        sst_meta: &RegionMetadataRef,
        expected_meta: Option<&RegionMetadata>,
        expr: &Expr,
    ) -> Option<Self> {
        let filter = SimpleFilterEvaluator::try_new(expr)?;
        let expr_str = format!("{expr:?}");
        let (column_metadata, maybe_filter) = match expected_meta {
            Some(meta) => {
                // Gets the column metadata from the expected metadata.
                let column = meta.column_by_name(filter.column_name())?;
                // Checks if the column is present in the SST metadata. We still uses the
                // column from the expected metadata.
                match sst_meta.column_by_id(column.column_id) {
                    Some(sst_column) => {
                        debug_assert_eq!(column.semantic_type, sst_column.semantic_type);
                        let maybe_filter = if sst_column.column_schema.data_type
                            == column.column_schema.data_type
                        {
                            MaybeFilter::Filter(filter)
                        } else {
                            // Schema evolution (altered field columns, or a
                            // widened time index unit) can make columns with
                            // the same id have different types across SSTs;
                            // evaluating the original filter may raise an
                            // invalid cross-type comparison. Timestamp
                            // predicates are not re-applied above this scan,
                            // so cast them into the file's unit instead of
                            // dropping them; other mismatches keep the
                            // conservative skip (the query layer re-applies
                            // those predicates).
                            match filter.cast_timestamp_unit(&sst_column.column_schema.data_type) {
                                Some(TimestampUnitCast::Filter(filter)) => {
                                    MaybeFilter::Filter(filter)
                                }
                                Some(TimestampUnitCast::Pruned) => MaybeFilter::Pruned,
                                Some(TimestampUnitCast::Matched) => MaybeFilter::Matched,
                                None => return None,
                            }
                        };
                        (column, maybe_filter)
                    }
                    None => {
                        // If the column is not present in the SST metadata, we evaluate the filter
                        // against the default value of the column.
                        // If we can't evaluate the filter, we return None.
                        if pruned_by_default(&filter, column)? {
                            (column, MaybeFilter::Pruned)
                        } else {
                            (column, MaybeFilter::Matched)
                        }
                    }
                }
            }
            None => {
                let column = sst_meta.column_by_name(filter.column_name())?;
                (column, MaybeFilter::Filter(filter))
            }
        };

        Some(Self {
            filter: maybe_filter,
            expr_str,
            column_id: column_metadata.column_id,
            semantic_type: column_metadata.semantic_type,
        })
    }

    /// Returns the filter to evaluate.
    pub(crate) fn filter(&self) -> &MaybeFilter {
        &self.filter
    }

    /// Returns the original logical expression string.
    pub(crate) fn expr_str(&self) -> &str {
        &self.expr_str
    }

    /// Returns the column id.
    pub(crate) fn column_id(&self) -> ColumnId {
        self.column_id
    }

    /// Returns the semantic type of the column.
    pub(crate) fn semantic_type(&self) -> SemanticType {
        self.semantic_type
    }
}

/// Context to evaluate a physical expression for a parquet file.
#[derive(Clone)]
pub(crate) struct PhysicalFilterContext {
    /// Filter to evaluate.
    filter: Arc<dyn PhysicalExpr>,
    /// Debug string of the original logical expression.
    expr_str: String,
    /// Id of the column to evaluate.
    column_id: ColumnId,
    /// Name of the column to evaluate.
    column_name: String,
    /// Semantic type of the column.
    semantic_type: SemanticType,
    /// Schema containing only the referenced column.
    schema: SchemaRef,
    /// Whether the original logical expression is immutable across queries.
    immutable: bool,
}

impl PhysicalFilterContext {
    /// Creates a context for the `expr`.
    ///
    /// Returns None if the expression doesn't reference exactly one column or the
    /// column to filter doesn't exist in the SST metadata or the expected metadata.
    pub(crate) fn new_opt(
        sst_meta: &RegionMetadataRef,
        expected_meta: Option<&RegionMetadata>,
        read_format: &FlatReadFormat,
        expr: &Expr,
    ) -> Option<Self> {
        if !Self::is_prefilter_candidate(expr) {
            return None;
        }
        let expr_str = format!("{expr:?}");
        let column_name = Self::single_column_name(expr)?;
        let column_metadata = match expected_meta {
            Some(meta) => {
                let column = meta.column_by_name(&column_name)?;
                let sst_column = sst_meta.column_by_id(column.column_id)?;
                // Physical expr requires the column name to match the SST column name.
                if sst_column.column_schema.name != column_name {
                    return None;
                }
                // Schema evolution (an altered field column, or a widened time
                // index unit) can make the column's file type differ from the
                // expected type; evaluating the original expr against the
                // file's column would raise a cross-type comparison error
                // (e.g. Timestamp(ms) >= Timestamp(µs)). Drop the prefilter;
                // the query layer re-applies the predicate above the scan.
                if sst_column.column_schema.data_type != column.column_schema.data_type {
                    return None;
                }
                column
            }
            None => sst_meta.column_by_name(&column_name)?,
        };

        // The column must be present in the projected arrow schema for the
        // prefilter to be able to read it.
        let (_, field) = read_format.arrow_schema().column_with_name(&column_name)?;
        let field = field.clone();
        let schema = Arc::new(ArrowSchema::new(vec![field]));
        let physical_expr = Predicate::to_physical_expr(expr, &schema)
            .inspect_err(|e| {
                error!(e; "Unable to build physical filter for {expr}, schema: {schema:?}");
            })
            .ok()?;
        let immutable = expr_is_immutable(expr);

        Some(Self {
            filter: physical_expr,
            expr_str,
            column_id: column_metadata.column_id,
            column_name,
            semantic_type: column_metadata.semantic_type,
            schema,
            immutable,
        })
    }

    /// Returns true if the expression is a variant we want to evaluate as a
    /// physical prefilter. Binary exprs are intentionally excluded because
    /// [`SimpleFilterEvaluator`] already handles them.
    // TODO(yingwen): extend more expressions if necessary. For example, allow some cheap scalar functions (e.g. `lower`, `length`, date truncations)
    fn is_prefilter_candidate(expr: &Expr) -> bool {
        if !matches!(
            expr,
            Expr::InList(_) | Expr::IsNull(_) | Expr::IsNotNull(_) | Expr::Between(_)
        ) {
            return false;
        }

        // If any functions are found in the expr, it will be not considered as worthy enough to
        // be evaluated in the prefilter. At last, prefilter reads the Parquet files one more time.
        !expr
            .exists(|e| Ok(matches!(e, Expr::ScalarFunction(_))))
            .unwrap_or(false)
    }

    fn single_column_name(expr: &Expr) -> Option<String> {
        let mut columns = HashSet::new();
        if expr_to_columns(expr, &mut columns).is_err() {
            return None;
        }
        if columns.len() != 1 {
            return None;
        }
        columns.iter().next().map(|column| column.name.clone())
    }

    /// Returns the filter to evaluate.
    pub(crate) fn filter(&self) -> &Arc<dyn PhysicalExpr> {
        &self.filter
    }

    /// Returns the original logical expression string.
    pub(crate) fn expr_str(&self) -> &str {
        &self.expr_str
    }

    /// Returns the column id.
    pub(crate) fn column_id(&self) -> ColumnId {
        self.column_id
    }

    /// Returns the column name.
    pub(crate) fn column_name(&self) -> &str {
        &self.column_name
    }

    /// Returns the semantic type of the column.
    pub(crate) fn semantic_type(&self) -> SemanticType {
        self.semantic_type
    }

    /// Returns the schema containing only the referenced column.
    pub(crate) fn schema(&self) -> &SchemaRef {
        &self.schema
    }

    /// Returns true if the original logical expression is immutable across queries.
    pub(crate) fn is_immutable(&self) -> bool {
        self.immutable
    }
}

fn expr_is_immutable(expr: &Expr) -> bool {
    let mut is_immutable = true;
    let _ = expr.apply(|expr| match expr {
        Expr::ScalarFunction(function)
            if function.func.signature().volatility != Volatility::Immutable =>
        {
            is_immutable = false;
            Ok(TreeNodeRecursion::Stop)
        }
        Expr::ScalarVariable(_, _) => {
            is_immutable = false;
            Ok(TreeNodeRecursion::Stop)
        }
        _ => Ok(TreeNodeRecursion::Continue),
    });
    is_immutable
}

/// Prune a column by its default value.
/// Returns false if we can't create the default value or evaluate the filter.
fn pruned_by_default(filter: &SimpleFilterEvaluator, column: &ColumnMetadata) -> Option<bool> {
    let value = column.column_schema.create_default().ok().flatten()?;
    let scalar_value = value
        .try_to_scalar_value(&column.column_schema.data_type)
        .ok()?;
    let matches = filter.evaluate_scalar(&scalar_value).ok()?;
    Some(!matches)
}

/// Parquet batch reader to read our SST format.
pub struct ParquetReader {
    /// File range context.
    context: FileRangeContextRef,
    /// Row group selection to read.
    selection: RowGroupSelection,
    /// Reader of current row group.
    reader: Option<FlatPruneReader>,
    /// Metrics for tracking row group fetch operations.
    fetch_metrics: ParquetFetchMetrics,
}

impl ParquetReader {
    #[tracing::instrument(
        skip_all,
        fields(
            region_id = %self.context.reader_builder().file_handle.region_id(),
            file_id = %self.context.reader_builder().file_handle.file_id()
        )
    )]
    pub async fn next_record_batch(&mut self) -> Result<Option<RecordBatch>> {
        loop {
            if let Some(reader) = &mut self.reader {
                if let Some(batch) = reader.next_batch().await? {
                    return Ok(Some(batch));
                }
                self.reader = None;
                continue;
            }

            let Some((row_group_idx, row_selection)) = self.selection.pop_first() else {
                return Ok(None);
            };

            let skip_fields = self.context.pre_filter_mode().skip_fields();
            let parquet_reader = self
                .context
                .reader_builder()
                .build(self.context.build_context(
                    row_group_idx,
                    Some(row_selection),
                    Some(&self.fetch_metrics),
                ))
                .await?;
            self.reader = Some(FlatPruneReader::new_with_row_group_reader(
                self.context.clone(),
                FlatRowGroupReader::new(self.context.clone(), parquet_reader),
                skip_fields,
            ));
        }
    }
    /// Creates a new reader.
    #[tracing::instrument(
        skip_all,
        fields(
            region_id = %context.reader_builder().file_handle.region_id(),
            file_id = %context.reader_builder().file_handle.file_id()
        )
    )]
    pub(crate) async fn new(
        context: FileRangeContextRef,
        mut selection: RowGroupSelection,
    ) -> Result<Self> {
        let fetch_metrics = ParquetFetchMetrics::default();
        let reader = if let Some((row_group_idx, row_selection)) = selection.pop_first() {
            let skip_fields = context.pre_filter_mode().skip_fields();
            let parquet_reader = context
                .reader_builder()
                .build(context.build_context(
                    row_group_idx,
                    Some(row_selection),
                    Some(&fetch_metrics),
                ))
                .await?;
            Some(FlatPruneReader::new_with_row_group_reader(
                context.clone(),
                FlatRowGroupReader::new(context.clone(), parquet_reader),
                skip_fields,
            ))
        } else {
            None
        };

        Ok(ParquetReader {
            context,
            selection,
            reader,
            fetch_metrics,
        })
    }

    /// Returns the metadata of the SST.
    pub fn metadata(&self) -> &RegionMetadataRef {
        self.context.read_format().metadata()
    }

    pub fn parquet_metadata(&self) -> Arc<ParquetMetaData> {
        self.context.reader_builder().parquet_meta.clone()
    }
}

/// Reader to read a row group of a parquet file in flat format, returning RecordBatch.
pub(crate) struct FlatRowGroupReader {
    /// Context for file ranges.
    context: FileRangeContextRef,
    /// Inner parquet record batch stream.
    stream: ProjectedRecordBatchStream,
    /// Cached sequence array to override sequences.
    override_sequence: Option<ArrayRef>,
}

impl FlatRowGroupReader {
    /// Creates a new flat reader from file range.
    pub(crate) fn new(context: FileRangeContextRef, stream: ProjectedRecordBatchStream) -> Self {
        // The batch length from the reader should be less than or equal to DEFAULT_READ_BATCH_SIZE.
        let override_sequence = context
            .read_format()
            .new_override_sequence_array(DEFAULT_READ_BATCH_SIZE);

        Self {
            context,
            stream,
            override_sequence,
        }
    }

    /// Returns the next RecordBatch.
    pub(crate) async fn next_batch(&mut self) -> Result<Option<RecordBatch>> {
        match self.stream.next().await {
            Some(batch_result) => {
                let record_batch = batch_result?;

                let record_batch = self
                    .context
                    .read_format()
                    .convert_batch(record_batch, self.override_sequence.as_ref())?;
                Ok(Some(record_batch))
            }
            None => Ok(None),
        }
    }
}

#[cfg(test)]
mod tests {
    use std::collections::HashMap;
    use std::fmt::{Debug, Formatter};
    use std::sync::{Arc, LazyLock};

    use api::v1::OpType;
    use common_error::ext::WhateverResult;
    use common_function::scalars::json::json_get::JsonGetWithType;
    use common_function::scalars::udf::create_udf;
    use common_recordbatch::ext::RecordBatchExt;
    use datafusion::arrow::datatypes::DataType;
    use datafusion_common::ScalarValue;
    use datafusion_expr::expr::ScalarFunction;
    use datafusion_expr::{
        ColumnarValue, Expr, ScalarFunctionArgs, ScalarUDF, ScalarUDFImpl, Signature, Volatility,
        col, lit,
    };
    use datatypes::arrow::array::{
        ArrayRef, BinaryDictionaryBuilder, Int64Array, StringArray, StructArray,
        TimestampMillisecondArray, UInt8Array, UInt64Array,
    };
    use datatypes::arrow::datatypes::{Fields, Schema, UInt32Type};
    use datatypes::arrow::record_batch::RecordBatch;
    use datatypes::extension::json::Json2ExtensionType;
    use datatypes::prelude::ConcreteDataType;
    use datatypes::schema::ColumnSchema;
    use object_store::services::Memory;
    use parquet::arrow::ArrowWriter;
    use parquet::arrow::arrow_reader::RowSelector;
    use parquet::file::properties::WriterProperties;
    use store_api::codec::PrimaryKeyEncoding;
    use store_api::metadata::{ColumnMetadata, RegionMetadata, RegionMetadataBuilder};
    use store_api::region_request::PathType;
    use store_api::storage::RegionId;
    use table::predicate::Predicate;

    use super::*;
    use crate::cache::CacheManager;
    use crate::sst::parquet::metadata::MetadataLoader;
    use crate::sst::parquet::prefilter::{build_reader_filter_plan, execute_prefilter};
    use crate::sst::parquet::read_columns::{ParquetReadColumn, ParquetReadColumns};
    use crate::sst::parquet::row_group::ParquetFetchMetrics;
    use crate::test_util::sst_util::{sst_file_handle, sst_region_metadata};

    async fn prefilter_test_builder(
        object_store: ObjectStore,
        predicate: Predicate,
        cache_strategy: CacheStrategy,
    ) -> (RowGroupReaderBuilder, Arc<RegionMetadata>) {
        let metadata = Arc::new(
            crate::test_util::sst_util::sst_region_metadata_with_encoding(
                PrimaryKeyEncoding::Sparse,
            ),
        );
        let batch = |start: i64, end: i64| {
            let mut primary_key = BinaryDictionaryBuilder::<UInt32Type>::new();
            let mut fields = Vec::new();
            let mut timestamps = Vec::new();
            for value in start..end {
                let tag = if value == 4 { "b" } else { "a" };
                primary_key
                    .append(crate::test_util::sst_util::new_sparse_primary_key(
                        &[tag, "x"],
                        &metadata,
                        1,
                        100,
                    ))
                    .unwrap();
                fields.push(value as u64);
                timestamps.push(value);
            }
            RecordBatch::try_new(
                crate::sst::to_flat_sst_arrow_schema(
                    &metadata,
                    &crate::sst::FlatSchemaOptions::from_encoding(PrimaryKeyEncoding::Sparse),
                ),
                vec![
                    Arc::new(UInt64Array::from(fields)) as ArrayRef,
                    Arc::new(TimestampMillisecondArray::from(timestamps)) as ArrayRef,
                    Arc::new(primary_key.finish()) as ArrayRef,
                    Arc::new(UInt64Array::from_value(1, (end - start) as usize)) as ArrayRef,
                    Arc::new(UInt8Array::from_value(
                        OpType::Put as u8,
                        (end - start) as usize,
                    )) as ArrayRef,
                ],
            )
            .unwrap()
        };
        let first_batch = batch(0, 3);
        let second_batch = batch(3, 6);
        let mut bytes = Vec::new();
        let mut writer = ArrowWriter::try_new(&mut bytes, first_batch.schema(), None).unwrap();
        writer.write(&first_batch).unwrap();
        writer.flush().unwrap();
        writer.write(&second_batch).unwrap();
        writer.close().unwrap();

        let file_handle = sst_file_handle(0, 6);
        let file_path = file_handle.file_path("prefilter_test", PathType::Bare);
        let file_size = bytes.len() as u64;
        object_store.write(&file_path, bytes).await.unwrap();

        let mut cache_metrics = MetadataCacheMetrics::default();
        let parquet_meta = Arc::new(
            MetadataLoader::new(object_store.clone(), &file_path, file_size)
                .load(&mut cache_metrics)
                .await
                .unwrap(),
        );
        let read_format = FlatReadFormat::new(
            metadata.clone(),
            ReadColumns::new(
                metadata
                    .column_metadatas
                    .iter()
                    .map(|column| column.column_id),
            ),
            None,
            &file_path,
            false,
        )
        .unwrap();
        let codec = build_primary_key_codec(metadata.as_ref());
        let filter_plan = build_reader_filter_plan(
            Some(&predicate),
            None,
            PreFilterMode::All,
            true,
            false,
            &read_format,
            &codec,
            &parquet_meta,
        );
        assert!(filter_plan.prefilter_builder.is_some());

        let output_schema = read_format.arrow_schema().clone();
        let parquet_schema = parquet_meta.file_metadata().schema_descr();
        let source_schema = parquet_to_arrow_schema(
            parquet_schema,
            parquet_meta.file_metadata().key_value_metadata(),
        )
        .unwrap();
        let projection = build_projection_plan(
            read_format.parquet_read_columns(),
            parquet_schema,
            &source_schema,
        )
        .unwrap();
        let arrow_metadata =
            ArrowReaderMetadata::try_new(parquet_meta.clone(), ArrowReaderOptions::new()).unwrap();
        (
            RowGroupReaderBuilder {
                file_handle: file_handle.clone(),
                file_path,
                parquet_meta,
                parquet_metadata_size: 0,
                arrow_metadata,
                output_schema,
                json2_rewrite_targets: HashMap::new(),
                object_store,
                projection,
                has_nested_projection: false,
                cache_strategy,
                prefilter_builder: filter_plan.prefilter_builder,
                batch_size: DEFAULT_READ_BATCH_SIZE,
            },
            metadata,
        )
    }

    #[tokio::test(flavor = "current_thread")]
    async fn test_execute_prefilter_proven_filters_preserve_selection_without_fetching() {
        let object_store = ObjectStore::new(Memory::default()).unwrap();
        let predicate = Predicate::new(vec![col("field_0").gt_eq(lit(0_u64))]);
        let (reader_builder, _) =
            prefilter_test_builder(object_store, predicate, CacheStrategy::Disabled).await;
        let prefilter_builder = reader_builder.prefilter_builder.as_ref().unwrap();

        for original_selection in [
            None,
            Some(RowSelection::from(vec![
                RowSelector::skip(1),
                RowSelector::select(1),
                RowSelector::skip(1),
            ])),
            Some(RowSelection::from(vec![])),
        ] {
            let mut prefilter_ctx = prefilter_builder.build(0);
            let fetch_metrics = ParquetFetchMetrics::default();
            let result = execute_prefilter(
                &mut prefilter_ctx,
                &reader_builder,
                &RowGroupBuildContext {
                    row_group_idx: 0,
                    row_selection: original_selection.clone(),
                    fetch_metrics: Some(&fetch_metrics),
                },
            )
            .await
            .unwrap();

            let expected = original_selection.unwrap_or_else(|| {
                RowSelection::from(vec![RowSelector::select(
                    reader_builder.parquet_meta.row_group(0).num_rows() as usize,
                )])
            });
            assert_eq!(result.refined_selection, expected);
            assert_eq!(result.filtered_rows, 0);
            let metrics = fetch_metrics.data.lock().unwrap();
            assert_eq!(metrics.pages_to_fetch_store, 0);
            assert_eq!(metrics.pages_to_fetch_mem, 0);
            assert_eq!(metrics.pages_to_fetch_write_cache, 0);
        }
    }

    #[tokio::test(flavor = "current_thread")]
    async fn test_execute_prefilter_mixed_filters_use_nonzero_row_group_and_cache() {
        let object_store = ObjectStore::new(Memory::default()).unwrap();
        let predicate = Predicate::new(vec![
            col("field_0").gt_eq(lit(3_u64)),
            col("ts").lt(lit(ScalarValue::TimestampMillisecond(Some(6), None))),
            col("field_0").in_list(vec![lit(3_u64), lit(4_u64)], false),
            col("tag_0").eq(lit("a")),
        ]);
        let cache = CacheStrategy::EnableAll(Arc::new(
            CacheManager::builder()
                .prefilter_result_cache_size(1024)
                .build(),
        ));
        let (reader_builder, _) =
            prefilter_test_builder(object_store, predicate, cache.clone()).await;
        let prefilter_builder = reader_builder.prefilter_builder.as_ref().unwrap();

        for pass in 0..2 {
            let mut prefilter_ctx = prefilter_builder.build(1);
            let fetch_metrics = ParquetFetchMetrics::default();
            let fetch_metrics_ref = (pass == 1).then_some(&fetch_metrics);
            let result = execute_prefilter(
                &mut prefilter_ctx,
                &reader_builder,
                &RowGroupBuildContext {
                    row_group_idx: 1,
                    row_selection: Some(RowSelection::from(vec![
                        RowSelector::select(2),
                        RowSelector::skip(1),
                    ])),
                    fetch_metrics: fetch_metrics_ref,
                },
            )
            .await
            .unwrap();
            assert_eq!(result.filtered_rows, 1);
            assert_eq!(
                result.refined_selection,
                RowSelection::from(vec![RowSelector::select(1), RowSelector::skip(2),])
            );
            if pass == 1 {
                assert_eq!(fetch_metrics.data.lock().unwrap().pages_to_fetch_store, 0);
            }
        }

        let disabled = prefilter_test_builder(
            ObjectStore::new(Memory::default()).unwrap(),
            Predicate::new(vec![
                col("field_0").gt_eq(lit(3_u64)),
                col("ts").lt(lit(ScalarValue::TimestampMillisecond(Some(6), None))),
                col("field_0").in_list(vec![lit(3_u64), lit(4_u64)], false),
                col("tag_0").eq(lit("a")),
            ]),
            CacheStrategy::Disabled,
        )
        .await;
        let mut prefilter_ctx = disabled.0.prefilter_builder.as_ref().unwrap().build(1);
        let result = execute_prefilter(
            &mut prefilter_ctx,
            &disabled.0,
            &RowGroupBuildContext {
                row_group_idx: 1,
                row_selection: None,
                fetch_metrics: None,
            },
        )
        .await
        .unwrap();
        assert_eq!(result.filtered_rows, 2);
        assert_eq!(
            result.refined_selection,
            RowSelection::from(vec![RowSelector::select(1), RowSelector::skip(2)])
        );
    }

    #[test]
    fn test_skip_prefilter_for_json_get() -> WhateverResult<()> {
        fn json_get_expr(base: Expr, path: &str) -> Expr {
            let json_get = Arc::new(create_udf(Arc::new(JsonGetWithType::default())));
            Expr::ScalarFunction(ScalarFunction::new_udf(json_get, vec![base, lit(path)]))
        }

        let metadata = Arc::new(sst_region_metadata());
        let format = FlatReadFormat::new(
            metadata.clone(),
            ReadColumns::new(metadata.column_metadatas.iter().map(|c| c.column_id)),
            None,
            "test",
            true,
        )?;
        let new_filter =
            |expr: Expr| PhysicalFilterContext::new_opt(&metadata, None, &format, &expr);

        let json_get = || json_get_expr(col("field_0"), "a.b");

        let regular_expr = col("field_0").is_null();
        assert!(new_filter(regular_expr).is_some());

        let is_null = json_get().is_null();
        assert!(new_filter(is_null).is_none());

        let is_not_null = json_get().is_not_null();
        assert!(new_filter(is_not_null).is_none());

        let in_list = json_get().in_list(vec![lit("value")], false);
        assert!(new_filter(in_list).is_none());

        let in_list_nested = col("field_0").in_list(vec![json_get()], false);
        assert!(new_filter(in_list_nested).is_none());

        let between = json_get().between(lit(1_u64), lit(10_u64));
        assert!(new_filter(between).is_none());

        let between_nested = col("field_0").between(json_get(), lit(10_u64));
        assert!(new_filter(between_nested).is_none());

        Ok(())
    }

    /// A physical prefilter expr must be dropped when the column's file type
    /// differs from the expected type (e.g. an old-unit SST after the time
    /// index unit was widened): evaluating the expected-unit literals against
    /// the file's column raises a cross-unit comparison error.
    #[test]
    fn test_physical_filter_dropped_on_file_type_mismatch() {
        let file_metadata: RegionMetadataRef = Arc::new(sst_region_metadata());
        // The region metadata after widening the `ts` unit to microsecond.
        let mut expected = (*file_metadata).clone();
        for column in expected.column_metadatas.iter_mut() {
            if column.column_schema.name == "ts" {
                column.column_schema.data_type = ConcreteDataType::timestamp_microsecond_datatype();
            }
        }
        let expected = Arc::new(expected);

        let format = FlatReadFormat::new(
            file_metadata.clone(),
            ReadColumns::new(file_metadata.column_metadatas.iter().map(|c| c.column_id)),
            None,
            "test",
            true,
        )
        .unwrap();

        let between = col("ts").between(
            lit(ScalarValue::TimestampMicrosecond(Some(1_000_000), None)),
            lit(ScalarValue::TimestampMicrosecond(Some(2_000_000), None)),
        );
        // Same type: the prefilter is kept.
        assert!(
            PhysicalFilterContext::new_opt(&file_metadata, Some(&file_metadata), &format, &between)
                .is_some()
        );
        // Widened unit: the prefilter is dropped.
        assert!(
            PhysicalFilterContext::new_opt(&file_metadata, Some(&expected), &format, &between)
                .is_none()
        );
    }

    #[tokio::test]
    async fn test_nested_projection_reads_partial_json2_physical_fields() -> WhateverResult<()> {
        // Write a full JSON2-like Arrow struct:
        // j: { a: { x: int, y: string }, b: string }.
        // The test later requests only j.a.x and verifies that the physical Parquet projection
        // does not materialize j.a.y or j.b.

        let xy_fields = Fields::from(vec![
            Arc::new(Field::new("x", DataType::Int64, true)),
            Arc::new(Field::new("y", DataType::Utf8, true)),
        ]);
        let a_field = Arc::new(Field::new("a", DataType::Struct(xy_fields.clone()), true));
        let b_field = Arc::new(Field::new("b", DataType::Utf8, true));
        let json_fields = Fields::from(vec![a_field, b_field]);
        let json_field = Field::new("j", DataType::Struct(json_fields.clone()), true)
            .with_extension_type(Json2ExtensionType::default());
        let schema = Arc::new(Schema::new(vec![json_field]));

        let a_array = Arc::new(StructArray::new(
            xy_fields,
            vec![
                Arc::new(Int64Array::from_iter_values([1, 2, 3])) as ArrayRef,
                Arc::new(StringArray::from_iter_values(["x1", "x2", "x3"])) as ArrayRef,
            ],
            None,
        )) as ArrayRef;
        let b_array = Arc::new(StringArray::from_iter_values(["b1", "b2", "b3"])) as ArrayRef;
        let j_array =
            Arc::new(StructArray::new(json_fields, vec![a_array, b_array], None)) as ArrayRef;
        let columns = vec![j_array];

        let batch = RecordBatch::try_new(schema, columns).map_err(|e| e.to_string())?;

        // Persist the complete nested schema to an in-memory Parquet file so the projection is
        // exercised through parquet-rs rather than a mock.

        let object_store = ObjectStore::new(Memory::default()).map_err(|e| e.to_string())?;
        let file_handle = sst_file_handle(0, 1);
        let file_path = file_handle.file_path("test_table", PathType::Bare);

        let mut parquet_bytes = Vec::new();
        ArrowWriter::try_new(&mut parquet_bytes, batch.schema(), None)
            .and_then(|mut w| {
                w.write(&batch)?;
                Ok(w)
            })
            .and_then(|w| w.close())
            .map_err(|e| e.to_string())?;
        let file_size = parquet_bytes.len() as u64;
        object_store
            .write(&file_path, parquet_bytes)
            .await
            .map_err(|e| e.to_string())?;

        let mut cache_metrics = MetadataCacheMetrics::default();
        let loader = MetadataLoader::new(object_store.clone(), &file_path, file_size);
        let parquet_meta = loader.load(&mut cache_metrics).await?;
        let parquet_schema = parquet_meta.file_metadata().schema_descr();
        assert_eq!(3, parquet_schema.num_columns());

        // Ask Parquet to read only the deepest requested JSON2 path. This should select the single
        // leaf j.a.x and avoid both sibling leaves j.a.y and j.b.

        let projection =
            ParquetReadColumns::from_deduped(vec![ParquetReadColumn::new(0).with_nested_paths(
                vec![vec!["j".to_string(), "a".to_string(), "x".to_string()]],
            )]);
        let projection_plan =
            build_projection_plan(&projection, parquet_schema, batch.schema_ref()).unwrap();
        assert_eq!(vec![true], projection_plan.projected_root_presence);
        assert_eq!(
            projection_plan.mask,
            ProjectionMask::leaves(parquet_schema, vec![0])
        );

        // Read through the low-level stream directly.

        let arrow_metadata =
            ArrowReaderMetadata::try_new(Arc::new(parquet_meta), ArrowReaderOptions::new())
                .map_err(|e| e.to_string())?;
        let fetcher = SstParquetRangeFetcher::new(
            file_handle.file_id(),
            file_path.clone(),
            object_store,
            CacheStrategy::Disabled,
            0,
            None,
        );
        let mut stream = build_sst_parquet_record_batch_stream(
            arrow_metadata,
            0,
            None,
            projection_plan.mask,
            fetcher,
            file_path,
            1024,
        )?;

        let Some(batch) = stream.next().await.transpose()? else {
            unreachable!()
        };
        let expected = r#"
+-------------+
| j           |
+-------------+
| {a: {x: 1}} |
| {a: {x: 2}} |
| {a: {x: 3}} |
+-------------+
"#;
        assert_eq!(batch.pretty_print(), expected.trim());
        Ok(())
    }

    #[tokio::test(flavor = "current_thread")]
    async fn test_minmax_predicate_key_not_built_when_index_result_cache_disabled() {
        #[derive(Eq, PartialEq, Hash)]
        struct PanicDebugUdf;

        impl Debug for PanicDebugUdf {
            fn fmt(&self, _f: &mut Formatter<'_>) -> std::fmt::Result {
                panic!("minmax predicate key should not format exprs when cache is disabled");
            }
        }

        impl ScalarUDFImpl for PanicDebugUdf {
            fn name(&self) -> &str {
                "panic_debug_udf"
            }

            fn signature(&self) -> &Signature {
                static SIGNATURE: LazyLock<Signature> =
                    LazyLock::new(|| Signature::variadic_any(Volatility::Immutable));
                &SIGNATURE
            }

            fn return_type(&self, _arg_types: &[DataType]) -> datafusion_common::Result<DataType> {
                Ok(DataType::Int64)
            }

            fn invoke_with_args(
                &self,
                _args: ScalarFunctionArgs,
            ) -> datafusion_common::Result<ColumnarValue> {
                Ok(ColumnarValue::Scalar(ScalarValue::Int64(Some(1))))
            }
        }

        let object_store = ObjectStore::new(Memory::default()).unwrap();
        let file_handle = sst_file_handle(0, 1);
        let table_dir = "test_table".to_string();
        let path_type = PathType::Bare;
        let file_path = file_handle.file_path(&table_dir, path_type);

        let col = Arc::new(Int64Array::from_iter_values([1, 2, 3])) as ArrayRef;
        let batch = RecordBatch::try_from_iter([("col", col)]).unwrap();
        let mut parquet_bytes = Vec::new();
        let mut writer = ArrowWriter::try_new(&mut parquet_bytes, batch.schema(), None).unwrap();
        writer.write(&batch).unwrap();
        writer.close().unwrap();
        let file_size = parquet_bytes.len() as u64;
        object_store.write(&file_path, parquet_bytes).await.unwrap();

        let region_metadata: RegionMetadataRef = Arc::new(sst_region_metadata());
        let read_format = FlatReadFormat::new(
            region_metadata.clone(),
            ReadColumns::new(
                region_metadata
                    .column_metadatas
                    .iter()
                    .map(|column| column.column_id),
            ),
            None,
            &file_path,
            false,
        )
        .unwrap();

        let mut cache_metrics = MetadataCacheMetrics::default();
        let loader = MetadataLoader::new(object_store.clone(), &file_path, file_size);
        let parquet_meta = loader.load(&mut cache_metrics).await.unwrap();

        let udf = Arc::new(ScalarUDF::new_from_impl(PanicDebugUdf));
        let predicate = Predicate::new(vec![Expr::ScalarFunction(ScalarFunction::new_udf(
            udf,
            vec![],
        ))]);
        let builder = ParquetReaderBuilder::new(table_dir, path_type, file_handle, object_store)
            .predicate(Some(predicate))
            .cache(CacheStrategy::Disabled);

        let row_group_size = parquet_meta.row_group(0).num_rows() as usize;
        let total_row_count = parquet_meta.file_metadata().num_rows() as usize;
        let mut metrics = ReaderFilterMetrics::default();
        let selection = builder.row_groups_by_minmax(
            &read_format,
            &parquet_meta,
            row_group_size,
            total_row_count,
            &mut metrics,
            false,
        );

        assert!(!selection.is_empty());
    }

    #[test]
    fn test_expr_is_immutable_checks_scalar_function_volatility() {
        #[derive(Debug, PartialEq, Eq, Hash)]
        struct TestVolatilityUdf {
            name: String,
            signature: Signature,
        }

        impl TestVolatilityUdf {
            fn new(name: &str, volatility: Volatility) -> Self {
                Self {
                    name: name.to_string(),
                    signature: Signature::variadic_any(volatility),
                }
            }
        }

        impl ScalarUDFImpl for TestVolatilityUdf {
            fn name(&self) -> &str {
                &self.name
            }

            fn signature(&self) -> &Signature {
                &self.signature
            }

            fn return_type(&self, _arg_types: &[DataType]) -> datafusion_common::Result<DataType> {
                Ok(DataType::Int64)
            }

            fn invoke_with_args(
                &self,
                _args: ScalarFunctionArgs,
            ) -> datafusion_common::Result<ColumnarValue> {
                Ok(ColumnarValue::Scalar(ScalarValue::Int64(Some(1))))
            }
        }

        let expr = |name: &str, volatility| {
            Expr::ScalarFunction(ScalarFunction::new_udf(
                Arc::new(ScalarUDF::new_from_impl(TestVolatilityUdf::new(
                    name, volatility,
                ))),
                vec![],
            ))
        };

        assert!(expr_is_immutable(&expr(
            "immutable_udf",
            Volatility::Immutable
        )));
        assert!(!expr_is_immutable(&expr("stable_udf", Volatility::Stable)));
        assert!(!expr_is_immutable(&expr(
            "volatile_udf",
            Volatility::Volatile
        )));

        let scalar_variable = Expr::ScalarVariable(
            Arc::new(Field::new("@@version", DataType::Utf8, false)),
            vec!["@@version".to_string()],
        );
        assert!(!expr_is_immutable(&scalar_variable));
    }

    #[tokio::test(flavor = "current_thread")]
    async fn test_has_row_level_selection() {
        let object_store = ObjectStore::new(Memory::default()).unwrap();
        let file_path = "row_level_selection.parquet";

        let col = Arc::new(Int64Array::from_iter_values([1, 2, 3, 4, 5])) as ArrayRef;
        let batch = RecordBatch::try_from_iter([("col", col)]).unwrap();
        let props = WriterProperties::builder()
            .set_max_row_group_row_count(Some(3))
            .build();
        let mut parquet_bytes = Vec::new();
        let mut writer =
            ArrowWriter::try_new(&mut parquet_bytes, batch.schema(), Some(props)).unwrap();
        writer.write(&batch).unwrap();
        writer.close().unwrap();
        let file_size = parquet_bytes.len() as u64;
        object_store.write(file_path, parquet_bytes).await.unwrap();

        let mut cache_metrics = MetadataCacheMetrics::default();
        let loader = MetadataLoader::new(object_store, file_path, file_size);
        let parquet_meta = loader.load(&mut cache_metrics).await.unwrap();
        assert_eq!(2, parquet_meta.num_row_groups());

        let full_row_groups = RowGroupSelection::from_full_row_group_ids([0, 1], 3, 5);
        assert!(!has_row_level_selection(&full_row_groups, &parquet_meta));

        let prefix_selection = RowGroupSelection::from_row_ranges(vec![(0, vec![0..1, 1..2])], 3);
        assert!(has_row_level_selection(&prefix_selection, &parquet_meta));

        let interior_selection = RowGroupSelection::from_row_ranges(vec![(0, vec![1..2, 2..3])], 3);
        assert!(has_row_level_selection(&interior_selection, &parquet_meta));
    }

    fn expected_metadata_with_reused_tag_name(
        old_metadata: &RegionMetadata,
    ) -> Arc<RegionMetadata> {
        let mut builder = RegionMetadataBuilder::new(old_metadata.region_id);
        builder
            .push_column_metadata(ColumnMetadata {
                column_schema: ColumnSchema::new(
                    "tag_0".to_string(),
                    ConcreteDataType::string_datatype(),
                    true,
                ),
                semantic_type: SemanticType::Tag,
                column_id: 10,
            })
            .push_column_metadata(ColumnMetadata {
                column_schema: ColumnSchema::new(
                    "tag_1".to_string(),
                    ConcreteDataType::string_datatype(),
                    true,
                ),
                semantic_type: SemanticType::Tag,
                column_id: 1,
            })
            .push_column_metadata(ColumnMetadata {
                column_schema: ColumnSchema::new(
                    "field_0".to_string(),
                    ConcreteDataType::uint64_datatype(),
                    true,
                ),
                semantic_type: SemanticType::Field,
                column_id: 2,
            })
            .push_column_metadata(ColumnMetadata {
                column_schema: ColumnSchema::new(
                    "ts".to_string(),
                    ConcreteDataType::timestamp_millisecond_datatype(),
                    false,
                ),
                semantic_type: SemanticType::Timestamp,
                column_id: 3,
            })
            .primary_key(vec![10, 1]);

        Arc::new(builder.build().unwrap())
    }

    #[test]
    fn test_simple_filter_context_uses_default_value_for_mismatched_expected_metadata() {
        let metadata: RegionMetadataRef = Arc::new(sst_region_metadata());
        let expected_metadata = expected_metadata_with_reused_tag_name(metadata.as_ref());
        let ctx = SimpleFilterContext::new_opt(
            &metadata,
            Some(expected_metadata.as_ref()),
            &col("tag_0").eq(lit("a")),
        )
        .unwrap();
        assert!(matches!(
            ctx.filter(),
            MaybeFilter::Matched | MaybeFilter::Pruned
        ));
    }

    #[test]
    fn test_simple_filter_context_drops_mismatched_field_filter() {
        let (sst_metadata, latest_metadata) = mock_metadata();
        let ctx = SimpleFilterContext::new_opt(
            &sst_metadata,
            Some(latest_metadata.as_ref()),
            &col("field_0").eq(lit(1_i64)),
        );

        assert!(ctx.is_none());
    }

    fn mock_metadata() -> (RegionMetadataRef, RegionMetadataRef) {
        let region_id = RegionId::new(1, 1);
        let make_tag_0 = || ColumnMetadata {
            column_schema: ColumnSchema::new(
                "tag_0".to_string(),
                ConcreteDataType::string_datatype(),
                true,
            ),
            semantic_type: SemanticType::Tag,
            column_id: 0,
        };
        let make_ts = || ColumnMetadata {
            column_schema: ColumnSchema::new(
                "ts".to_string(),
                ConcreteDataType::timestamp_millisecond_datatype(),
                false,
            ),
            semantic_type: SemanticType::Timestamp,
            column_id: 2,
        };
        let make_field_0 = |data_type| ColumnMetadata {
            column_schema: ColumnSchema::new("field_0".to_string(), data_type, true),
            semantic_type: SemanticType::Field,
            column_id: 1,
        };

        let mut sst_builder = RegionMetadataBuilder::new(region_id);
        sst_builder
            .push_column_metadata(make_tag_0())
            .push_column_metadata(make_field_0(ConcreteDataType::uint64_datatype()))
            .push_column_metadata(make_ts())
            .primary_key(vec![0]);
        let sst_metadata = Arc::new(sst_builder.build().unwrap());

        let mut expected_builder = RegionMetadataBuilder::new(region_id);
        expected_builder
            .push_column_metadata(make_tag_0())
            .push_column_metadata(make_field_0(ConcreteDataType::int64_datatype()))
            .push_column_metadata(make_ts())
            .primary_key(vec![0]);

        let expected_metadata = Arc::new(expected_builder.build().unwrap());

        (sst_metadata, expected_metadata)
    }

    #[test]
    fn test_physical_filter_context_skips_renamed_column() {
        let metadata: RegionMetadataRef = Arc::new(sst_region_metadata());
        let expected_metadata = expected_metadata_with_reused_tag_name(metadata.as_ref());
        let read_format = FlatReadFormat::new(
            metadata.clone(),
            ReadColumns::new(metadata.column_metadatas.iter().map(|c| c.column_id)),
            None,
            "test",
            true,
        )
        .unwrap();

        let ctx = PhysicalFilterContext::new_opt(
            &metadata,
            Some(expected_metadata.as_ref()),
            &read_format,
            &col("tag_0").in_list(vec![lit("a"), lit("b")], false),
        );

        assert!(ctx.is_none());
    }

    #[test]
    fn test_physical_filter_context_only_accepts_prefilter_candidates() {
        let metadata: RegionMetadataRef = Arc::new(sst_region_metadata());
        let read_format = FlatReadFormat::new(
            metadata.clone(),
            ReadColumns::new(metadata.column_metadatas.iter().map(|c| c.column_id)),
            None,
            "test",
            true,
        )
        .unwrap();

        // InList is on the allowlist — should build a context.
        let in_list = col("tag_0").in_list(vec![lit("a"), lit("b")], false);
        assert!(PhysicalFilterContext::new_opt(&metadata, None, &read_format, &in_list).is_some());

        // NOT IN uses the same variant with `negated: true` — also accepted.
        let not_in = col("tag_0").in_list(vec![lit("a"), lit("b")], true);
        assert!(PhysicalFilterContext::new_opt(&metadata, None, &read_format, &not_in).is_some());

        // IS NULL / IS NOT NULL are accepted.
        let is_null = col("tag_0").is_null();
        assert!(PhysicalFilterContext::new_opt(&metadata, None, &read_format, &is_null).is_some());
        let is_not_null = col("tag_0").is_not_null();
        assert!(
            PhysicalFilterContext::new_opt(&metadata, None, &read_format, &is_not_null).is_some()
        );

        // BETWEEN is accepted.
        let between = col("field_0").between(lit(1_u64), lit(10_u64));
        assert!(PhysicalFilterContext::new_opt(&metadata, None, &read_format, &between).is_some());

        // Binary expr is handled by SimpleFilterEvaluator — rejected here.
        let binary = col("tag_0").eq(lit("a"));
        assert!(PhysicalFilterContext::new_opt(&metadata, None, &read_format, &binary).is_none());
    }

    fn write_test_parquet_with_pk_column(values: &[&[u8]]) -> bytes::Bytes {
        use datatypes::arrow::array::{
            BinaryArray, TimestampMillisecondArray, UInt8Array, UInt64Array,
        };
        use datatypes::arrow::datatypes::{Field as ArrowField, Schema as ArrowSchema, TimeUnit};
        use store_api::storage::consts::{
            OP_TYPE_COLUMN_NAME, PRIMARY_KEY_COLUMN_NAME, SEQUENCE_COLUMN_NAME,
        };

        let n = values.len();
        let schema = Arc::new(ArrowSchema::new(vec![
            ArrowField::new(
                "ts",
                DataType::Timestamp(TimeUnit::Millisecond, None),
                false,
            ),
            ArrowField::new(PRIMARY_KEY_COLUMN_NAME, DataType::Binary, false),
            ArrowField::new(SEQUENCE_COLUMN_NAME, DataType::UInt64, false),
            ArrowField::new(OP_TYPE_COLUMN_NAME, DataType::UInt8, false),
        ]));
        let ts: ArrayRef = Arc::new(TimestampMillisecondArray::from(vec![0_i64; n]));
        let pk: ArrayRef = Arc::new(BinaryArray::from_iter_values(values.iter().copied()));
        let seq: ArrayRef = Arc::new(UInt64Array::from(vec![0_u64; n]));
        let op: ArrayRef = Arc::new(UInt8Array::from(vec![0_u8; n]));
        let batch = RecordBatch::try_new(schema.clone(), vec![ts, pk, seq, op]).unwrap();

        let mut bytes = Vec::new();
        let mut writer = ArrowWriter::try_new(&mut bytes, schema, None).unwrap();
        writer.write(&batch).unwrap();
        writer.close().unwrap();
        bytes::Bytes::from(bytes)
    }

    fn load_parquet_meta(bytes: bytes::Bytes) -> Arc<ParquetMetaData> {
        let builder =
            parquet::arrow::arrow_reader::ParquetRecordBatchReaderBuilder::try_new(bytes).unwrap();
        builder.metadata().clone()
    }

    #[test]
    fn test_should_read_pk_as_binary_small_chunk_returns_false() {
        let bytes = write_test_parquet_with_pk_column(&[b"a", b"b", b"c"]);
        let meta = load_parquet_meta(bytes);

        assert!(!should_read_pk_as_binary_with_limit(&meta, 1024));
    }

    #[test]
    fn test_should_read_pk_as_binary_large_chunk_returns_true() {
        let owned: Vec<Vec<u8>> = (0..512u32)
            .map(|i| {
                let mut v = vec![0u8; 16];
                v[..4].copy_from_slice(&i.to_le_bytes());
                v
            })
            .collect();
        let refs: Vec<&[u8]> = owned.iter().map(|v| v.as_slice()).collect();
        let bytes = write_test_parquet_with_pk_column(&refs);
        let meta = load_parquet_meta(bytes);

        assert!(should_read_pk_as_binary_with_limit(&meta, 1024));
    }
}