be/src/format_v2/parquet/parquet_scan.cpp

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

#include "format_v2/parquet/parquet_scan.h"

#include <algorithm>
#include <iterator>
#include <limits>
#include <memory>
#include <optional>
#include <ranges>
#include <set>
#include <span>
#include <unordered_set>
#include <utility>

#include "common/exception.h"
#include "common/status.h"
#include "core/assert_cast.h"
#include "core/block/block.h"
#include "core/column/column_decimal.h"
#include "core/column/column_vector.h"
#include "exprs/expr_zonemap_filter.h"
#include "exprs/vcompound_pred.h"
#include "exprs/vectorized_fn_call.h"
#include "exprs/vexpr_context.h"
#include "format_v2/parquet/parquet_column_schema.h"
#include "format_v2/parquet/parquet_file_context.h"
#include "format_v2/parquet/parquet_statistics.h"
#include "format_v2/parquet/reader/global_rowid_column_reader.h"
#include "format_v2/parquet/reader/native/column_chunk_reader.h"
#include "format_v2/parquet/reader/native_column_reader.h"
#include "format_v2/parquet/reader/row_position_column_reader.h"
#include "util/defer_op.h"
#include "util/time.h"

namespace doris::format::parquet {

namespace detail {

std::vector<size_t> order_adaptive_predicates(
        const std::vector<size_t>& positions,
        const std::unordered_map<size_t, AdaptivePredicateStats>& stats) {
    if (std::ranges::any_of(positions, [&](size_t position) {
            const auto it = stats.find(position);
            return it == stats.end() || it->second.samples == 0;
        })) {
        return positions;
    }
    auto ordered = positions;
    std::stable_sort(ordered.begin(), ordered.end(), [&](size_t left, size_t right) {
        const auto score = [&](size_t position) {
            const auto& sample = stats.at(position);
            return sample.cost_per_input_row_ns / std::max(1.0 - sample.survival_ratio, 0.01);
        };
        return score(left) < score(right);
    });
    return ordered;
}

std::vector<size_t> adaptive_prefetch_prefix(
        const std::vector<size_t>& ordered_positions,
        const std::unordered_map<size_t, AdaptivePredicateStats>& stats,
        double minimum_reach_probability) {
    if (std::ranges::any_of(ordered_positions, [&](size_t position) {
            const auto it = stats.find(position);
            return it == stats.end() || it->second.samples == 0;
        })) {
        return ordered_positions;
    }
    std::vector<size_t> result;
    double reach_probability = 1;
    for (const size_t position : ordered_positions) {
        if (!result.empty() && reach_probability < minimum_reach_probability) {
            break;
        }
        result.push_back(position);
        reach_probability *= stats.at(position).survival_ratio;
    }
    return result;
}

bool should_sample_adaptive_predicate(size_t samples, size_t batch_sequence) {
    constexpr size_t WARMUP_SAMPLES = 8;
    constexpr size_t STEADY_STATE_INTERVAL = 16;
    return samples < WARMUP_SAMPLES || batch_sequence % STEADY_STATE_INTERVAL == 0;
}

} // namespace detail

namespace {

detail::PredicateConjunctSchedule build_predicate_conjunct_schedule(
        const format::FileScanRequest& request);

bool is_dictionary_data_encoding(tparquet::Encoding::type encoding) {
    return encoding == tparquet::Encoding::PLAIN_DICTIONARY ||
           encoding == tparquet::Encoding::RLE_DICTIONARY;
}

bool is_level_encoding(tparquet::Encoding::type encoding) {
    return encoding == tparquet::Encoding::RLE || encoding == tparquet::Encoding::BIT_PACKED;
}

bool is_data_page_type(tparquet::PageType::type page_type) {
    return page_type == tparquet::PageType::DATA_PAGE ||
           page_type == tparquet::PageType::DATA_PAGE_V2;
}

bool is_fully_dictionary_encoded_chunk(const tparquet::ColumnMetaData& column_metadata) {
    if (!column_metadata.__isset.dictionary_page_offset ||
        column_metadata.dictionary_page_offset < 0) {
        return false;
    }

    const auto& encoding_stats = column_metadata.encoding_stats;
    if (!encoding_stats.empty()) {
        bool has_dictionary_data_page = false;
        for (const auto& encoding_stat : encoding_stats) {
            if (!is_data_page_type(encoding_stat.page_type) || encoding_stat.count <= 0) {
                continue;
            }
            if (!is_dictionary_data_encoding(encoding_stat.encoding)) {
                return false;
            }
            has_dictionary_data_page = true;
        }
        return has_dictionary_data_page;
    }

    bool has_dictionary_encoding = false;
    for (const auto encoding : column_metadata.encodings) {
        if (is_dictionary_data_encoding(encoding)) {
            has_dictionary_encoding = true;
            continue;
        }
        if (!is_level_encoding(encoding)) {
            return false;
        }
    }
    return has_dictionary_encoding;
}

bool supports_row_level_dictionary_filter(const ParquetColumnSchema& column_schema,
                                          const tparquet::ColumnMetaData& column_metadata) {
    if (column_schema.kind != ParquetColumnSchemaKind::PRIMITIVE || column_schema.type == nullptr ||
        column_schema.max_repetition_level > 0) {
        return false;
    }
    bool is_supported_physical_type = false;
    switch (column_metadata.type) {
    case tparquet::Type::BYTE_ARRAY:
        is_supported_physical_type = column_schema.type_descriptor.is_string_like;
        break;
    case tparquet::Type::INT32:
    case tparquet::Type::INT64:
    case tparquet::Type::INT96:
    case tparquet::Type::FLOAT:
    case tparquet::Type::DOUBLE:
    case tparquet::Type::FIXED_LEN_BYTE_ARRAY:
        is_supported_physical_type = true;
        break;
    case tparquet::Type::BOOLEAN:
        // Parquet booleans are PLAIN encoded and cannot have a dictionary page.
        break;
    }
    if (!is_supported_physical_type) {
        return false;
    }
    if (remove_nullable(column_schema.type)->get_primitive_type() == TYPE_VARBINARY) {
        // A table STRING predicate can be rewritten through a raw VARBINARY file slot. Evaluating
        // it on dictionary Fields before the mapping expression is neither type-safe nor exact.
        return false;
    }
    // The row filter consumes dictionary ids rather than decoded values, so a plain data page
    // cannot resume this reader without changing its output domain. Keep mixed chunks on the
    // normal decoded-value path to preserve one representation for the complete column chunk.
    return is_fully_dictionary_encoded_chunk(column_metadata);
}

void collect_all_leaf_column_ids(const ParquetColumnSchema& column_schema,
                                 std::unordered_set<int>* leaf_column_ids) {
    DORIS_CHECK(leaf_column_ids != nullptr);
    if (column_schema.kind == ParquetColumnSchemaKind::PRIMITIVE) {
        if (column_schema.leaf_column_id >= 0) {
            leaf_column_ids->insert(column_schema.leaf_column_id);
        }
        return;
    }
    for (const auto& child : column_schema.children) {
        DORIS_CHECK(child != nullptr);
        collect_all_leaf_column_ids(*child, leaf_column_ids);
    }
}

void collect_projected_leaf_column_ids(const ParquetColumnSchema& column_schema,
                                       const format::LocalColumnIndex& projection,
                                       std::unordered_set<int>* leaf_column_ids) {
    DORIS_CHECK(leaf_column_ids != nullptr);
    if (projection.project_all_children || projection.children.empty()) {
        collect_all_leaf_column_ids(column_schema, leaf_column_ids);
        return;
    }
    for (const auto& child_projection : projection.children) {
        const auto child_it =
                std::ranges::find_if(column_schema.children, [&](const auto& child_schema) {
                    return child_schema->local_id == child_projection.local_id();
                });
        DORIS_CHECK(child_it != column_schema.children.end());
        collect_projected_leaf_column_ids(**child_it, child_projection, leaf_column_ids);
    }
}

std::vector<format::LocalColumnIndex> request_scan_columns(const format::FileScanRequest& request) {
    std::vector<format::LocalColumnIndex> scan_columns;
    scan_columns.reserve(request.predicate_columns.size() + request.non_predicate_columns.size());
    scan_columns.insert(scan_columns.end(), request.predicate_columns.begin(),
                        request.predicate_columns.end());
    for (const auto& column : request.non_predicate_columns) {
        if (!request.is_count_star_placeholder(column.column_id())) {
            scan_columns.push_back(column);
        }
    }
    return scan_columns;
}

std::vector<format::LocalColumnIndex> physical_non_predicate_columns(
        const format::FileScanRequest& request) {
    std::vector<format::LocalColumnIndex> columns;
    columns.reserve(request.non_predicate_columns.size());
    for (const auto& column : request.non_predicate_columns) {
        if (!request.is_count_star_placeholder(column.column_id())) {
            columns.push_back(column);
        }
    }
    return columns;
}

void materialize_count_star_placeholders(const format::FileScanRequest& request, size_t rows,
                                         Block* file_block) {
    DORIS_CHECK(file_block != nullptr);
    for (const auto& column : request.non_predicate_columns) {
        if (!request.is_count_star_placeholder(column.column_id())) {
            continue;
        }
        const auto block_position = request.local_positions.at(column.column_id()).value();
        auto placeholder = file_block->get_by_position(block_position).column->assert_mutable();
        DCHECK(placeholder->empty());
        placeholder->insert_many_defaults(rows);
        file_block->replace_by_position(block_position, std::move(placeholder));
    }
}

} // namespace

namespace detail {

Status build_native_prefetch_ranges(
        const tparquet::FileMetaData& metadata,
        const std::vector<std::unique_ptr<ParquetColumnSchema>>& file_schema,
        const std::vector<format::LocalColumnIndex>& scan_columns, int row_group_idx,
        size_t file_size, bool parquet_816_padding, std::vector<ParquetPageCacheRange>* ranges) {
    DORIS_CHECK(ranges != nullptr);
    ranges->clear();
    std::unordered_set<int> leaf_column_ids;
    for (const auto& projection : scan_columns) {
        const auto local_id = projection.local_id();
        if (local_id == format::ROW_POSITION_COLUMN_ID ||
            local_id == format::GLOBAL_ROWID_COLUMN_ID) {
            continue;
        }
        if (local_id < 0 || local_id >= static_cast<int32_t>(file_schema.size()) ||
            file_schema[local_id] == nullptr) {
            return Status::Corruption("Invalid Parquet projected column id {}", local_id);
        }
        // Prefetch and merge-reader ranges must be physical leaf chunks, not Doris logical slots.
        // Example: for a struct column s<a:int,b:string>, projecting only s.a should include only
        // the Parquet leaf chunk of a. Projecting the whole struct includes both a and b.
        collect_projected_leaf_column_ids(*file_schema[local_id], projection, &leaf_column_ids);
    }

    if (row_group_idx < 0 || row_group_idx >= static_cast<int>(metadata.row_groups.size())) {
        return Status::Corruption("Invalid Parquet row group index {}", row_group_idx);
    }
    const auto& row_group_metadata = metadata.row_groups[row_group_idx];
    std::vector<int> ordered_leaf_column_ids(leaf_column_ids.begin(), leaf_column_ids.end());
    std::ranges::sort(ordered_leaf_column_ids);

    ranges->reserve(ordered_leaf_column_ids.size());
    for (const auto leaf_column_id : ordered_leaf_column_ids) {
        if (leaf_column_id < 0 ||
            leaf_column_id >= static_cast<int>(row_group_metadata.columns.size())) {
            return Status::Corruption("Invalid Parquet leaf column id {}", leaf_column_id);
        }
        const auto& chunk = row_group_metadata.columns[leaf_column_id];
        if (!chunk.__isset.meta_data) {
            return Status::Corruption("Parquet leaf column {} has no chunk metadata",
                                      leaf_column_id);
        }
        native::ColumnChunkRange chunk_range;
        RETURN_IF_ERROR(native::compute_column_chunk_range(chunk.meta_data, file_size,
                                                           parquet_816_padding, &chunk_range));
        if (chunk_range.length > 0) {
            if (chunk_range.offset > static_cast<size_t>(std::numeric_limits<int64_t>::max()) ||
                chunk_range.length > static_cast<size_t>(std::numeric_limits<int64_t>::max())) {
                return Status::Corruption("Parquet column chunk range exceeds int64 coordinates");
            }
            // Prefetch must use the same checked chunk extent as the decoder, including the
            // PARQUET-816 compatibility padding, so warm-up cannot target different bytes.
            ranges->push_back(
                    ParquetPageCacheRange {.offset = static_cast<int64_t>(chunk_range.offset),
                                           .size = static_cast<int64_t>(chunk_range.length)});
        }
    }
    return Status::OK();
}

} // namespace detail

namespace detail {

Status select_native_row_groups_by_scan_range(const tparquet::FileMetaData& metadata,
                                              const ParquetScanRange& scan_range,
                                              std::vector<int64_t>* row_group_first_rows,
                                              std::vector<int>* selected_row_groups) {
    DORIS_CHECK(row_group_first_rows != nullptr && selected_row_groups != nullptr);
    if (scan_range.start_offset < 0 || scan_range.size < -1 ||
        (scan_range.size >= 0 &&
         scan_range.start_offset > std::numeric_limits<int64_t>::max() - scan_range.size)) {
        return Status::Corruption("Invalid Parquet scan range [{}, {})", scan_range.start_offset,
                                  scan_range.size);
    }
    const uint64_t range_start = static_cast<uint64_t>(scan_range.start_offset);
    const uint64_t range_end = scan_range.size < 0
                                       ? std::numeric_limits<uint64_t>::max()
                                       : range_start + static_cast<uint64_t>(scan_range.size);
    const size_t file_size = scan_range.file_size < 0 ? std::numeric_limits<size_t>::max()
                                                      : static_cast<size_t>(scan_range.file_size);
    const bool full_file_range =
            scan_range.size < 0 || (range_start == 0 && scan_range.file_size >= 0 &&
                                    range_end >= static_cast<uint64_t>(scan_range.file_size));
    const auto compat = native::parquet_reader_compat(
            metadata.__isset.created_by ? metadata.created_by : std::string {});
    row_group_first_rows->assign(metadata.row_groups.size(), 0);
    selected_row_groups->clear();
    selected_row_groups->reserve(metadata.row_groups.size());
    int64_t next_first_row = 0;
    for (size_t row_group_idx = 0; row_group_idx < metadata.row_groups.size(); ++row_group_idx) {
        (*row_group_first_rows)[row_group_idx] = next_first_row;
        const auto& row_group = metadata.row_groups[row_group_idx];
        if (row_group.num_rows < 0) {
            return Status::Corruption("Invalid negative row count in parquet row group {}",
                                      row_group_idx);
        }
        if (row_group.num_rows > std::numeric_limits<int64_t>::max() - next_first_row) {
            return Status::Corruption("Parquet row counts overflow at row group {}", row_group_idx);
        }
        next_first_row += row_group.num_rows;
        bool selected = full_file_range;
        if (!full_file_range) {
            if (row_group.columns.empty()) {
                return Status::Corruption("Parquet row group {} has no column chunks",
                                          row_group_idx);
            }
            size_t group_start = std::numeric_limits<size_t>::max();
            size_t group_end = 0;
            for (size_t column_idx = 0; column_idx < row_group.columns.size(); ++column_idx) {
                const auto& chunk = row_group.columns[column_idx];
                if (!chunk.__isset.meta_data) {
                    return Status::Corruption("Parquet row group {} column {} has no metadata",
                                              row_group_idx, column_idx);
                }
                native::ColumnChunkRange chunk_range;
                RETURN_IF_ERROR(native::compute_column_chunk_range(
                        chunk.meta_data, file_size, compat.parquet_816_padding, &chunk_range));
                group_start = std::min(group_start, chunk_range.offset);
                group_end = std::max(group_end, chunk_range.offset + chunk_range.length);
            }
            // Checked chunk ranges make end >= start; this midpoint form cannot overflow even
            // when footer offsets are close to the host coordinate limit.
            const uint64_t group_mid =
                    static_cast<uint64_t>(group_start) + (group_end - group_start) / 2;
            selected = group_mid >= range_start && group_mid < range_end;
        }
        if (selected) {
            selected_row_groups->push_back(cast_set<int>(row_group_idx));
        }
    }
    return Status::OK();
}

} // namespace detail

namespace {

std::vector<RowRange> intersect_row_ranges(const std::vector<RowRange>& left,
                                           const std::vector<RowRange>& right) {
    std::vector<RowRange> result;
    size_t left_idx = 0;
    size_t right_idx = 0;
    while (left_idx < left.size() && right_idx < right.size()) {
        const int64_t left_end = left[left_idx].start + left[left_idx].length;
        const int64_t right_end = right[right_idx].start + right[right_idx].length;
        const int64_t start = std::max(left[left_idx].start, right[right_idx].start);
        const int64_t end = std::min(left_end, right_end);
        if (start < end) {
            result.push_back({.start = start, .length = end - start});
        }
        if (left_end < right_end) {
            ++left_idx;
        } else {
            ++right_idx;
        }
    }
    return result;
}

Status finalize_native_row_group_read_plan(
        const NativeParquetMetadata& metadata,
        const std::vector<std::unique_ptr<ParquetColumnSchema>>& file_schema,
        const format::FileScanRequest& request, bool enable_bloom_filter,
        RowGroupReadPlan* row_group_plan, ParquetPruningStats* pruning_stats,
        const cctz::time_zone* timezone, const RuntimeState* runtime_state,
        ParquetFileContext* file_context, const ParquetColumnReaderProfile& column_reader_profile,
        bool* selected) {
    DORIS_CHECK(row_group_plan != nullptr && pruning_stats != nullptr && file_context != nullptr &&
                selected != nullptr);
    *selected = true;
    if (!row_group_plan->expensive_pruning_pending) {
        return Status::OK();
    }
    row_group_plan->expensive_pruning_pending = false;
    const auto& thrift = metadata.to_thrift();
    const std::vector<int> candidate {row_group_plan->row_group_id};
    std::vector<int> metadata_selected;
    RETURN_IF_ERROR(select_row_groups_by_metadata(
            thrift, file_schema, request, &candidate, &metadata_selected, enable_bloom_filter,
            pruning_stats, timezone, runtime_state, file_context, column_reader_profile,
            ParquetMetadataProbeMode::EXPENSIVE_ONLY));
    if (metadata_selected.empty()) {
        *selected = false;
        return Status::OK();
    }

    std::unordered_set<int> requested_leaf_ids;
    for (const auto& projection : request_scan_columns(request)) {
        const auto local_id = projection.local_id();
        if (local_id < 0 || local_id >= static_cast<int32_t>(file_schema.size())) {
            continue;
        }
        collect_projected_leaf_column_ids(*file_schema[local_id], projection, &requested_leaf_ids);
    }
    std::unordered_map<int, NativeParquetPageIndex> page_indexes;
    if (can_use_parquet_page_index(request, runtime_state)) {
        RETURN_IF_ERROR(file_context->load_native_page_indexes(
                row_group_plan->row_group_id, requested_leaf_ids, &page_indexes,
                &pruning_stats->read_page_index_time, &pruning_stats->parse_page_index_time));
    }
    std::vector<RowRange> page_selected_ranges;
    std::map<int, ParquetPageSkipPlan> page_skip_plans;
    RETURN_IF_ERROR(select_row_group_ranges_by_native_page_index(
            thrift, page_indexes, file_schema, request, row_group_plan->row_group_rows,
            &page_selected_ranges, &page_skip_plans, pruning_stats, timezone, runtime_state));
    row_group_plan->selected_ranges =
            intersect_row_ranges(row_group_plan->selected_ranges, page_selected_ranges);
    row_group_plan->page_skip_plans = std::move(page_skip_plans);
    for (auto& [leaf_column_id, indexes] : page_indexes) {
        row_group_plan->offset_indexes.emplace(leaf_column_id, std::move(indexes.offset_index));
    }
    if (row_group_plan->selected_ranges.empty()) {
        *selected = false;
        return Status::OK();
    }
    pruning_stats->selected_row_ranges += row_group_plan->selected_ranges.size();
    return Status::OK();
}

Status build_native_row_group_read_plans(
        const NativeParquetMetadata& metadata,
        const std::vector<std::unique_ptr<ParquetColumnSchema>>& file_schema,
        const format::FileScanRequest& request, const std::vector<int>& selected_row_groups,
        const std::vector<int64_t>& row_group_first_rows, RowGroupScanPlan* plan,
        const cctz::time_zone* timezone, const RuntimeState* runtime_state,
        ParquetFileContext* file_context) {
    DORIS_CHECK(plan != nullptr && file_context != nullptr);
    const auto& thrift = metadata.to_thrift();
    plan->row_groups.reserve(selected_row_groups.size());
    for (const int row_group_idx : selected_row_groups) {
        const auto& row_group = thrift.row_groups[row_group_idx];
        if (row_group.num_rows == 0) {
            continue;
        }
        RowGroupReadPlan row_group_plan;
        row_group_plan.row_group_id = row_group_idx;
        row_group_plan.first_file_row = row_group_first_rows[row_group_idx];
        row_group_plan.row_group_rows = row_group.num_rows;
        row_group_plan.selected_ranges = {{.start = 0, .length = row_group.num_rows}};
        row_group_plan.expensive_pruning_pending = true;
        plan->row_groups.push_back(std::move(row_group_plan));
    }
    return Status::OK();
}

} // namespace

Status plan_parquet_row_groups(const NativeParquetMetadata& metadata,
                               const std::vector<std::unique_ptr<ParquetColumnSchema>>& file_schema,
                               const format::FileScanRequest& request,
                               const ParquetScanRange& scan_range, bool enable_bloom_filter,
                               RowGroupScanPlan* plan, const cctz::time_zone* timezone,
                               const RuntimeState* runtime_state, ParquetFileContext* file_context,
                               const ParquetColumnReaderProfile& column_reader_profile) {
    DORIS_CHECK(plan != nullptr && file_context != nullptr);
    plan->row_groups.clear();
    plan->pruning_stats = {};
    plan->enable_bloom_filter = enable_bloom_filter;
    std::vector<int64_t> row_group_first_rows;
    std::vector<int> scan_range_selected;
    RETURN_IF_ERROR(detail::select_native_row_groups_by_scan_range(
            metadata.to_thrift(), scan_range, &row_group_first_rows, &scan_range_selected));
    std::vector<int> metadata_selected;
    RETURN_IF_ERROR(select_row_groups_by_metadata(
            metadata.to_thrift(), file_schema, request, &scan_range_selected, &metadata_selected,
            enable_bloom_filter, &plan->pruning_stats, timezone, runtime_state, file_context,
            column_reader_profile, ParquetMetadataProbeMode::FOOTER_ONLY));
    RETURN_IF_ERROR(build_native_row_group_read_plans(metadata, file_schema, request,
                                                      metadata_selected, row_group_first_rows, plan,
                                                      timezone, runtime_state, file_context));
    plan->pruning_stats.selected_row_groups = plan->row_groups.size();
    return Status::OK();
}

Status finalize_parquet_row_group_plans(
        const NativeParquetMetadata& metadata,
        const std::vector<std::unique_ptr<ParquetColumnSchema>>& file_schema,
        const format::FileScanRequest& request, bool enable_bloom_filter, RowGroupScanPlan* plan,
        const cctz::time_zone* timezone, const RuntimeState* runtime_state,
        ParquetFileContext* file_context, const ParquetColumnReaderProfile& column_reader_profile,
        const ParquetProfile* parquet_profile) {
    DORIS_CHECK(plan != nullptr && file_context != nullptr);
    std::vector<RowGroupReadPlan> selected_plans;
    selected_plans.reserve(plan->row_groups.size());
    for (auto& row_group_plan : plan->row_groups) {
        ParquetPruningStats deferred_stats;
        bool selected = false;
        RETURN_IF_ERROR(finalize_native_row_group_read_plan(
                metadata, file_schema, request, enable_bloom_filter, &row_group_plan,
                &deferred_stats, timezone, runtime_state, file_context, column_reader_profile,
                &selected));
        if (parquet_profile != nullptr) {
            parquet_profile->update_deferred_pruning_stats(deferred_stats, selected);
        }
        if (selected) {
            selected_plans.push_back(std::move(row_group_plan));
        }
    }
    plan->row_groups = std::move(selected_plans);
    plan->pruning_stats.selected_row_groups = plan->row_groups.size();
    return Status::OK();
}

namespace {

using OwnedExpressionConjunct = std::pair<VExprContextSPtr, VExprSPtr>;
using OwnedExpressionConjuncts = std::vector<OwnedExpressionConjunct>;

void update_counter_if_not_null(RuntimeProfile::Counter* counter, int64_t value) {
    if (counter != nullptr) {
        COUNTER_UPDATE(counter, value);
    }
}

uint16_t apply_filter_to_selection(const IColumn::Filter& filter, SelectionVector* selection,
                                   uint16_t selected_rows) {
    uint16_t new_selected_rows = 0;
    for (uint16_t selection_idx = 0; selection_idx < selected_rows; ++selection_idx) {
        const auto row_idx = selection->get_index(selection_idx);
        if (filter[row_idx] != 0) {
            selection->set_index(new_selected_rows++, static_cast<SelectionVector::Index>(row_idx));
        }
    }
    return new_selected_rows;
}

Status execute_compact_filter_conjuncts(const VExprContextSPtrs& conjuncts, size_t rows,
                                        Block* file_block, IColumn::Filter* compact_filter,
                                        bool* can_filter_all) {
    DORIS_CHECK(compact_filter != nullptr);
    DORIS_CHECK(can_filter_all != nullptr);
    compact_filter->resize_fill(rows, 1);
    *can_filter_all = false;
    for (const auto& conjunct : conjuncts) {
        DORIS_CHECK(conjunct != nullptr);
        IColumn::Filter filter(rows, 1);
        bool conjunct_can_filter_all = false;
        RETURN_IF_ERROR(conjunct->execute_filter(file_block, filter.data(), rows, false,
                                                 &conjunct_can_filter_all));
        if (conjunct_can_filter_all) {
            std::ranges::fill(*compact_filter, 0);
            *can_filter_all = true;
            break;
        }
        for (size_t row = 0; row < rows; ++row) {
            (*compact_filter)[row] &= filter[row];
        }
    }
    return Status::OK();
}

Status execute_compact_owned_conjuncts(std::span<const OwnedExpressionConjunct> conjuncts,
                                       size_t rows, Block* file_block,
                                       IColumn::Filter* compact_filter, bool* can_filter_all) {
    DORIS_CHECK(compact_filter != nullptr);
    DORIS_CHECK(can_filter_all != nullptr);
    compact_filter->resize_fill(rows, 1);
    *can_filter_all = false;
    for (const auto& [owner_context, residual_expr] : conjuncts) {
        DORIS_CHECK(owner_context != nullptr);
        DORIS_CHECK(residual_expr != nullptr);
        IColumn::Filter filter(rows, 1);
        bool conjunct_can_filter_all = false;
        RETURN_IF_ERROR(residual_expr->execute_filter(owner_context.get(), file_block,
                                                      filter.data(), rows, false,
                                                      &conjunct_can_filter_all));
        if (conjunct_can_filter_all) {
            std::ranges::fill(*compact_filter, 0);
            *can_filter_all = true;
            break;
        }
        for (size_t row = 0; row < rows; ++row) {
            (*compact_filter)[row] &= filter[row];
        }
    }
    return Status::OK();
}

Status execute_compact_delete_conjuncts(const VExprContextSPtrs& delete_conjuncts, size_t rows,
                                        Block* file_block, IColumn::Filter* compact_filter,
                                        bool* can_filter_all) {
    DORIS_CHECK(compact_filter != nullptr);
    DORIS_CHECK(can_filter_all != nullptr);
    compact_filter->resize_fill(rows, 1);
    *can_filter_all = false;
    for (const auto& delete_conjunct : delete_conjuncts) {
        DORIS_CHECK(delete_conjunct != nullptr);
        const size_t original_columns = file_block->columns();
        int result_column_id = -1;
        RETURN_IF_ERROR(delete_conjunct->root()->execute(delete_conjunct.get(), file_block,
                                                         &result_column_id));
        RETURN_IF_ERROR(detail::validate_ephemeral_expr_result_column(
                original_columns, result_column_id, file_block->columns()));
        const auto& delete_filter = assert_cast<const ColumnUInt8&>(
                                            *file_block->get_by_position(result_column_id).column)
                                            .get_data();
        DORIS_CHECK(delete_filter.size() == rows);
        bool has_kept_row = false;
        for (size_t row = 0; row < rows; ++row) {
            (*compact_filter)[row] &= !delete_filter[row];
            has_kept_row |= (*compact_filter)[row] != 0;
        }
        file_block->erase(result_column_id);
        if (!has_kept_row) {
            *can_filter_all = true;
            break;
        }
    }
    return Status::OK();
}

Status execute_filter_conjuncts(const format::FileScanRequest& request, int64_t batch_rows,
                                Block* file_block, SelectionVector* selection,
                                uint16_t* selected_rows) {
    for (const auto& conjunct : request.conjuncts) {
        if (*selected_rows == 0) {
            break;
        }
        DORIS_CHECK(conjunct != nullptr);
        IColumn::Filter filter(static_cast<size_t>(batch_rows), 1);
        bool can_filter_all = false;
        RETURN_IF_ERROR(conjunct->execute_filter(file_block, filter.data(),
                                                 static_cast<size_t>(batch_rows), false,
                                                 &can_filter_all));
        *selected_rows =
                can_filter_all ? 0 : apply_filter_to_selection(filter, selection, *selected_rows);
    }
    return Status::OK();
}

Status execute_delete_conjuncts(const format::FileScanRequest& request, int64_t batch_rows,
                                Block* file_block, SelectionVector* selection,
                                uint16_t* selected_rows) {
    for (const auto& delete_conjunct : request.delete_conjuncts) {
        if (*selected_rows == 0) {
            break;
        }
        DORIS_CHECK(delete_conjunct != nullptr);
        const size_t original_columns = file_block->columns();
        int result_column_id = -1;
        RETURN_IF_ERROR(delete_conjunct->root()->execute(delete_conjunct.get(), file_block,
                                                         &result_column_id));
        RETURN_IF_ERROR(detail::validate_ephemeral_expr_result_column(
                original_columns, result_column_id, file_block->columns()));
        const auto& delete_filter = assert_cast<const ColumnUInt8&>(
                                            *file_block->get_by_position(result_column_id).column)
                                            .get_data();
        DORIS_CHECK(delete_filter.size() == static_cast<size_t>(batch_rows));
        IColumn::Filter keep_filter(static_cast<size_t>(batch_rows), 1);
        bool has_kept_row = false;
        for (size_t row = 0; row < static_cast<size_t>(batch_rows); ++row) {
            keep_filter[row] = !delete_filter[row];
            has_kept_row |= keep_filter[row] != 0;
        }
        file_block->erase(result_column_id);
        *selected_rows =
                !has_kept_row ? 0
                              : apply_filter_to_selection(keep_filter, selection, *selected_rows);
    }
    return Status::OK();
}

} // namespace

Status detail::validate_ephemeral_expr_result_column(size_t original_columns, int result_column_id,
                                                     size_t current_columns) {
    // Delete predicates may erase only a temporary expression result. A bare SlotRef returns an
    // input column id, which must remain in the block for later predicates and materialization.
    if (UNLIKELY(result_column_id < 0 || static_cast<size_t>(result_column_id) < original_columns ||
                 static_cast<size_t>(result_column_id) >= current_columns)) {
        return Status::InternalError(
                "Delete conjunct result column {} is not ephemeral (original={}, current={})",
                result_column_id, original_columns, current_columns);
    }
    return Status::OK();
}

uint16_t apply_compact_filter_to_selection(const IColumn::Filter& filter,
                                           SelectionVector* selection, uint16_t selected_rows) {
    DORIS_CHECK(selection != nullptr);
    DORIS_CHECK(filter.size() == selected_rows);
    uint16_t new_selected_rows = 0;
    for (uint16_t selection_idx = 0; selection_idx < selected_rows; ++selection_idx) {
        if (filter[selection_idx] != 0) {
            selection->set_index(new_selected_rows++, static_cast<SelectionVector::Index>(
                                                              selection->get_index(selection_idx)));
        }
    }
    return new_selected_rows;
}

IColumn::Filter selection_to_filter(const SelectionVector& selection, uint16_t selected_rows,
                                    int64_t batch_rows) {
    IColumn::Filter filter(static_cast<size_t>(batch_rows), 0);
    for (uint16_t selection_idx = 0; selection_idx < selected_rows; ++selection_idx) {
        filter[selection.get_index(selection_idx)] = 1;
    }
    return filter;
}

Status execute_batch_filters(const format::FileScanRequest& request, int64_t batch_rows,
                             Block* file_block, SelectionVector* selection, uint16_t* selected_rows,
                             int64_t* conjunct_filtered_rows) {
    if (request.conjuncts.empty() && request.delete_conjuncts.empty()) {
        return Status::OK();
    }
    const auto selected_rows_before_conjunct = *selected_rows;
    RETURN_IF_ERROR(
            execute_filter_conjuncts(request, batch_rows, file_block, selection, selected_rows));
    if (conjunct_filtered_rows != nullptr) {
        *conjunct_filtered_rows += static_cast<int64_t>(selected_rows_before_conjunct) -
                                   static_cast<int64_t>(*selected_rows);
    }
    if (*selected_rows == 0) {
        return Status::OK();
    }
    return execute_delete_conjuncts(request, batch_rows, file_block, selection, selected_rows);
}

namespace {
int64_t count_range_rows(const std::vector<RowRange>& ranges) {
    int64_t rows = 0;
    for (const auto& range : ranges) {
        rows += range.length;
    }
    return rows;
}

void append_intersection(const RowRange& left, const RowRange& right,
                         std::vector<RowRange>& result) {
    const int64_t start = std::max(left.start, right.start);
    const int64_t end = std::min(left.start + left.length, right.start + right.length);
    if (start < end) {
        // Cache granules are only filter coordinates. Merge adjacent survivors so cache hits
        // preserve the original read-range batch boundaries, matching V1 RowRanges semantics.
        if (!result.empty() && result.back().start + result.back().length == start) {
            result.back().length = end - result.back().start;
            return;
        }
        result.push_back(RowRange {.start = start, .length = end - start});
    }
}

std::vector<RowRange> filter_ranges_by_condition_cache(const std::vector<RowRange>& ranges,
                                                       const std::vector<bool>& cache,
                                                       int64_t row_group_first_row,
                                                       int64_t base_granule) {
    std::vector<RowRange> result;
    if (cache.empty()) {
        return ranges;
    }

    // Cache coordinates are file-global granules; RowRange coordinates are row-group-relative.
    // Walk every selected range in order and split it by granule. Granules covered by the bitmap
    // are kept only when the bit is true. Granules outside the bitmap are kept conservatively, so
    // an undersized or old-format cache entry cannot skip valid rows.
    for (const auto& range : ranges) {
        const int64_t global_start = row_group_first_row + range.start;
        const int64_t global_end = global_start + range.length;
        for (int64_t granule = global_start / ConditionCacheContext::GRANULE_SIZE;
             granule <= (global_end - 1) / ConditionCacheContext::GRANULE_SIZE; ++granule) {
            const int64_t cache_idx = granule - base_granule;
            const bool keep = cache_idx < 0 || static_cast<size_t>(cache_idx) >= cache.size() ||
                              cache[static_cast<size_t>(cache_idx)];
            if (!keep) {
                continue;
            }
            const int64_t granule_start = granule * ConditionCacheContext::GRANULE_SIZE;
            const int64_t granule_end = granule_start + ConditionCacheContext::GRANULE_SIZE;
            const RowRange file_granule_range {.start = granule_start - row_group_first_row,
                                               .length = granule_end - granule_start};
            append_intersection(range, file_granule_range, result);
        }
    }
    return result;
}

} // namespace

void ParquetScanScheduler::set_plan(RowGroupScanPlan plan) {
    _enable_bloom_filter = plan.enable_bloom_filter;
    _row_group_plans = std::move(plan.row_groups);
    _condition_cache_filtered_rows = 0;
    _predicate_filtered_rows = 0;
    reset();
}

void ParquetScanScheduler::set_condition_cache_context(std::shared_ptr<ConditionCacheContext> ctx) {
    _condition_cache_ctx = std::move(ctx);
    if (!_condition_cache_ctx || !_condition_cache_ctx->filter_result || _row_group_plans.empty()) {
        return;
    }

    if (!_condition_cache_ctx->is_hit) {
        _condition_cache_ctx->base_granule =
                _row_group_plans.front().first_file_row / ConditionCacheContext::GRANULE_SIZE;
        const auto& last_plan = _row_group_plans.back();
        const int64_t end_granule = (last_plan.first_file_row + last_plan.row_group_rows +
                                     ConditionCacheContext::GRANULE_SIZE - 1) /
                                    ConditionCacheContext::GRANULE_SIZE;
        DORIS_CHECK(end_granule > _condition_cache_ctx->base_granule);
        _condition_cache_ctx->num_granules =
                std::min(_condition_cache_ctx->filter_result->size(),
                         static_cast<size_t>(end_granule - _condition_cache_ctx->base_granule));
        return;
    }

    std::vector<RowGroupReadPlan> filtered_plans;
    filtered_plans.reserve(_row_group_plans.size());
    for (auto& plan : _row_group_plans) {
        const int64_t old_rows = count_range_rows(plan.selected_ranges);
        plan.selected_ranges = filter_ranges_by_condition_cache(
                plan.selected_ranges, *_condition_cache_ctx->filter_result, plan.first_file_row,
                _condition_cache_ctx->base_granule);
        const int64_t new_rows = count_range_rows(plan.selected_ranges);
        _condition_cache_filtered_rows += old_rows - new_rows;
        if (!plan.selected_ranges.empty()) {
            filtered_plans.push_back(std::move(plan));
        }
    }
    _row_group_plans = std::move(filtered_plans);
    reset();
}

void ParquetScanScheduler::reset() {
    _next_row_group_plan_idx = 0;
    _raw_rows_read = 0;
    _predicate_schedule_request = nullptr;
    _predicate_schedule = {};
    _predicate_positions_scratch.clear();
    _predicate_indices_by_position_scratch.clear();
    _materialized_predicate_positions_scratch.clear();
    _ordered_predicate_positions_scratch.clear();
    _predicate_batch_sequence = 0;
    reset_current_row_group();
}

void ParquetScanScheduler::reset_current_row_group() {
    // RuntimeProfile updates are amortized on the batch path, but a row-group transition destroys
    // the reader tree. Force the final delta out before clearing it so short row groups and early
    // EOF paths cannot lose their last decode/IO timings.
    flush_current_reader_profiles();
    _batches_since_profile_flush = 0;
    _has_current_row_group = false;
    _current_predicate_columns.clear();
    _current_non_predicate_columns.clear();
    _current_dictionary_filters.clear();
    _current_dictionary_residual_conjuncts.clear();
    _current_row_group_rows = 0;
    _current_row_group_id = -1;
    _current_row_group_rows_read = 0;
    _current_row_group_first_row = 0;
    _current_selected_ranges.clear();
    _current_offset_indexes.clear();
    _current_range_idx = 0;
    _current_range_rows_read = 0;
    // Readers are row-group scoped. If every remaining row was filtered, no future output can
    // observe the non-predicate readers' position, so dropping them together with their pending lag
    // avoids a useless end-of-row-group SkipRecords call. Example: predicate readers advance from 0
    // to 10,000 while lazy readers stay at 0; clearing both readers here is sufficient because the
    // next row group constructs a new set starting at its own row 0.
    _pending_non_predicate_skip_rows = 0;
    _pending_predicate_batch_rows = 0;
    _pending_predicate_batch_rows_consumed = 0;
    _pending_predicate_selected_offset = 0;
    _pending_predicate_selection.clear();
    _pending_predicate_columns.clear();
    _pending_output_selection.clear();
    _current_predicate_prefetched = false;
    _current_non_predicate_prefetched = false;
    _current_merge_range_active = false;
}

void ParquetScanScheduler::flush_current_reader_profiles() {
    for (const auto& reader : _current_predicate_columns | std::views::values) {
        reader->flush_profile();
    }
    for (const auto& reader : _current_non_predicate_columns | std::views::values) {
        reader->flush_profile();
    }
}

bool ParquetScanScheduler::finish_current_reader_batch_profiles() {
    bool crossed_page = false;
    // A scheduler batch is counted once even when several projected leaves cross page boundaries.
    for (const auto& reader : _current_predicate_columns | std::views::values) {
        crossed_page |= reader->crossed_page_since_last_batch();
    }
    for (const auto& reader : _current_non_predicate_columns | std::views::values) {
        crossed_page |= reader->crossed_page_since_last_batch();
    }
    return crossed_page;
}

const detail::PredicateConjunctSchedule& ParquetScanScheduler::predicate_conjunct_schedule(
        const format::FileScanRequest& request) {
    if (_predicate_schedule_request == &request) {
        return _predicate_schedule;
    }

    // FileScanRequest is frozen by ParquetReader::open(). Its address therefore identifies both
    // the conjunct set and local-position mapping for the scheduler lifetime.
    _predicate_schedule = build_predicate_conjunct_schedule(request);
    _predicate_schedule_request = &request;
    _predicate_positions_scratch.clear();
    _predicate_indices_by_position_scratch.clear();
    _materialized_predicate_positions_scratch.clear();
    _predicate_positions_scratch.reserve(request.predicate_columns.size());
    _predicate_indices_by_position_scratch.reserve(request.predicate_columns.size());
    _materialized_predicate_positions_scratch.reserve(request.predicate_columns.size());
    for (size_t idx = 0; idx < request.predicate_columns.size(); ++idx) {
        const auto position_it =
                request.local_positions.find(request.predicate_columns[idx].column_id());
        DORIS_CHECK(position_it != request.local_positions.end());
        const size_t position = position_it->second.value();
        _predicate_positions_scratch.push_back(position);
        _predicate_indices_by_position_scratch.emplace(position, idx);
    }
    return _predicate_schedule;
}

std::vector<format::LocalColumnIndex> ParquetScanScheduler::adaptive_predicate_prefetch_columns(
        const format::FileScanRequest& request) {
    std::vector<size_t> positions;
    std::unordered_map<size_t, const format::LocalColumnIndex*> columns_by_position;
    columns_by_position.reserve(request.predicate_columns.size());
    for (const auto& column : request.predicate_columns) {
        const auto position_it = request.local_positions.find(column.column_id());
        DORIS_CHECK(position_it != request.local_positions.end());
        const size_t position = position_it->second.value();
        columns_by_position.emplace(position, &column);
    }
    const auto& schedule = predicate_conjunct_schedule(request);
    if (!schedule.supports_lazy_materialization) {
        positions.reserve(request.predicate_columns.size());
        for (const auto& column : request.predicate_columns) {
            positions.push_back(request.local_positions.at(column.column_id()).value());
        }
    } else if (!schedule.single_column_conjuncts.empty()) {
        positions.reserve(schedule.single_column_conjuncts.size());
        for (const auto& column : request.predicate_columns) {
            const size_t position = request.local_positions.at(column.column_id()).value();
            if (schedule.single_column_conjuncts.contains(position)) {
                // Cold adaptive statistics intentionally preserve request order; iterating the
                // hash map here would make the first decoded predicate depend on bucket layout.
                positions.push_back(position);
            }
        }
    } else if (!schedule.remaining_stages.empty()) {
        // Match execution's first reachable stage. Warming columns owned only by later residuals
        // would turn lazy decode into eager remote IO before an earlier conjunct can reject rows.
        positions = schedule.remaining_stages.front().required_positions;
    } else {
        positions.reserve(request.predicate_columns.size());
        for (const auto& column : request.predicate_columns) {
            positions.push_back(request.local_positions.at(column.column_id()).value());
        }
    }
    auto ordered = detail::order_adaptive_predicates(positions, _predicate_runtime_stats);
    ordered = detail::adaptive_prefetch_prefix(ordered, _predicate_runtime_stats, 0.25);
    std::vector<format::LocalColumnIndex> result;
    result.reserve(ordered.size());
    for (const size_t position : ordered) {
        result.push_back(*columns_by_position.at(position));
    }
    return result;
}

Status ParquetScanScheduler::open_next_row_group(
        ParquetFileContext& file_context,
        const std::vector<std::unique_ptr<ParquetColumnSchema>>& file_schema,
        const format::FileScanRequest& request, bool* has_row_group) {
    *has_row_group = false;
    RowGroupReadPlan* selected_plan = nullptr;
    while (_next_row_group_plan_idx < _row_group_plans.size()) {
        RowGroupReadPlan& candidate_plan = _row_group_plans[_next_row_group_plan_idx++];
        // Probe only the row group about to execute. This keeps LIMIT/cancellation latency
        // independent of the number of later remote row groups while preserving eager footer
        // statistics pruning during open.
        file_context.reset_random_access_ranges();
        _current_merge_range_active = false;
        ParquetPruningStats deferred_stats;
        bool selected = false;
        RETURN_IF_ERROR(finalize_native_row_group_read_plan(
                *file_context.native_metadata, file_schema, request, _enable_bloom_filter,
                &candidate_plan, &deferred_stats, _timezone, _runtime_state, &file_context,
                _scan_profile.column_reader_profile, &selected));
        if (_parquet_profile != nullptr) {
            _parquet_profile->update_deferred_pruning_stats(deferred_stats, selected);
        }
        if (!selected) {
            continue;
        }
        selected_plan = &candidate_plan;
        break;
    }
    if (selected_plan == nullptr) {
        // The last row group's native readers have already been released by
        // reset_current_row_group(). Flush the shared merge reader now so its counters are visible
        // when EOF is returned and its bounded scratch does not survive until file close.
        file_context.reset_random_access_ranges();
        _current_merge_range_active = false;
        return Status::OK();
    }
    RowGroupReadPlan& row_group_plan = *selected_plan;
    const int row_group_idx = row_group_plan.row_group_id;
    // Dictionary probes and data-page readers share the native metadata tree. Reset the previous
    // row-group merge reader before probing because dictionary-page offsets are not scan ordered.
    file_context.reset_random_access_ranges();
    _current_merge_range_active = false;

    const auto& row_group_metadata =
            file_context.native_metadata->to_thrift().row_groups[row_group_idx];
    _current_row_group_rows = row_group_metadata.num_rows;
    DORIS_CHECK(_current_row_group_rows == row_group_plan.row_group_rows);
    DORIS_CHECK(_current_row_group_rows > 0);
    _current_row_group_id = row_group_idx;
    _has_current_row_group = true;
    DORIS_CHECK(!row_group_plan.selected_ranges.empty());
    _current_row_group_first_row = row_group_plan.first_file_row;
    _current_row_group_rows_read = 0;
    _current_selected_ranges = row_group_plan.selected_ranges;
    _current_offset_indexes = std::move(row_group_plan.offset_indexes);
    // Condition Cache and split planning can narrow logical ranges without a physical OffsetIndex.
    // Native readers must keep the sequential level/value cursor path valid in that case; only a
    // PageIndex-derived skip plan requires the transferred indexes below.
    for (const auto& [leaf_column_id, skip_plan] : row_group_plan.page_skip_plans) {
        if (!_current_offset_indexes.contains(leaf_column_id)) {
            continue;
        }
        for (size_t page = 0; page < skip_plan.skipped_pages.size(); ++page) {
            if (!skip_plan.should_skip_page(page)) {
                continue;
            }
            if (_page_skip_profile.skipped_pages != nullptr) {
                COUNTER_UPDATE(_page_skip_profile.skipped_pages, 1);
            }
            if (_page_skip_profile.skipped_bytes != nullptr) {
                COUNTER_UPDATE(_page_skip_profile.skipped_bytes,
                               skip_plan.skipped_page_compressed_size(page));
            }
        }
    }
    _current_range_idx = 0;
    _current_range_rows_read = 0;
    _current_predicate_columns.clear();
    _current_non_predicate_columns.clear();
    _current_dictionary_filters.clear();
    RETURN_IF_ERROR(prepare_current_dictionary_filters(file_context, file_schema, request,
                                                       row_group_idx, row_group_metadata));
    // Dictionary probing is complete, so the native data-page readers can now share the same
    // row-group-scoped MergeRangeFileReader policy as v1. Sharing one wrapper is important: a
    // separate merge reader per leaf would duplicate its 128MB scratch capacity and defeat lazy
    // materialization for wide schemas.
    const auto& thrift_metadata = file_context.native_metadata->to_thrift();
    const auto compat = native::parquet_reader_compat(
            thrift_metadata.__isset.created_by ? thrift_metadata.created_by : std::string {});
    std::vector<ParquetPageCacheRange> native_ranges;
    RETURN_IF_ERROR(detail::build_native_prefetch_ranges(
            thrift_metadata, file_schema, request_scan_columns(request), row_group_idx,
            file_context.native_file->size(), compat.parquet_816_padding, &native_ranges));
    _current_merge_range_active = file_context.set_native_random_access_ranges(
            native_ranges, detail::average_prefetch_range_size(native_ranges), _profile,
            _merge_read_slice_size);

    for (const auto& col : request.predicate_columns) {
        const auto local_id = col.column_id();
        if (_current_predicate_columns.contains(local_id)) {
            continue;
        }
        if (local_id == format::LocalColumnId(format::ROW_POSITION_COLUMN_ID)) {
            _current_predicate_columns[local_id] = std::make_unique<RowPositionColumnReader>(
                    _current_row_group_first_row, _scan_profile.column_reader_profile);
            continue;
        }
        if (local_id == format::LocalColumnId(format::GLOBAL_ROWID_COLUMN_ID)) {
            DORIS_CHECK(_global_rowid_context.has_value());
            _current_predicate_columns[local_id] = std::make_unique<GlobalRowIdColumnReader>(
                    *_global_rowid_context, _current_row_group_first_row,
                    _scan_profile.column_reader_profile);
            continue;
        }

        DORIS_CHECK(local_id.is_valid() &&
                    local_id.value() < static_cast<int32_t>(file_schema.size()));
        const auto& column_schema = file_schema[local_id.value()];
        DORIS_CHECK(column_schema != nullptr);
        std::unique_ptr<ParquetColumnReader> column_reader;
        RETURN_IF_ERROR(NativeColumnReader::create(
                *column_schema, &col, file_context.native_data_file(), file_context.native_metadata,
                row_group_idx, _current_selected_ranges, _current_offset_indexes, _timezone,
                file_context.native_io_ctx, _runtime_state, file_context.native_page_cache_enabled,
                file_context.native_page_cache_file_key,
                _current_dictionary_filters.contains(local_id), _scan_profile.column_reader_profile,
                &column_reader));
        _current_predicate_columns[local_id] = std::move(column_reader);
    }
    // Start warming filter-column chunks as soon as their row group is selected. The native
    // BufferedFileStreamReader later consumes the same Doris file-cache blocks; prefetch never
    // changes row/column materialization order.
    if (!_current_merge_range_active) {
        const auto prefetch_columns = adaptive_predicate_prefetch_columns(request);
        RETURN_IF_ERROR(prefetch_current_row_group_columns(
                file_context, file_schema, prefetch_columns, &_current_predicate_prefetched));
    }
    for (const auto& col : request.non_predicate_columns) {
        const auto local_id = col.column_id();
        if (request.is_count_star_placeholder(col.column_id())) {
            continue;
        }
        if (local_id == format::LocalColumnId(format::ROW_POSITION_COLUMN_ID)) {
            _current_non_predicate_columns[local_id] = std::make_unique<RowPositionColumnReader>(
                    _current_row_group_first_row, _scan_profile.column_reader_profile);
            continue;
        }
        if (local_id == format::LocalColumnId(format::GLOBAL_ROWID_COLUMN_ID)) {
            DORIS_CHECK(_global_rowid_context.has_value());
            _current_non_predicate_columns[local_id] = std::make_unique<GlobalRowIdColumnReader>(
                    *_global_rowid_context, _current_row_group_first_row,
                    _scan_profile.column_reader_profile);
            continue;
        }
        DORIS_CHECK(local_id.is_valid() &&
                    local_id.value() < static_cast<int32_t>(file_schema.size()));
        const auto& column_schema = file_schema[local_id.value()];
        DORIS_CHECK(column_schema != nullptr);
        std::unique_ptr<ParquetColumnReader> column_reader;
        RETURN_IF_ERROR(NativeColumnReader::create(
                *column_schema, &col, file_context.native_data_file(), file_context.native_metadata,
                row_group_idx, _current_selected_ranges, _current_offset_indexes, _timezone,
                file_context.native_io_ctx, _runtime_state, file_context.native_page_cache_enabled,
                file_context.native_page_cache_file_key, false, _scan_profile.column_reader_profile,
                &column_reader));
        _current_non_predicate_columns[local_id] = std::move(column_reader);
    }
    if (!_current_merge_range_active &&
        ((request.conjuncts.empty() && request.delete_conjuncts.empty()) ||
         _predicate_survival_ratio >= 0.8)) {
        // With no row-level filters there is no lazy-read decision to wait for, so start warming
        // output chunks immediately after their readers are created. Filtered scans still defer
        // this until at least one row survives the predicate phase.
        RETURN_IF_ERROR(prefetch_current_row_group_columns(file_context, file_schema,
                                                           physical_non_predicate_columns(request),
                                                           &_current_non_predicate_prefetched));
    }
    *has_row_group = true;
    return Status::OK();
}

Status ParquetScanScheduler::skip_current_row_group_rows(int64_t rows) {
    DORIS_CHECK(rows >= 0);
    if (rows == 0) {
        return Status::OK();
    }
    if (_scan_profile.range_gap_skipped_rows != nullptr) {
        COUNTER_UPDATE(_scan_profile.range_gap_skipped_rows, rows);
    }
    for (const auto& column_reader : _current_predicate_columns | std::views::values) {
        RETURN_IF_ERROR(column_reader->skip(rows));
    }
    // Keep page-index/condition-cache gaps pending for lazy columns as well. For example, after a
    // fully filtered [0, 32) batch and a pruned [32, 96) gap, predicate readers are at 96 while lazy
    // readers remain at 0; one later skip(96) is cheaper than skip(32) followed by skip(64).
    DORIS_CHECK(_pending_non_predicate_skip_rows <= std::numeric_limits<int64_t>::max() - rows);
    _pending_non_predicate_skip_rows += rows;
    _current_row_group_rows_read += rows;
    return Status::OK();
}

Status ParquetScanScheduler::flush_pending_non_predicate_skip_rows() {
    if (_pending_non_predicate_skip_rows == 0) {
        return Status::OK();
    }
    for (const auto& column_reader : _current_non_predicate_columns | std::views::values) {
        RETURN_IF_ERROR(column_reader->skip(_pending_non_predicate_skip_rows));
    }
    _pending_non_predicate_skip_rows = 0;
    return Status::OK();
}

namespace {

bool append_residual_stages(const VExprContextSPtr& owner_context, const VExprSPtr& expression,
                            const std::unordered_set<size_t>& predicate_block_positions,
                            std::vector<detail::PredicateConjunctStage>* stages) {
    DORIS_CHECK(owner_context != nullptr);
    DORIS_CHECK(expression != nullptr);
    DORIS_CHECK(stages != nullptr);
    const auto* compound_predicate = dynamic_cast<const VCompoundPred*>(expression.get());
    if (compound_predicate != nullptr && compound_predicate->op() == TExprOpcode::COMPOUND_AND) {
        for (const auto& child : expression->children()) {
            if (!append_residual_stages(owner_context, child, predicate_block_positions, stages)) {
                return false;
            }
        }
        return true;
    }

    std::set<int> referenced_positions;
    expression->collect_slot_column_ids(referenced_positions);
    auto& stage = stages->emplace_back();
    stage.owner_context = owner_context;
    stage.expression = expression;
    for (const int position : referenced_positions) {
        if (position < 0 || !predicate_block_positions.contains(cast_set<size_t>(position))) {
            stages->pop_back();
            return false;
        }
        stage.required_positions.push_back(cast_set<size_t>(position));
    }
    return true;
}

detail::PredicateConjunctSchedule build_predicate_conjunct_schedule(
        const format::FileScanRequest& request) {
    std::unordered_set<size_t> predicate_block_positions;
    predicate_block_positions.reserve(request.predicate_columns.size());
    for (const auto& col : request.predicate_columns) {
        const auto position_it = request.local_positions.find(col.column_id());
        DORIS_CHECK(position_it != request.local_positions.end());
        predicate_block_positions.insert(position_it->second.value());
    }

    detail::PredicateConjunctSchedule schedule;
    for (const auto& conjunct : request.conjuncts) {
        DORIS_CHECK(conjunct != nullptr);
        DORIS_CHECK(conjunct->root() != nullptr);
        if (!conjunct->root()->is_safe_to_execute_on_selected_rows()) {
            // Round-by-round filtering can compact later predicate columns before evaluating
            // remaining expressions. Stateful functions such as random(1) and error-preserving
            // functions such as assert_true() must see the same full batch they saw before this
            // optimization, so any unsafe conjunct disables the per-column schedule for the batch.
            schedule.remaining_conjuncts = request.conjuncts;
            schedule.single_column_conjuncts.clear();
            schedule.remaining_stages.clear();
            schedule.supports_lazy_materialization = false;
            return schedule;
        }
        std::set<int> referenced_positions;
        conjunct->root()->collect_slot_column_ids(referenced_positions);
        if (referenced_positions.size() != 1) {
            schedule.remaining_conjuncts.push_back(conjunct);
            if (!append_residual_stages(conjunct, conjunct->root(), predicate_block_positions,
                                        &schedule.remaining_stages)) {
                schedule.supports_lazy_materialization = false;
                schedule.remaining_conjuncts = request.conjuncts;
                schedule.single_column_conjuncts.clear();
                schedule.remaining_stages.clear();
                return schedule;
            }
            continue;
        }
        const auto block_position = static_cast<size_t>(*referenced_positions.begin());
        if (!predicate_block_positions.contains(block_position)) {
            schedule.supports_lazy_materialization = false;
            schedule.remaining_conjuncts = request.conjuncts;
            schedule.single_column_conjuncts.clear();
            schedule.remaining_stages.clear();
            return schedule;
        }
        schedule.single_column_conjuncts[block_position].push_back(conjunct);
    }
    return schedule;
}

bool can_evaluate_all_with_dictionary(const VExprContextSPtrs& conjuncts) {
    if (conjuncts.empty()) {
        return false;
    }
    return std::ranges::all_of(conjuncts, [](const auto& conjunct) {
        return conjunct != nullptr && conjunct->root() != nullptr &&
               conjunct->root()->can_evaluate_dictionary_filter();
    });
}

bool can_evaluate_dictionary_exactly(const VExprSPtr& expr) {
    DORIS_CHECK(expr != nullptr);
    const auto* compound_pred = dynamic_cast<const VCompoundPred*>(expr.get());
    if (compound_pred == nullptr) {
        return expr->can_evaluate_dictionary_filter();
    }
    if (compound_pred->op() != TExprOpcode::COMPOUND_AND &&
        compound_pred->op() != TExprOpcode::COMPOUND_OR) {
        return false;
    }
    return !expr->children().empty() &&
           std::ranges::all_of(expr->children(), [](const auto& child) {
               return can_evaluate_dictionary_exactly(child);
           });
}

void collect_dictionary_residual_exprs(const VExprContextSPtr& owner_context, const VExprSPtr& expr,
                                       OwnedExpressionConjuncts* residual_conjuncts) {
    DORIS_CHECK(owner_context != nullptr);
    DORIS_CHECK(expr != nullptr);
    DORIS_CHECK(residual_conjuncts != nullptr);

    if (can_evaluate_dictionary_exactly(expr)) {
        return;
    }

    // VCompoundPred dictionary evaluation is a conservative prefilter for AND when only some
    // children are dictionary-aware. Split AND so exact dictionary children are not executed again
    // on materialized rows. Do not split a non-exact OR: its branches cannot be evaluated
    // independently after a dictionary prefilter without changing the original boolean semantics.
    const auto* compound_pred = dynamic_cast<const VCompoundPred*>(expr.get());
    if (compound_pred != nullptr && compound_pred->op() == TExprOpcode::COMPOUND_AND) {
        for (const auto& child : expr->children()) {
            collect_dictionary_residual_exprs(owner_context, child, residual_conjuncts);
        }
        return;
    }

    residual_conjuncts->emplace_back(owner_context, expr);
}

OwnedExpressionConjuncts build_dictionary_residual_conjuncts(const VExprContextSPtrs& conjuncts) {
    OwnedExpressionConjuncts residual_conjuncts;
    for (const auto& conjunct : conjuncts) {
        DORIS_CHECK(conjunct != nullptr);
        collect_dictionary_residual_exprs(conjunct, conjunct->root(), &residual_conjuncts);
    }
    return residual_conjuncts;
}

uint16_t count_selected_rows(const IColumn::Filter& filter) {
    uint16_t selected_rows = 0;
    for (const auto value : filter) {
        selected_rows += value != 0;
    }
    return selected_rows;
}

enum class DictionaryEntryFilterKernel {
    GENERIC,
    TYPED_FIXED_WIDTH,
    TYPED_STRING,
};

template <typename ColumnType>
bool get_fixed_dictionary_raw_values(const IColumn& dictionary, const uint8_t** values,
                                     size_t* value_width) {
    const auto* typed_dictionary = check_and_get_column<ColumnType>(dictionary);
    if (typed_dictionary == nullptr) {
        return false;
    }
    *values = reinterpret_cast<const uint8_t*>(typed_dictionary->get_data().data());
    *value_width = sizeof(typename ColumnType::value_type);
    return true;
}

bool get_typed_dictionary_raw_values(PrimitiveType primitive_type, const IColumn& dictionary,
                                     const uint8_t** values, size_t* value_width) {
    switch (primitive_type) {
#define GET_TYPED_DICTIONARY_VALUES(TYPE)                                                       \
    case TYPE:                                                                                  \
        return get_fixed_dictionary_raw_values<typename PrimitiveTypeTraits<TYPE>::ColumnType>( \
                dictionary, values, value_width)
        GET_TYPED_DICTIONARY_VALUES(TYPE_BOOLEAN);
        GET_TYPED_DICTIONARY_VALUES(TYPE_TINYINT);
        GET_TYPED_DICTIONARY_VALUES(TYPE_SMALLINT);
        GET_TYPED_DICTIONARY_VALUES(TYPE_INT);
        GET_TYPED_DICTIONARY_VALUES(TYPE_BIGINT);
        GET_TYPED_DICTIONARY_VALUES(TYPE_LARGEINT);
        GET_TYPED_DICTIONARY_VALUES(TYPE_FLOAT);
        GET_TYPED_DICTIONARY_VALUES(TYPE_DOUBLE);
        GET_TYPED_DICTIONARY_VALUES(TYPE_DATE);
        GET_TYPED_DICTIONARY_VALUES(TYPE_DATETIME);
        GET_TYPED_DICTIONARY_VALUES(TYPE_DATEV2);
        GET_TYPED_DICTIONARY_VALUES(TYPE_DATETIMEV2);
        GET_TYPED_DICTIONARY_VALUES(TYPE_TIMESTAMPTZ);
        GET_TYPED_DICTIONARY_VALUES(TYPE_DECIMAL32);
        GET_TYPED_DICTIONARY_VALUES(TYPE_DECIMAL64);
        GET_TYPED_DICTIONARY_VALUES(TYPE_DECIMALV2);
        GET_TYPED_DICTIONARY_VALUES(TYPE_DECIMAL128I);
        GET_TYPED_DICTIONARY_VALUES(TYPE_DECIMAL256);
        GET_TYPED_DICTIONARY_VALUES(TYPE_IPV4);
        GET_TYPED_DICTIONARY_VALUES(TYPE_IPV6);
#undef GET_TYPED_DICTIONARY_VALUES
    default:
        return false;
    }
}

enum class StringDictionaryCompareOp {
    EQ,
    NE,
    LT,
    LE,
    GT,
    GE,
};

std::optional<StringDictionaryCompareOp> string_dictionary_compare_op(std::string_view name,
                                                                      bool reverse) {
    StringDictionaryCompareOp op;
    if (name == "eq") {
        op = StringDictionaryCompareOp::EQ;
    } else if (name == "ne") {
        op = StringDictionaryCompareOp::NE;
    } else if (name == "lt") {
        op = StringDictionaryCompareOp::LT;
    } else if (name == "le") {
        op = StringDictionaryCompareOp::LE;
    } else if (name == "gt") {
        op = StringDictionaryCompareOp::GT;
    } else if (name == "ge") {
        op = StringDictionaryCompareOp::GE;
    } else {
        return std::nullopt;
    }
    if (!reverse || op == StringDictionaryCompareOp::EQ || op == StringDictionaryCompareOp::NE) {
        return op;
    }
    switch (op) {
    case StringDictionaryCompareOp::LT:
        return StringDictionaryCompareOp::GT;
    case StringDictionaryCompareOp::LE:
        return StringDictionaryCompareOp::GE;
    case StringDictionaryCompareOp::GT:
        return StringDictionaryCompareOp::LT;
    case StringDictionaryCompareOp::GE:
        return StringDictionaryCompareOp::LE;
    default:
        __builtin_unreachable();
    }
}

bool string_compare_matches(int comparison, StringDictionaryCompareOp op) {
    switch (op) {
    case StringDictionaryCompareOp::EQ:
        return comparison == 0;
    case StringDictionaryCompareOp::NE:
        return comparison != 0;
    case StringDictionaryCompareOp::LT:
        return comparison < 0;
    case StringDictionaryCompareOp::LE:
        return comparison <= 0;
    case StringDictionaryCompareOp::GT:
        return comparison > 0;
    case StringDictionaryCompareOp::GE:
        return comparison >= 0;
    }
    __builtin_unreachable();
}

bool try_apply_string_dictionary_conjunct(size_t block_position, const DataTypePtr& column_type,
                                          const VExprSPtr& root, const IColumn& dictionary,
                                          IColumn::Filter* dictionary_filter) {
    const auto fn = std::dynamic_pointer_cast<VectorizedFnCall>(root);
    if (fn == nullptr || (!dictionary.is_column_string() && !dictionary.is_column_string64())) {
        return false;
    }
    const auto slot_literal = expr_zonemap::extract_slot_and_literal(fn->children());
    if (!slot_literal.has_value() || slot_literal->slot_index != block_position ||
        slot_literal->literal.get_type() != TYPE_STRING ||
        !remove_nullable(slot_literal->slot_type)->equals(*remove_nullable(column_type)) ||
        !remove_nullable(slot_literal->literal_type)->equals(*remove_nullable(column_type))) {
        return false;
    }
    const auto op =
            string_dictionary_compare_op(fn->function_name(), slot_literal->literal_on_left);
    if (!op.has_value()) {
        return false;
    }
    const auto& literal = slot_literal->literal.get<TYPE_STRING>();
    const StringRef literal_ref(literal.data(), literal.size());
    for (size_t dictionary_id = 0; dictionary_id < dictionary.size(); ++dictionary_id) {
        const int comparison = dictionary.get_data_at(dictionary_id).compare(literal_ref);
        (*dictionary_filter)[dictionary_id] &= string_compare_matches(comparison, *op) ? 1 : 0;
    }
    return true;
}

Status build_dictionary_entry_filter(size_t block_position,
                                     const ParquetColumnSchema& column_schema,
                                     const VExprContextSPtrs& conjuncts, const IColumn& dictionary,
                                     IColumn::Filter* dictionary_filter,
                                     DictionaryEntryFilterKernel* kernel) {
    DORIS_CHECK(dictionary_filter != nullptr);
    DORIS_CHECK(kernel != nullptr);
    dictionary_filter->clear();
    dictionary_filter->resize_fill(dictionary.size(), 1);
    *kernel = DictionaryEntryFilterKernel::GENERIC;
    // Block positions are expression slot IDs here; validate the narrowing once so every
    // dictionary evaluation path uses the same representable ID.
    const int expression_column_id = cast_set<int>(block_position);
    const auto typed_data_type = remove_nullable(column_schema.type);
    const uint8_t* raw_values = nullptr;
    size_t value_width = 0;
    if (std::ranges::all_of(conjuncts,
                            [&](const auto& conjunct) {
                                return conjunct->root()->can_execute_on_raw_fixed_values(
                                        column_schema.type, expression_column_id);
                            }) &&
        get_typed_dictionary_raw_values(typed_data_type->get_primitive_type(), dictionary,
                                        &raw_values, &value_width)) {
        // A dictionary is immutable for the row group, so compare its contiguous typed values once
        // and reuse the resulting id bitmap for every data page.
        for (const auto& conjunct : conjuncts) {
            RETURN_IF_ERROR(conjunct->root()->execute_on_raw_fixed_values(
                    raw_values, dictionary.size(), value_width, column_schema.type,
                    expression_column_id, dictionary_filter->data()));
        }
        *kernel = DictionaryEntryFilterKernel::TYPED_FIXED_WIDTH;
        return Status::OK();
    }

    if (std::ranges::all_of(conjuncts, [&](const auto& conjunct) {
            return try_apply_string_dictionary_conjunct(block_position, column_schema.type,
                                                        conjunct->root(), dictionary,
                                                        dictionary_filter);
        })) {
        *kernel = DictionaryEntryFilterKernel::TYPED_STRING;
        return Status::OK();
    }

    dictionary_filter->clear();
    dictionary_filter->resize_fill(dictionary.size(), 1);
    DictionaryEvalContext ctx;
    auto& slot = ctx.slots
                         .emplace(expression_column_id,
                                  DictionaryEvalContext::SlotDictionary {
                                          .data_type = column_schema.type, .values = {}})
                         .first->second;
    slot.values.reserve(1);
    for (size_t dictionary_id = 0; dictionary_id < dictionary.size(); ++dictionary_id) {
        Field value;
        dictionary.get(dictionary_id, value);
        slot.values.clear();
        slot.values.push_back(std::move(value));
        (*dictionary_filter)[dictionary_id] =
                VExprContext::evaluate_dictionary_filter(conjuncts, ctx) ==
                                ZoneMapFilterResult::kNoMatch
                        ? 0
                        : 1;
    }
    return Status::OK();
}

} // namespace

Status ParquetScanScheduler::prepare_current_dictionary_filters(
        ParquetFileContext& file_context,
        const std::vector<std::unique_ptr<ParquetColumnSchema>>& file_schema,
        const format::FileScanRequest& request, int row_group_idx,
        const tparquet::RowGroup& row_group_metadata) {
    _current_dictionary_filters.clear();
    _current_dictionary_residual_conjuncts.clear();
    if (request.conjuncts.empty()) {
        return Status::OK();
    }
    detail::PredicateConjunctSchedule schedule;
    {
        SCOPED_TIMER(_scan_profile.dict_filter_expr_rewrite_time);
        schedule = predicate_conjunct_schedule(request);
    }
    if (schedule.single_column_conjuncts.empty()) {
        return Status::OK();
    }

    SCOPED_TIMER(_scan_profile.dict_filter_rewrite_time);
    for (const auto& col : request.predicate_columns) {
        const auto local_id = col.column_id();
        if (!local_id.is_valid() || local_id.value() >= static_cast<int32_t>(file_schema.size())) {
            continue;
        }
        const auto position_it = request.local_positions.find(col.column_id());
        DORIS_CHECK(position_it != request.local_positions.end());
        const auto block_position = static_cast<size_t>(position_it->second.value());
        const auto conjunct_it = schedule.single_column_conjuncts.find(block_position);
        if (conjunct_it == schedule.single_column_conjuncts.end() ||
            !can_evaluate_all_with_dictionary(conjunct_it->second)) {
            continue;
        }
        update_counter_if_not_null(_scan_profile.dict_filter_candidate_columns, 1);

        // This optimization is deliberately limited to single-column predicates with a dictionary
        // evaluable part. Mixed AND predicates are split so dictionary-covered children run as a
        // dict-id prefilter and residual children keep the normal row-level expression path.
        const auto& column_schema = file_schema[local_id.value()];
        DORIS_CHECK(column_schema != nullptr);
        if (column_schema->leaf_column_id < 0 ||
            column_schema->leaf_column_id >= static_cast<int>(row_group_metadata.columns.size())) {
            update_counter_if_not_null(_scan_profile.dict_filter_unsupported_columns, 1);
            continue;
        }
        const auto& column_chunk = row_group_metadata.columns[column_schema->leaf_column_id];
        if (!column_chunk.__isset.meta_data ||
            !supports_row_level_dictionary_filter(*column_schema, column_chunk.meta_data)) {
            update_counter_if_not_null(_scan_profile.dict_filter_unsupported_columns, 1);
            continue;
        }

        std::unique_ptr<ParquetColumnReader> column_reader;
        RETURN_IF_ERROR(NativeColumnReader::create(
                *column_schema, &col, file_context.native_file, file_context.native_metadata,
                row_group_idx, _current_selected_ranges, _current_offset_indexes, _timezone,
                file_context.native_io_ctx, _runtime_state, file_context.native_page_cache_enabled,
                file_context.native_page_cache_file_key, true, _scan_profile.column_reader_profile,
                &column_reader));
        MutableColumnPtr dictionary_values;
        {
            SCOPED_TIMER(_scan_profile.dict_filter_read_dict_time);
            auto dictionary_result = column_reader->dictionary_values();
            if (!dictionary_result.has_value()) {
                update_counter_if_not_null(_scan_profile.dict_filter_read_failures, 1);
                // Dictionary filtering is optional: a probe failure must not reject a file that
                // the normal native read path can still decode.
                continue;
            }
            dictionary_values = std::move(dictionary_result).value();
        }

        // Build a safe dictionary prefilter from the dictionary-filter interface instead of
        // executing the row expression on a temporary dictionary block. For compound AND,
        // VCompoundPred intentionally evaluates only dictionary-capable children, so residual
        // predicates still run later on surviving rows.
        IColumn::Filter dictionary_filter;
        OwnedExpressionConjuncts residual_conjuncts;
        {
            SCOPED_TIMER(_scan_profile.dict_filter_build_time);
            DictionaryEntryFilterKernel filter_kernel = DictionaryEntryFilterKernel::GENERIC;
            RETURN_IF_ERROR(build_dictionary_entry_filter(block_position, *column_schema,
                                                          conjunct_it->second, *dictionary_values,
                                                          &dictionary_filter, &filter_kernel));
            if (filter_kernel == DictionaryEntryFilterKernel::TYPED_FIXED_WIDTH) {
                update_counter_if_not_null(_scan_profile.dict_filter_typed_compare_columns, 1);
            } else if (filter_kernel == DictionaryEntryFilterKernel::TYPED_STRING) {
                update_counter_if_not_null(_scan_profile.dict_filter_string_compare_columns, 1);
            }
            residual_conjuncts = build_dictionary_residual_conjuncts(conjunct_it->second);
        }

        // The bitmap is keyed by Parquet dictionary id. Later data-page reads evaluate the
        // predicate with an integer lookup and materialize typed values only for surviving rows.
        _current_dictionary_filters.emplace(local_id, std::move(dictionary_filter));
        _current_dictionary_residual_conjuncts.emplace(local_id, std::move(residual_conjuncts));
        _current_predicate_columns.emplace(local_id, std::move(column_reader));
        update_counter_if_not_null(_scan_profile.dict_filter_columns, 1);
    }
    return Status::OK();
}

Status ParquetScanScheduler::read_filter_columns(int64_t batch_rows,
                                                 const format::FileScanRequest& request,
                                                 Block* file_block, SelectionVector* selection,
                                                 uint16_t* selected_rows,
                                                 int64_t* conjunct_filtered_rows,
                                                 bool* predicate_columns_filtered) {
    DORIS_CHECK(predicate_columns_filtered != nullptr);
    *predicate_columns_filtered = false;
    if (!request.conjuncts.empty() || !request.delete_conjuncts.empty()) {
        selection->resize(static_cast<size_t>(batch_rows));
    }
    const auto& schedule = predicate_conjunct_schedule(request);
    std::unordered_set<size_t> residual_predicate_positions;
    auto remember_residual_positions = [&](const VExprContextSPtrs& conjuncts) {
        for (const auto& conjunct : conjuncts) {
            std::set<int> positions;
            conjunct->root()->collect_slot_column_ids(positions);
            for (const int position : positions) {
                if (position >= 0) {
                    residual_predicate_positions.insert(cast_set<size_t>(position));
                }
            }
        }
    };
    remember_residual_positions(schedule.remaining_conjuncts);
    remember_residual_positions(request.delete_conjuncts);
    const size_t predicate_batch_sequence = _predicate_batch_sequence++;
    const bool can_read_predicate_columns_round_by_round = schedule.supports_lazy_materialization;
    auto& read_column_positions = _read_column_positions_scratch;
    read_column_positions.clear();
    read_column_positions.reserve(request.predicate_columns.size());
    auto& materialized_positions = _materialized_predicate_positions_scratch;
    materialized_positions.clear();
    for (auto& rows : _predicate_column_selection_scratch | std::views::values) {
        rows.clear();
    }
    // A generation becomes dirty only when filtering changes SelectionVector. Columns read after
    // an all-pass stage already share its coordinates, so rewalking every prior mapping is wasted.
    bool predicate_columns_need_alignment = false;

    auto remember_column_selection = [&](uint32_t position) {
        auto& rows = _predicate_column_selection_scratch[position];
        rows.resize(*selected_rows);
        for (uint16_t row = 0; row < *selected_rows; ++row) {
            // SelectionVector and the scanner batch contract both bound row ordinals to uint16_t;
            // keep the checked conversion explicit when persisting the coordinate mapping.
            rows[row] = cast_set<uint16_t>(selection->get_index(row));
        }
    };

    auto compact_predicate_columns = [&](bool discard_predicate_only_payload) -> Status {
        bool compacted = false;
        int64_t compacted_bytes = 0;
        update_counter_if_not_null(_scan_profile.predicate_alignment_columns,
                                   cast_set<int64_t>(read_column_positions.size()));
        for (const uint32_t position : read_column_positions) {
            auto& source_rows = _predicate_column_selection_scratch[position];
            const auto& old_column = file_block->get_by_position(position).column;
            if (old_column->size() != source_rows.size()) {
                return Status::Corruption(
                        "Predicate column {} has {} values but {} remembered source rows", position,
                        old_column->size(), source_rows.size());
            }
            bool predicate_only = false;
            if (discard_predicate_only_payload) {
                predicate_only = std::ranges::any_of(
                        request.predicate_only_columns, [&](format::LocalColumnId local_id) {
                            const auto position_it = request.local_positions.find(local_id);
                            return position_it != request.local_positions.end() &&
                                   position_it->second.value() == position;
                        });
            }
            if (predicate_only) {
                auto placeholder = old_column->clone_empty();
                // Hidden predicate values are dead after the last filter, but every file-block
                // column must retain the selected row count until TableReader drops hidden slots.
                placeholder->insert_many_defaults(*selected_rows);
                file_block->replace_by_position(position, std::move(placeholder));
                remember_column_selection(position);
                continue;
            }
            bool already_compact = source_rows.size() == *selected_rows &&
                                   old_column->size() == static_cast<size_t>(*selected_rows);
            for (uint16_t row = 0; already_compact && row < *selected_rows; ++row) {
                already_compact = source_rows[row] == selection->get_index(row);
            }
            if (already_compact) {
                continue;
            }
            auto& filter = _predicate_compaction_filter_scratch;
            // resize_fill() preserves bytes when the next predicate column is smaller. Clear the
            // whole reusable mask so survivors from an earlier coordinate space cannot reappear.
            filter.resize(source_rows.size());
            std::ranges::fill(filter, 0);
            size_t source_idx = 0;
            uint16_t selected_idx = 0;
            while (source_idx < source_rows.size() && selected_idx < *selected_rows) {
                const auto source_row = source_rows[source_idx];
                const auto selected_row = selection->get_index(selected_idx);
                if (source_row < selected_row) {
                    ++source_idx;
                    continue;
                }
                DORIS_CHECK_EQ(source_row, selected_row);
                filter[source_idx++] = 1;
                ++selected_idx;
            }
            DORIS_CHECK_EQ(selected_idx, *selected_rows);
            compacted_bytes += static_cast<int64_t>(old_column->byte_size());
            RETURN_IF_CATCH_EXCEPTION(file_block->replace_by_position(
                    position, old_column->filter(filter, *selected_rows)));
            remember_column_selection(position);
            compacted = true;
        }
        if (compacted) {
            update_counter_if_not_null(_scan_profile.predicate_compaction_bytes, compacted_bytes);
            update_counter_if_not_null(_scan_profile.predicate_compaction_count, 1);
        }
        // The output path must not apply a batch-coordinate filter to columns that now use compact
        // coordinates. The loop above establishes this invariant even when no bytes moved because
        // every column was already aligned.
        *predicate_columns_filtered = !read_column_positions.empty();
        return Status::OK();
    };

    auto read_predicate_column =
            [&](ParquetColumnReader* column_reader, size_t block_position,
                format::LocalColumnId local_id, const VExprContextSPtrs* single_column_conjuncts,
                bool* used_dictionary_filter, bool* used_fixed_width_filter) -> Status {
        DORIS_CHECK(used_dictionary_filter != nullptr);
        DORIS_CHECK(used_fixed_width_filter != nullptr);
        *used_dictionary_filter = false;
        *used_fixed_width_filter = false;
        DCHECK(remove_nullable(column_reader->type())
                       ->equals(*remove_nullable(file_block->get_by_position(block_position).type)))
                << column_reader->type()->get_name() << " "
                << file_block->get_by_position(block_position).type->get_name() << " "
                << column_reader->name() << " " << file_block->get_by_position(block_position).name;
        auto column = file_block->get_by_position(block_position).column->assert_mutable();
        SCOPED_TIMER(_scan_profile.column_read_time);
        const auto dictionary_filter_it = _current_dictionary_filters.find(local_id);
        if (dictionary_filter_it != _current_dictionary_filters.end()) {
            const uint16_t selected_rows_before = *selected_rows;
            IColumn::Filter compact_filter;
            uint16_t new_selected_rows = 0;
            bool used_filter = false;
            const bool predicate_only = request.is_predicate_only(local_id);
            // Dictionary ids are sufficient for predicate-only slots; skipping typed survivor
            // gathers preserves the block row shape without materializing an unobservable payload.
            IColumn* projected_column = predicate_only ? nullptr : column.get();
            RETURN_IF_ERROR(column_reader->select_with_dictionary_filter(
                    *selection, *selected_rows, batch_rows, dictionary_filter_it->second,
                    projected_column, &compact_filter, &new_selected_rows, &used_filter));
            if (used_filter) {
                DORIS_CHECK(compact_filter.size() == selected_rows_before);
                DORIS_CHECK(new_selected_rows <= selected_rows_before);
                update_counter_if_not_null(_scan_profile.dictionary_predicate_direct_batches, 1);
                update_counter_if_not_null(_scan_profile.dictionary_predicate_direct_rows,
                                           selected_rows_before);
                // The decoder already observes every keep bit while producing compact_filter, so
                // reuse its count instead of adding another full filter scan at this boundary.
                if (!predicate_only) {
                    update_counter_if_not_null(_scan_profile.dictionary_predicate_projected_rows,
                                               new_selected_rows);
                }
                const auto filtered_rows = static_cast<int64_t>(selected_rows_before) -
                                           static_cast<int64_t>(new_selected_rows);
                if (conjunct_filtered_rows != nullptr) {
                    *conjunct_filtered_rows += filtered_rows;
                }
                update_counter_if_not_null(_scan_profile.rows_filtered_by_dict_filter,
                                           filtered_rows);
                if (new_selected_rows != selected_rows_before) {
                    // The dictionary reader already appended only survivors for this column. Keep
                    // older predicate columns in their original coordinate spaces and compact all
                    // of them once at the expression/output boundary below.
                    *selected_rows = apply_compact_filter_to_selection(compact_filter, selection,
                                                                       selected_rows_before);
                }
                if (predicate_only) {
                    auto placeholder = column->clone_empty();
                    placeholder->insert_many_defaults(*selected_rows);
                    file_block->replace_by_position(block_position, std::move(placeholder));
                } else {
                    file_block->replace_by_position(block_position, std::move(column));
                }
                read_column_positions.push_back(cast_set<uint32_t>(block_position));
                remember_column_selection(cast_set<uint32_t>(block_position));
                *used_dictionary_filter = true;
                return Status::OK();
            }
        }

        if (single_column_conjuncts != nullptr &&
            !residual_predicate_positions.contains(block_position)) {
            VExprSPtrs direct_conjuncts;
            direct_conjuncts.reserve(single_column_conjuncts->size());
            std::ranges::transform(*single_column_conjuncts, std::back_inserter(direct_conjuncts),
                                   [](const auto& context) { return context->root(); });
            if (!direct_conjuncts.empty()) {
                const uint16_t selected_rows_before = *selected_rows;
                IColumn::Filter compact_filter;
                bool used_filter = false;
                const bool predicate_only = request.is_predicate_only(local_id);
                // The raw decoder cannot rewind after evaluating encoded fixed-width values.
                // Project survivors in that pass when output still needs the predicate column.
                IColumn* projected_column = predicate_only ? nullptr : column.get();
                RETURN_IF_ERROR(column_reader->select_with_fixed_width_filter(
                        *selection, *selected_rows, batch_rows, direct_conjuncts,
                        cast_set<int>(block_position), projected_column, &compact_filter,
                        &used_filter));
                if (used_filter) {
                    DORIS_CHECK_EQ(compact_filter.size(), selected_rows_before);
                    update_counter_if_not_null(_scan_profile.fixed_width_predicate_direct_batches,
                                               1);
                    update_counter_if_not_null(_scan_profile.fixed_width_predicate_direct_rows,
                                               selected_rows_before);
                    const uint16_t new_selected_rows = count_selected_rows(compact_filter);
                    const auto filtered_rows = static_cast<int64_t>(selected_rows_before) -
                                               static_cast<int64_t>(new_selected_rows);
                    if (conjunct_filtered_rows != nullptr) {
                        *conjunct_filtered_rows += filtered_rows;
                    }
                    if (new_selected_rows != selected_rows_before) {
                        *selected_rows = apply_compact_filter_to_selection(
                                compact_filter, selection, selected_rows_before);
                    }
                    if (predicate_only) {
                        // This slot is absent from every residual/delete conjunct, so no later
                        // expression can observe its payload. Keep only the block row-shape contract.
                        auto placeholder = column->clone_empty();
                        placeholder->insert_many_defaults(*selected_rows);
                        file_block->replace_by_position(block_position, std::move(placeholder));
                    } else {
                        file_block->replace_by_position(block_position, std::move(column));
                    }
                    read_column_positions.push_back(cast_set<uint32_t>(block_position));
                    remember_column_selection(cast_set<uint32_t>(block_position));
                    *predicate_columns_filtered = true;
                    *used_fixed_width_filter = true;
                    return Status::OK();
                }
            }
        }

        if (*selected_rows == batch_rows) {
            int64_t column_rows = 0;
            RETURN_IF_ERROR(column_reader->read(batch_rows, column, &column_rows));
            if (column_rows != batch_rows) {
                return Status::Corruption(
                        "Parquet filter column {} returned {} rows, expected {} rows",
                        column_reader->name(), column_rows, batch_rows);
            }
        } else {
            [[maybe_unused]] auto old_size = column->size();
            RETURN_IF_ERROR(column_reader->select(*selection, *selected_rows, batch_rows, column));
            if (column->size() != old_size + *selected_rows) {
                return Status::Corruption(
                        "Parquet selected filter column {} returned {} rows, expected {} rows",
                        column_reader->name(), column->size(), old_size + *selected_rows);
            }
            *predicate_columns_filtered = true;
        }
        file_block->replace_by_position(block_position, std::move(column));
        read_column_positions.push_back(cast_set<uint32_t>(block_position));
        remember_column_selection(cast_set<uint32_t>(block_position));
        return Status::OK();
    };

    auto execute_scheduled_conjuncts = [&](const VExprContextSPtrs& conjuncts) -> Status {
        if (conjuncts.empty() || *selected_rows == 0) {
            return Status::OK();
        }
        const uint16_t selected_rows_before = *selected_rows;
        IColumn::Filter compact_filter;
        bool can_filter_all = false;
        RETURN_IF_ERROR(execute_compact_filter_conjuncts(
                conjuncts, selected_rows_before, file_block, &compact_filter, &can_filter_all));
        if (can_filter_all) {
            compact_filter.resize_fill(selected_rows_before, 0);
        }
        const uint16_t new_selected_rows = can_filter_all ? 0 : count_selected_rows(compact_filter);
        if (conjunct_filtered_rows != nullptr) {
            *conjunct_filtered_rows += static_cast<int64_t>(selected_rows_before) -
                                       static_cast<int64_t>(new_selected_rows);
        }
        if (new_selected_rows != selected_rows_before) {
            predicate_columns_need_alignment = true;
            *selected_rows = can_filter_all
                                     ? 0
                                     : apply_compact_filter_to_selection(compact_filter, selection,
                                                                         selected_rows_before);
        }
        return Status::OK();
    };

    auto execute_scheduled_owned_conjuncts =
            [&](std::span<const OwnedExpressionConjunct> conjuncts) -> Status {
        if (conjuncts.empty() || *selected_rows == 0) {
            return Status::OK();
        }
        const uint16_t selected_rows_before = *selected_rows;
        IColumn::Filter compact_filter;
        bool can_filter_all = false;
        RETURN_IF_ERROR(execute_compact_owned_conjuncts(conjuncts, selected_rows_before, file_block,
                                                        &compact_filter, &can_filter_all));
        if (can_filter_all) {
            compact_filter.resize_fill(selected_rows_before, 0);
        }
        const uint16_t new_selected_rows = can_filter_all ? 0 : count_selected_rows(compact_filter);
        if (conjunct_filtered_rows != nullptr) {
            *conjunct_filtered_rows += static_cast<int64_t>(selected_rows_before) -
                                       static_cast<int64_t>(new_selected_rows);
        }
        if (new_selected_rows != selected_rows_before) {
            predicate_columns_need_alignment = true;
            *selected_rows = can_filter_all
                                     ? 0
                                     : apply_compact_filter_to_selection(compact_filter, selection,
                                                                         selected_rows_before);
        }
        return Status::OK();
    };

    auto execute_scheduled_conjuncts_with_profile =
            [&](const VExprContextSPtrs& conjuncts) -> Status {
        if (_scan_profile.predicate_filter_time == nullptr) {
            return execute_scheduled_conjuncts(conjuncts);
        }
        SCOPED_TIMER(_scan_profile.predicate_filter_time);
        return execute_scheduled_conjuncts(conjuncts);
    };

    auto execute_scheduled_owned_conjuncts_with_profile =
            [&](std::span<const OwnedExpressionConjunct> conjuncts) -> Status {
        if (_scan_profile.predicate_filter_time == nullptr) {
            return execute_scheduled_owned_conjuncts(conjuncts);
        }
        SCOPED_TIMER(_scan_profile.predicate_filter_time);
        return execute_scheduled_owned_conjuncts(conjuncts);
    };

    auto execute_scheduled_delete_conjuncts = [&]() -> Status {
        if (request.delete_conjuncts.empty() || *selected_rows == 0) {
            return Status::OK();
        }
        const uint16_t selected_rows_before = *selected_rows;
        IColumn::Filter compact_filter;
        bool can_filter_all = false;
        RETURN_IF_ERROR(execute_compact_delete_conjuncts(request.delete_conjuncts,
                                                         selected_rows_before, file_block,
                                                         &compact_filter, &can_filter_all));
        if (can_filter_all) {
            compact_filter.resize_fill(selected_rows_before, 0);
        }
        if (can_filter_all || count_selected_rows(compact_filter) != selected_rows_before) {
            predicate_columns_need_alignment = true;
            *selected_rows = can_filter_all
                                     ? 0
                                     : apply_compact_filter_to_selection(compact_filter, selection,
                                                                         selected_rows_before);
        }
        return Status::OK();
    };

    auto read_all_predicate_columns = [&]() -> Status {
        for (const auto& [fid, column_reader] : _current_predicate_columns) {
            auto position_it = request.local_positions.find(fid);
            DORIS_CHECK(position_it != request.local_positions.end());
            bool used_dictionary_filter = false;
            bool used_fixed_width_filter = false;
            RETURN_IF_ERROR(read_predicate_column(column_reader.get(), position_it->second.value(),
                                                  fid, nullptr, &used_dictionary_filter,
                                                  &used_fixed_width_filter));
            materialized_positions.insert(position_it->second.value());
        }
        return Status::OK();
    };

    if (!can_read_predicate_columns_round_by_round) {
        RETURN_IF_ERROR(read_all_predicate_columns());
        if (_scan_profile.predicate_filter_time == nullptr) {
            return execute_batch_filters(request, batch_rows, file_block, selection, selected_rows,
                                         conjunct_filtered_rows);
        }
        SCOPED_TIMER(_scan_profile.predicate_filter_time);
        return execute_batch_filters(request, batch_rows, file_block, selection, selected_rows,
                                     conjunct_filtered_rows);
    }

    auto read_round_by_round = [&]() -> Status {
        // Single-column conjuncts can be evaluated immediately after their column is read. Once
        // selection shrinks, later predicate columns use ParquetColumnReader::select() so the
        // reader skips rows already rejected by earlier predicates instead of materializing them.
        _ordered_predicate_positions_scratch.clear();
        _ordered_predicate_positions_scratch.reserve(schedule.single_column_conjuncts.size());
        for (const auto& column : request.predicate_columns) {
            const size_t position = request.local_positions.at(column.column_id()).value();
            if (schedule.single_column_conjuncts.contains(position)) {
                // The request order is the stable cold-start policy until measured costs can
                // reorder predicates; unordered-map iteration can defeat an early selective filter.
                _ordered_predicate_positions_scratch.push_back(position);
            }
        }
        _ordered_predicate_positions_scratch = detail::order_adaptive_predicates(
                _ordered_predicate_positions_scratch, _predicate_runtime_stats);
        const auto& ordered_positions = _ordered_predicate_positions_scratch;
        for (size_t order_idx = 0; order_idx < ordered_positions.size(); ++order_idx) {
            const size_t position = ordered_positions[order_idx];
            const size_t idx = _predicate_indices_by_position_scratch.at(position);
            const auto& col = request.predicate_columns[idx];
            const auto fid = col.column_id();
            auto reader_it = _current_predicate_columns.find(fid);
            DORIS_CHECK(reader_it != _current_predicate_columns.end());
            auto position_it = request.local_positions.find(col.column_id());
            DORIS_CHECK(position_it != request.local_positions.end());
            const auto block_position = position_it->second.value();
            const uint16_t rows_before = *selected_rows;
            auto& stats = _predicate_runtime_stats[position];
            const bool sample = detail::should_sample_adaptive_predicate(stats.samples,
                                                                         predicate_batch_sequence);
            const int64_t start_ns = sample ? MonotonicNanos() : 0;
            bool used_dictionary_filter = false;
            bool used_fixed_width_filter = false;
            const auto conjunct_it = schedule.single_column_conjuncts.find(block_position);
            const VExprContextSPtrs* column_conjuncts =
                    conjunct_it == schedule.single_column_conjuncts.end() ? nullptr
                                                                          : &conjunct_it->second;
            RETURN_IF_ERROR(read_predicate_column(reader_it->second.get(), block_position, fid,
                                                  column_conjuncts, &used_dictionary_filter,
                                                  &used_fixed_width_filter));
            materialized_positions.insert(block_position);
            if (*selected_rows != 0 && conjunct_it != schedule.single_column_conjuncts.end()) {
                if (used_dictionary_filter) {
                    const auto residual_it = _current_dictionary_residual_conjuncts.find(fid);
                    DORIS_CHECK(residual_it != _current_dictionary_residual_conjuncts.end());
                    RETURN_IF_ERROR(
                            execute_scheduled_owned_conjuncts_with_profile(residual_it->second));
                } else if (!used_fixed_width_filter) {
                    RETURN_IF_ERROR(execute_scheduled_conjuncts_with_profile(conjunct_it->second));
                }
            }
            if (*selected_rows != rows_before) {
                predicate_columns_need_alignment = true;
            }
            if (sample) {
                const double cost_per_row = static_cast<double>(MonotonicNanos() - start_ns) /
                                            std::max<uint16_t>(rows_before, 1);
                const double survival =
                        static_cast<double>(*selected_rows) / std::max<uint16_t>(rows_before, 1);
                constexpr double ADAPTIVE_ALPHA = 0.25;
                if (stats.samples == 0) {
                    stats.cost_per_input_row_ns = cost_per_row;
                    stats.survival_ratio = survival;
                } else {
                    stats.cost_per_input_row_ns =
                            ADAPTIVE_ALPHA * cost_per_row +
                            (1 - ADAPTIVE_ALPHA) * stats.cost_per_input_row_ns;
                    stats.survival_ratio =
                            ADAPTIVE_ALPHA * survival + (1 - ADAPTIVE_ALPHA) * stats.survival_ratio;
                }
                ++stats.samples;
            }
            if (*selected_rows != 0) {
                continue;
            }
            return Status::OK();
        }
        return Status::OK();
    };

    auto materialize_predicate_positions = [&](const std::vector<size_t>& positions) -> Status {
        for (const size_t position : positions) {
            if (materialized_positions.contains(position)) {
                continue;
            }
            const auto index_it = _predicate_indices_by_position_scratch.find(position);
            DORIS_CHECK(index_it != _predicate_indices_by_position_scratch.end());
            const auto fid = request.predicate_columns[index_it->second].column_id();
            const auto reader_it = _current_predicate_columns.find(fid);
            DORIS_CHECK(reader_it != _current_predicate_columns.end());
            bool used_dictionary_filter = false;
            bool used_fixed_width_filter = false;
            RETURN_IF_ERROR(read_predicate_column(reader_it->second.get(), position, fid, nullptr,
                                                  &used_dictionary_filter,
                                                  &used_fixed_width_filter));
            materialized_positions.insert(position);
        }
        return Status::OK();
    };

    auto skip_unmaterialized_predicate_columns = [&]() -> Status {
        for (const auto& col : request.predicate_columns) {
            const auto position_it = request.local_positions.find(col.column_id());
            DORIS_CHECK(position_it != request.local_positions.end());
            if (materialized_positions.contains(position_it->second.value())) {
                continue;
            }
            const auto reader_it = _current_predicate_columns.find(col.column_id());
            DORIS_CHECK(reader_it != _current_predicate_columns.end());
            RETURN_IF_ERROR(reader_it->second->skip(batch_rows));
        }
        // Every skipped column has an empty payload in the block. Suppress the caller's
        // batch-coordinate filter because there is no materialized batch-sized column left.
        *predicate_columns_filtered = true;
        return Status::OK();
    };

    auto compact_predicate_columns_with_profile =
            [&](bool discard_predicate_only_payload) -> Status {
        if (!discard_predicate_only_payload && !predicate_columns_need_alignment) {
            return Status::OK();
        }
        const int64_t start_ns = MonotonicNanos();
        auto status = compact_predicate_columns(discard_predicate_only_payload);
        update_counter_if_not_null(_scan_profile.predicate_compaction_time,
                                   MonotonicNanos() - start_ns);
        if (status.ok()) {
            predicate_columns_need_alignment = false;
        }
        return status;
    };

    RETURN_IF_ERROR(read_round_by_round());
    if (*selected_rows == 0) {
        RETURN_IF_ERROR(skip_unmaterialized_predicate_columns());
        return compact_predicate_columns_with_profile(true);
    }

    // Complex residuals keep their original conjunct order. Materialize only the columns needed
    // by the next reachable expression, then compact previously read columns into the same row
    // space before evaluating it. This is the scanner-side equivalent of expression-triggered
    // lazy columns: a conjunct that rejects the batch prevents later-only columns from decoding.
    for (const auto& stage : schedule.remaining_stages) {
        RETURN_IF_ERROR(materialize_predicate_positions(stage.required_positions));
        RETURN_IF_ERROR(compact_predicate_columns_with_profile(false));
        const OwnedExpressionConjunct stage_conjunct {stage.owner_context, stage.expression};
        RETURN_IF_ERROR(execute_scheduled_owned_conjuncts_with_profile(
                std::span<const OwnedExpressionConjunct>(&stage_conjunct, 1)));
        if (*selected_rows == 0) {
            RETURN_IF_ERROR(skip_unmaterialized_predicate_columns());
            return compact_predicate_columns_with_profile(true);
        }
    }

    if (!request.delete_conjuncts.empty()) {
        std::set<int> delete_positions;
        for (const auto& conjunct : request.delete_conjuncts) {
            DORIS_CHECK(conjunct != nullptr && conjunct->root() != nullptr);
            conjunct->root()->collect_slot_column_ids(delete_positions);
        }
        std::vector<size_t> required_delete_positions;
        required_delete_positions.reserve(delete_positions.size());
        for (const int position : delete_positions) {
            DORIS_CHECK(position >= 0);
            required_delete_positions.push_back(cast_set<size_t>(position));
        }
        if (required_delete_positions.empty() && !_predicate_positions_scratch.empty()) {
            // An all-literal equality-delete predicate has no slot dependency, but its hidden
            // row-count carrier must still be materialized so the result matches selected_rows.
            required_delete_positions.push_back(_predicate_positions_scratch.front());
        }
        RETURN_IF_ERROR(materialize_predicate_positions(required_delete_positions));
        RETURN_IF_ERROR(compact_predicate_columns_with_profile(false));
    }
    if (_scan_profile.predicate_filter_time == nullptr) {
        RETURN_IF_ERROR(execute_scheduled_delete_conjuncts());
    } else {
        SCOPED_TIMER(_scan_profile.predicate_filter_time);
        RETURN_IF_ERROR(execute_scheduled_delete_conjuncts());
    }
    if (*selected_rows == 0) {
        RETURN_IF_ERROR(skip_unmaterialized_predicate_columns());
        return compact_predicate_columns_with_profile(true);
    }
    RETURN_IF_ERROR(materialize_predicate_positions(_predicate_positions_scratch));
    return compact_predicate_columns_with_profile(true);
}

Status ParquetScanScheduler::prefetch_current_row_group_columns(
        ParquetFileContext& file_context,
        const std::vector<std::unique_ptr<ParquetColumnSchema>>& file_schema,
        const std::vector<format::LocalColumnIndex>& scan_columns, bool* prefetched) {
    DORIS_CHECK(prefetched != nullptr);
    if (_current_merge_range_active || *prefetched || scan_columns.empty() ||
        _current_row_group_id < 0 || file_context.native_metadata == nullptr) {
        return Status::OK();
    }
    *prefetched = true;
    // The scanner request separates predicate and non-predicate columns so Parquet can read
    // predicate columns first and lazily materialize the rest. Keep the same contract for
    // prefetch: callers decide which side to warm, and this helper only translates that selected
    // projection into physical column-chunk byte ranges for the current row group.
    const auto& metadata = file_context.native_metadata->to_thrift();
    const auto compat = native::parquet_reader_compat(
            metadata.__isset.created_by ? metadata.created_by : std::string {});
    std::vector<ParquetPageCacheRange> ranges;
    RETURN_IF_ERROR(detail::build_native_prefetch_ranges(
            metadata, file_schema, scan_columns, _current_row_group_id,
            file_context.native_file->size(), compat.parquet_816_padding, &ranges));
    file_context.prefetch_ranges(ranges, nullptr);
    return Status::OK();
}

Status ParquetScanScheduler::read_current_row_group_batch(
        ParquetFileContext& file_context,
        const std::vector<std::unique_ptr<ParquetColumnSchema>>& file_schema, int64_t batch_rows,
        const format::FileScanRequest& request, int64_t batch_first_file_row, Block* file_block,
        size_t* rows) {
    // Reader statistics are cumulative plain integers. Publishing their delta recursively for
    // every tiny batch is measurable on wide/nested scans, so flush periodically and force the
    // tail at row-group reset/close.
    Defer profile_flush {[this, batch_rows]() {
        // A widened predicate batch can be emitted in several output slices. Its lazy readers
        // have not consumed the whole physical batch until the last slice is drained.
        if (_pending_predicate_selection.empty() && finish_current_reader_batch_profiles() &&
            _scan_profile.column_reader_profile.page_crossing_batches != nullptr) {
            COUNTER_UPDATE(_scan_profile.column_reader_profile.page_crossing_batches, 1);
        }
        const bool finishes_row_group = _current_range_idx + 1 == _current_selected_ranges.size() &&
                                        _current_range_rows_read + batch_rows ==
                                                _current_selected_ranges[_current_range_idx].length;
        if (++_batches_since_profile_flush >= PROFILE_FLUSH_BATCH_INTERVAL || finishes_row_group) {
            flush_current_reader_profiles();
            _batches_since_profile_flush = 0;
        }
    }};
    if (_scan_profile.total_batches != nullptr) {
        COUNTER_UPDATE(_scan_profile.total_batches, 1);
    }
    if (_scan_profile.raw_rows_read != nullptr) {
        COUNTER_UPDATE(_scan_profile.raw_rows_read, batch_rows);
    }
    _raw_rows_read += batch_rows;
    if (_current_predicate_columns.empty() && _current_non_predicate_columns.empty()) {
        *rows = static_cast<size_t>(batch_rows);
        materialize_count_star_placeholders(request, *rows, file_block);
        if (_scan_profile.selected_rows != nullptr) {
            COUNTER_UPDATE(_scan_profile.selected_rows, batch_rows);
        }
        return Status::OK();
    }
    auto& selection = _selection;
    DORIS_CHECK(batch_rows <= std::numeric_limits<uint16_t>::max());
    uint16_t selected_rows = static_cast<uint16_t>(batch_rows);
    int64_t conjunct_filtered_rows = 0;
    bool predicate_columns_filtered = false;
    RETURN_IF_ERROR(read_filter_columns(batch_rows, request, file_block, &selection, &selected_rows,
                                        &conjunct_filtered_rows, &predicate_columns_filtered));
    _predicate_filtered_rows += conjunct_filtered_rows;
    mark_condition_cache_granules(selection, selected_rows, batch_first_file_row);

    const bool need_filter_output = selected_rows != batch_rows;
    const double batch_survival = static_cast<double>(selected_rows) / batch_rows;
    _predicate_survival_ratio = _predicate_survival_ratio < 0
                                        ? batch_survival
                                        : 0.25 * batch_survival + 0.75 * _predicate_survival_ratio;
    if (_scan_profile.selected_rows != nullptr) {
        COUNTER_UPDATE(_scan_profile.selected_rows, selected_rows);
    }
    if (_scan_profile.rows_filtered_by_conjunct != nullptr) {
        COUNTER_UPDATE(_scan_profile.rows_filtered_by_conjunct, conjunct_filtered_rows);
    }
    if (!_current_non_predicate_columns.empty() &&
        _scan_profile.lazy_read_filtered_rows != nullptr) {
        COUNTER_UPDATE(_scan_profile.lazy_read_filtered_rows, batch_rows - selected_rows);
    }
    if (selected_rows == 0 && _scan_profile.empty_selection_batches != nullptr) {
        COUNTER_UPDATE(_scan_profile.empty_selection_batches, 1);
    } else if (static_cast<int64_t>(selected_rows) == batch_rows &&
               _scan_profile.dense_batches != nullptr) {
        COUNTER_UPDATE(_scan_profile.dense_batches, 1);
    } else if (_scan_profile.selected_batches != nullptr) {
        COUNTER_UPDATE(_scan_profile.selected_batches, 1);
    }
    if (need_filter_output && !predicate_columns_filtered) {
        IColumn::Filter output_filter = selection_to_filter(selection, selected_rows, batch_rows);
        for (const auto& col : request.predicate_columns) {
            auto position_it = request.local_positions.find(col.column_id());
            DORIS_CHECK(position_it != request.local_positions.end());
            const auto block_position = position_it->second.value();
            RETURN_IF_CATCH_EXCEPTION(file_block->replace_by_position(
                    block_position, file_block->get_by_position(block_position)
                                            .column->filter(output_filter, selected_rows)));
        }
    }
    if (selected_rows == 0) {
        // Predicate readers have consumed this physical batch, but touching every lazy column here
        // turns a long rejected prefix into `empty_batches * lazy_columns` native calls. Record only
        // the positional lag. If [0, 32), [32, 64), and [64, 96) are empty, the first surviving
        // batch performs one skip(96) per lazy column. If the row group ends instead, reset drops the
        // lazy readers without flushing because no value from them can be observed.
        DORIS_CHECK(_pending_non_predicate_skip_rows <=
                    std::numeric_limits<int64_t>::max() - batch_rows);
        _pending_non_predicate_skip_rows += batch_rows;
        *rows = 0;
        return Status::OK();
    }
    if (!_current_merge_range_active && selected_rows > 0 &&
        !_current_non_predicate_columns.empty()) {
        // Do not prefetch lazy output columns until at least one row survives filtering. This is
        // the same decision point where the v2 reader switches from predicate-only reads to
        // materializing non-predicate columns, so fully filtered batches avoid unnecessary IO.
        RETURN_IF_ERROR(prefetch_current_row_group_columns(file_context, file_schema,
                                                           physical_non_predicate_columns(request),
                                                           &_current_non_predicate_prefetched));
    }

    if (selected_rows > _batch_size) {
        DORIS_CHECK(_pending_predicate_selection.empty());
        _pending_predicate_batch_rows = batch_rows;
        _pending_predicate_batch_rows_consumed = 0;
        _pending_predicate_selected_offset = 0;
        _pending_predicate_selection.resize(selected_rows);
        for (uint16_t idx = 0; idx < selected_rows; ++idx) {
            _pending_predicate_selection[idx] =
                    static_cast<SelectionVector::Index>(selection.get_index(idx));
        }
        for (const auto& col : request.predicate_columns) {
            const auto position_it = request.local_positions.find(col.column_id());
            DORIS_CHECK(position_it != request.local_positions.end());
            const size_t block_position = position_it->second.value();
            const auto& column = file_block->get_by_position(block_position).column;
            DORIS_CHECK_EQ(column->size(), selected_rows);
            _pending_predicate_columns.emplace(block_position, column);
        }
        return materialize_pending_predicate_batch(request, file_block, rows);
    }

    {
        SCOPED_TIMER(_scan_profile.column_read_time);
        // Bring lazy readers to the first row of the current physical batch before interpreting its
        // selection vector. This also merges pending range gaps with fully filtered batches.
        RETURN_IF_ERROR(flush_pending_non_predicate_skip_rows());
        for (const auto& [fid, column_reader] : _current_non_predicate_columns) {
            auto position_it = request.local_positions.find(fid);
            DORIS_CHECK(position_it != request.local_positions.end());
            const auto block_position = position_it->second.value();
            auto column = file_block->get_by_position(block_position).column->assert_mutable();
            DCHECK_EQ(file_block->get_by_position(block_position).type->get_primitive_type(),
                      column_reader->type()->get_primitive_type())
                    << type_to_string(file_block->get_by_position(block_position)
                                              .type->get_primitive_type())
                    << " " << type_to_string(column_reader->type()->get_primitive_type()) << " "
                    << column_reader->name() << " " << fid << " " << block_position;
            if (need_filter_output) {
                [[maybe_unused]] auto old_size = column->size();
                RETURN_IF_ERROR(
                        column_reader->select(selection, selected_rows, batch_rows, column));
                if (column->size() != old_size + selected_rows) {
                    return Status::Corruption(
                            "Parquet selected output column {} returned {} rows, expected {} rows",
                            column_reader->name(), column->size(), old_size + selected_rows);
                }
            } else {
                int64_t column_rows = 0;
                RETURN_IF_ERROR(column_reader->read(batch_rows, column, &column_rows));
                if (column_rows != batch_rows) {
                    return Status::Corruption(
                            "Parquet output column {} returned {} rows, expected {} rows",
                            column_reader->name(), column_rows, batch_rows);
                }
            }
            file_block->replace_by_position(block_position, std::move(column));
        }
    }
    materialize_count_star_placeholders(request, selected_rows, file_block);
    *rows = static_cast<size_t>(selected_rows);
    return Status::OK();
}

Status ParquetScanScheduler::materialize_pending_predicate_batch(
        const format::FileScanRequest& request, Block* file_block, size_t* rows) {
    DORIS_CHECK(!_pending_predicate_selection.empty());
    DORIS_CHECK(_pending_predicate_selected_offset < _pending_predicate_selection.size());
    const size_t remaining_selected =
            _pending_predicate_selection.size() - _pending_predicate_selected_offset;
    const size_t output_rows =
            std::min<size_t>(static_cast<size_t>(_batch_size), remaining_selected);
    const size_t output_end = _pending_predicate_selected_offset + output_rows;
    const int64_t physical_end =
            output_end == _pending_predicate_selection.size()
                    ? _pending_predicate_batch_rows
                    : static_cast<int64_t>(_pending_predicate_selection[output_end - 1]) + 1;
    DORIS_CHECK(physical_end > _pending_predicate_batch_rows_consumed);
    const int64_t physical_rows = physical_end - _pending_predicate_batch_rows_consumed;

    _pending_output_selection.resize(output_rows);
    for (size_t idx = 0; idx < output_rows; ++idx) {
        const int64_t physical_row =
                _pending_predicate_selection[_pending_predicate_selected_offset + idx];
        DORIS_CHECK(physical_row >= _pending_predicate_batch_rows_consumed);
        _pending_output_selection.set_index(
                idx, static_cast<SelectionVector::Index>(physical_row -
                                                         _pending_predicate_batch_rows_consumed));
    }

    for (const auto& [block_position, column] : _pending_predicate_columns) {
        file_block->replace_by_position(
                block_position, column->cut(_pending_predicate_selected_offset, output_rows));
    }
    {
        SCOPED_TIMER(_scan_profile.column_read_time);
        RETURN_IF_ERROR(flush_pending_non_predicate_skip_rows());
        for (const auto& [fid, column_reader] : _current_non_predicate_columns) {
            auto position_it = request.local_positions.find(fid);
            DORIS_CHECK(position_it != request.local_positions.end());
            const auto block_position = position_it->second.value();
            auto column = file_block->get_by_position(block_position).column->assert_mutable();
            [[maybe_unused]] const auto old_size = column->size();
            RETURN_IF_ERROR(column_reader->select(_pending_output_selection,
                                                  static_cast<uint16_t>(output_rows), physical_rows,
                                                  column));
            if (column->size() != old_size + output_rows) {
                return Status::Corruption(
                        "Parquet pending output column {} returned {} rows, expected {} rows",
                        column_reader->name(), column->size(), old_size + output_rows);
            }
            file_block->replace_by_position(block_position, std::move(column));
        }
    }
    materialize_count_star_placeholders(request, output_rows, file_block);
    *rows = output_rows;
    _pending_predicate_batch_rows_consumed = physical_end;
    _pending_predicate_selected_offset = output_end;
    if (_pending_predicate_selected_offset == _pending_predicate_selection.size()) {
        DORIS_CHECK_EQ(_pending_predicate_batch_rows_consumed, _pending_predicate_batch_rows);
        if (finish_current_reader_batch_profiles() &&
            _scan_profile.column_reader_profile.page_crossing_batches != nullptr) {
            COUNTER_UPDATE(_scan_profile.column_reader_profile.page_crossing_batches, 1);
        }
        _pending_predicate_batch_rows = 0;
        _pending_predicate_batch_rows_consumed = 0;
        _pending_predicate_selected_offset = 0;
        _pending_predicate_selection.clear();
        _pending_predicate_columns.clear();
        _pending_output_selection.clear();
    }
    return Status::OK();
}

void ParquetScanScheduler::mark_condition_cache_granules(const SelectionVector& selection,
                                                         uint16_t selected_rows,
                                                         int64_t batch_first_file_row) {
    if (!_condition_cache_ctx || _condition_cache_ctx->is_hit ||
        !_condition_cache_ctx->filter_result) {
        return;
    }
    auto& cache = *_condition_cache_ctx->filter_result;
    for (uint16_t selection_idx = 0; selection_idx < selected_rows; ++selection_idx) {
        const int64_t file_row = batch_first_file_row + selection.get_index(selection_idx);
        const int64_t granule = file_row / ConditionCacheContext::GRANULE_SIZE;
        const int64_t cache_idx = granule - _condition_cache_ctx->base_granule;
        if (cache_idx >= 0 && static_cast<size_t>(cache_idx) < cache.size()) {
            cache[static_cast<size_t>(cache_idx)] = true;
        }
    }
}

Status ParquetScanScheduler::read_next_batch(
        ParquetFileContext& file_context,
        const std::vector<std::unique_ptr<ParquetColumnSchema>>& file_schema,
        const format::FileScanRequest& request, Block* file_block, size_t* rows, bool* eof) {
    *rows = 0;
    if (!_pending_predicate_selection.empty()) {
        RETURN_IF_ERROR(materialize_pending_predicate_batch(request, file_block, rows));
        *eof = false;
        return Status::OK();
    }
    int64_t predicate_batch_rows = _batch_size;
    const int64_t max_predicate_batch_rows = std::min<int64_t>(
            std::numeric_limits<uint16_t>::max(),
            std::max<int64_t>(DEFAULT_READ_BATCH_SIZE, _runtime_state == nullptr
                                                               ? DEFAULT_READ_BATCH_SIZE
                                                               : _runtime_state->batch_size()));
    auto grow_empty_predicate_batch = [max_predicate_batch_rows](int64_t current) {
        for (const int64_t target :
             {int64_t {256}, int64_t {1024}, int64_t {4096}, max_predicate_batch_rows}) {
            if (current < target) {
                return std::min(target, max_predicate_batch_rows);
            }
        }
        return max_predicate_batch_rows;
    };
    while (true) {
        if (!_has_current_row_group) {
            bool has_row_group = false;
            RETURN_IF_ERROR(
                    open_next_row_group(file_context, file_schema, request, &has_row_group));
            if (!has_row_group) {
                *eof = true;
                return Status::OK();
            }
        }

        if (_current_range_idx >= _current_selected_ranges.size()) {
            // Current row group finished, try next row group.
            reset_current_row_group();
            continue;
        }

        const RowRange& current_range = _current_selected_ranges[_current_range_idx];
        DORIS_CHECK(current_range.start >= 0);
        DORIS_CHECK(current_range.length > 0);
        DORIS_CHECK(current_range.start + current_range.length <= _current_row_group_rows);

        if (_current_row_group_rows_read < current_range.start) {
            // Skip filtered rows according to row group level pruning.
            RETURN_IF_ERROR(skip_current_row_group_rows(current_range.start -
                                                        _current_row_group_rows_read));
        }
        DORIS_CHECK(_current_row_group_rows_read == current_range.start + _current_range_rows_read);
        const int64_t remaining_rows = current_range.length - _current_range_rows_read;
        if (remaining_rows <= 0) {
            // Current range finished, try next range in the same row group.
            ++_current_range_idx;
            _current_range_rows_read = 0;
            continue;
        }

        const int64_t batch_rows = std::min<int64_t>(predicate_batch_rows, remaining_rows);
        const int64_t physical_rows_read = batch_rows;
        const int64_t batch_first_file_row =
                _current_row_group_first_row + _current_row_group_rows_read;
        RETURN_IF_ERROR(read_current_row_group_batch(file_context, file_schema, batch_rows, request,
                                                     batch_first_file_row, file_block, rows));
        _current_row_group_rows_read += physical_rows_read;
        _current_range_rows_read += physical_rows_read;
        if (_current_range_rows_read >= current_range.length) {
            ++_current_range_idx;
            _current_range_rows_read = 0;
        }
        if (*rows == 0) {
            // Fully rejected probes carry no output-width sample. Widen predicate work to cross
            // long empty prefixes cheaply; a later non-empty probe is sliced before lazy columns
            // are materialized, so this internal width cannot escape the caller's row cap.
            predicate_batch_rows = grow_empty_predicate_batch(predicate_batch_rows);
            continue;
        }
        *eof = false;
        return Status::OK();
    }
}

} // namespace doris::format::parquet

Coverage Report

Created: 2026-07-29 15:18