dbx_core/sql/

interface.rs

//! SQL Execution Pipeline — SQL query execution methods

use crate::engine::Database;
use crate::sql::StringCaseExt;

use crate::error::{DbxError, DbxResult};
use crate::sql::executor::{
    FilterOperator, HashAggregateOperator, HashJoinOperator, LimitOperator, PhysicalOperator,
    ProjectionOperator, SortOperator, TableScanOperator,
};
use crate::sql::planner::{LogicalPlanner, PhysicalExpr, PhysicalPlan, PhysicalPlanner};
use crate::storage::columnar_cache::ColumnarCache;
use arrow::array::RecordBatch;
use arrow::datatypes::Schema;
use arrow::ipc::writer::StreamWriter;
use std::collections::HashMap;
use std::sync::Arc;

// ════════════════════════════════════════════
// Arrow IPC Serialization
// ════════════════════════════════════════════

/// Infer Arrow schema from row values
fn infer_schema_from_values(values: &[PhysicalExpr]) -> DbxResult<Schema> {
    use crate::storage::columnar::ScalarValue;
    use arrow::datatypes::{DataType, Field};

    let fields: Vec<Field> = values
        .iter()
        .enumerate()
        .map(|(i, expr)| {
            let name = format!("col_{}", i);
            let data_type = match expr {
                PhysicalExpr::Literal(scalar) => match scalar {
                    ScalarValue::Int32(_) => DataType::Int32,
                    ScalarValue::Int64(_) => DataType::Int64,
                    ScalarValue::Float64(_) => DataType::Float64,
                    ScalarValue::Utf8(_) => DataType::Utf8,
                    ScalarValue::Boolean(_) => DataType::Boolean,
                    ScalarValue::Binary(_) => DataType::Binary,
                    ScalarValue::Null => DataType::Utf8, // Default to string for nulls
                },
                _ => DataType::Utf8, // Fallback to string
            };
            Field::new(name, data_type, true) // All fields nullable
        })
        .collect();

    Ok(Schema::new(fields))
}

/// Serialize row values to Arrow IPC format
fn serialize_to_arrow_ipc(schema: &Schema, row_values: &[PhysicalExpr]) -> DbxResult<Vec<u8>> {
    use crate::storage::columnar::ScalarValue;
    use arrow::array::*;

    // Build arrays for each field
    let mut arrays: Vec<ArrayRef> = Vec::new();

    for (field, expr) in schema.fields().iter().zip(row_values) {
        let array: ArrayRef = match expr {
            PhysicalExpr::Literal(scalar) => {
                match (field.data_type(), scalar) {
                    (arrow::datatypes::DataType::Int32, ScalarValue::Int32(v)) => {
                        Arc::new(Int32Array::from(vec![*v]))
                    }
                    (arrow::datatypes::DataType::Int64, ScalarValue::Int64(v)) => {
                        Arc::new(Int64Array::from(vec![*v]))
                    }
                    // Auto-convert Int32 to Int64
                    (arrow::datatypes::DataType::Int64, ScalarValue::Int32(v)) => {
                        Arc::new(Int64Array::from(vec![*v as i64]))
                    }
                    (arrow::datatypes::DataType::Float64, ScalarValue::Float64(v)) => {
                        Arc::new(Float64Array::from(vec![*v]))
                    }
                    (arrow::datatypes::DataType::Utf8, ScalarValue::Utf8(s)) => {
                        Arc::new(StringArray::from(vec![s.as_str()]))
                    }
                    (arrow::datatypes::DataType::Boolean, ScalarValue::Boolean(b)) => {
                        Arc::new(BooleanArray::from(vec![*b]))
                    }
                    (arrow::datatypes::DataType::Binary, ScalarValue::Binary(b)) => {
                        Arc::new(BinaryArray::from(vec![b.as_slice()]))
                    }
                    (_, ScalarValue::Null) => {
                        // Create null array of appropriate type
                        match field.data_type() {
                            arrow::datatypes::DataType::Int32 => {
                                Arc::new(Int32Array::from(vec![None as Option<i32>]))
                            }
                            arrow::datatypes::DataType::Int64 => {
                                Arc::new(Int64Array::from(vec![None as Option<i64>]))
                            }
                            arrow::datatypes::DataType::Float64 => {
                                Arc::new(Float64Array::from(vec![None as Option<f64>]))
                            }
                            arrow::datatypes::DataType::Utf8 => {
                                Arc::new(StringArray::from(vec![None as Option<&str>]))
                            }
                            arrow::datatypes::DataType::Boolean => {
                                Arc::new(BooleanArray::from(vec![None as Option<bool>]))
                            }
                            arrow::datatypes::DataType::Binary => {
                                Arc::new(BinaryArray::from(vec![None as Option<&[u8]>]))
                            }
                            _ => {
                                return Err(DbxError::NotImplemented(format!(
                                    "Unsupported null type: {:?}",
                                    field.data_type()
                                )));
                            }
                        }
                    }
                    _ => {
                        return Err(DbxError::NotImplemented(format!(
                            "Type mismatch: field {:?} vs scalar {:?}",
                            field.data_type(),
                            scalar
                        )));
                    }
                }
            }
            _ => {
                return Err(DbxError::NotImplemented(
                    "Non-literal value in INSERT".to_string(),
                ));
            }
        };
        arrays.push(array);
    }

    // Create RecordBatch
    let batch = RecordBatch::try_new(Arc::new(schema.clone()), arrays)
        .map_err(|e| DbxError::Storage(e.to_string()))?;

    // Serialize to Arrow IPC Stream format
    let mut buffer = Vec::new();
    {
        let mut writer = StreamWriter::try_new(&mut buffer, schema)
            .map_err(|e| DbxError::Serialization(e.to_string()))?;
        writer
            .write(&batch)
            .map_err(|e| DbxError::Serialization(e.to_string()))?;
        writer
            .finish()
            .map_err(|e| DbxError::Serialization(e.to_string()))?;
    }

    Ok(buffer)
}

// ════════════════════════════════════════════
// Helper Functions for WHERE Clause Evaluation
// ════════════════════════════════════════════

/// Reconstruct a full record by prepending the key (col 0) to the stored value JSON.
/// INSERT stores `row_values[1..]` as JSON, excluding the key column.
/// This function re-adds the key as the first element so that schema column indices
/// align correctly with the data during filter evaluation.
fn reconstruct_full_record(key: &[u8], value_bytes: &[u8]) -> Vec<u8> {
    // Try to interpret the key as an integer (le_bytes from INSERT)
    let key_json = if key.len() == 4 {
        let val = i32::from_le_bytes(key.try_into().unwrap_or([0; 4]));
        serde_json::Value::Number(val.into())
    } else if key.len() == 8 {
        let val = i64::from_le_bytes(key.try_into().unwrap_or([0; 8]));
        serde_json::Value::Number(val.into())
    } else {
        // Treat as UTF-8 string key
        let s = String::from_utf8_lossy(key).to_string();
        serde_json::Value::String(s)
    };

    // Parse existing values and prepend key
    let mut values: Vec<serde_json::Value> =
        serde_json::from_slice(value_bytes).unwrap_or_else(|_| vec![]);
    values.insert(0, key_json);
    serde_json::to_vec(&values).unwrap_or_else(|_| value_bytes.to_vec())
}

/// Convert a single JSON record to a RecordBatch for filter evaluation
pub fn json_record_to_batch(value_bytes: &[u8]) -> DbxResult<RecordBatch> {
    use arrow::array::{ArrayRef, BooleanArray, Float64Array, Int64Array, StringArray};
    use arrow::datatypes::{DataType, Field, Schema};

    let current_values: Vec<serde_json::Value> =
        serde_json::from_slice(value_bytes).unwrap_or_else(|_| vec![]);

    let mut fields = Vec::new();
    let mut columns: Vec<ArrayRef> = Vec::new();

    for (i, val) in current_values.iter().enumerate() {
        match val {
            serde_json::Value::Number(n) => {
                if n.is_i64() {
                    fields.push(Field::new(format!("col_{}", i), DataType::Int64, true));
                    columns.push(Arc::new(Int64Array::from(vec![n.as_i64()])));
                } else if n.is_f64() {
                    fields.push(Field::new(format!("col_{}", i), DataType::Float64, true));
                    columns.push(Arc::new(Float64Array::from(vec![n.as_f64()])));
                } else {
                    fields.push(Field::new(format!("col_{}", i), DataType::Int64, true));
                    columns.push(Arc::new(Int64Array::from(vec![n.as_i64()])));
                }
            }
            serde_json::Value::String(s) => {
                fields.push(Field::new(format!("col_{}", i), DataType::Utf8, true));
                columns.push(Arc::new(StringArray::from(vec![Some(s.as_str())])));
            }
            serde_json::Value::Bool(b) => {
                fields.push(Field::new(format!("col_{}", i), DataType::Boolean, true));
                columns.push(Arc::new(BooleanArray::from(vec![Some(*b)])));
            }
            serde_json::Value::Null => {
                fields.push(Field::new(format!("col_{}", i), DataType::Int64, true));
                columns.push(Arc::new(Int64Array::from(vec![None::<i64>])));
            }
            _ => {
                fields.push(Field::new(format!("col_{}", i), DataType::Int64, true));
                columns.push(Arc::new(Int64Array::from(vec![None::<i64>])));
            }
        }
    }

    let schema = Arc::new(Schema::new(fields));
    RecordBatch::try_new(schema, columns).map_err(DbxError::from)
}

/// Evaluate a filter expression for a single record
/// Returns true if the record matches the filter (or if no filter is provided)
fn evaluate_filter_for_record(
    filter_expr: Option<&PhysicalExpr>,
    value_bytes: &[u8],
) -> DbxResult<bool> {
    if let Some(expr) = filter_expr {
        let batch = json_record_to_batch(value_bytes)?;

        use crate::sql::executor::evaluate_expr;
        let result = evaluate_expr(expr, &batch)?;

        use arrow::array::BooleanArray;
        let bool_array = result
            .as_any()
            .downcast_ref::<BooleanArray>()
            .ok_or_else(|| DbxError::TypeMismatch {
                expected: "BooleanArray".to_string(),
                actual: format!("{:?}", result.data_type()),
            })?;

        Ok(bool_array.value(0))
    } else {
        Ok(true) // No filter, all records match
    }
}

impl Database {
    // ════════════════════════════════════════════
    // Phase 3: 파티셔닝 (Partition Pruning Helper)
    // ════════════════════════════════════════════

    /// 파티셔닝 라우팅 리스트를 가져옵니다.
    /// 조건이 없으면 모든 파티션을 스캔, 조건이 있으면 (향후) 필요한 파티션만 스캔합니다.
    pub(crate) fn get_tables_to_scan(
        &self,
        table: &str,
        _filter: Option<&PhysicalExpr>,
    ) -> Vec<String> {
        let maps = self.partition_maps.read().unwrap();
        if let Some(map) = maps.get(table) {
            // MVP: 조건절을 분석하여 특정 파티션만 추출하는 프루닝 로직 대신 전체 파티션을 반환.
            // TODO: Extract PartitionValue from _filter and prune.
            map.all_partitions()
        } else {
            // 파티션 매핑이 없으면 셔딩 라우터에 쓰인 모든 샤드 가상 테이블을 조회
            // ════════════════════════════════════════════
            // Phase 6: Sharding Scatter Read (Scatter-Gather)
            // ════════════════════════════════════════════
            self.sharding_router
                .all_shards()
                .iter()
                .map(|shard| format!("{}__shard_{}", table, shard.id))
                .collect()
        }
    }

    // ════════════════════════════════════════════
    // SQL Execution Pipeline
    // ════════════════════════════════════════════

    /// Register table data for SQL queries.
    ///
    /// Tables registered here can be queried via `execute_sql()`.
    pub fn register_table(&self, name: &str, batches: Vec<RecordBatch>) {
        // Store schema from first batch
        if let Some(first_batch) = batches.first() {
            let schema = first_batch.schema();
            let mut schemas = self.table_schemas.write().unwrap();
            schemas.insert(name.to_string(), schema);
        }

        // Store batches
        let mut tables = self.tables.write().unwrap();
        tables.insert(name.to_string(), batches);
    }

    /// Append a RecordBatch to an existing registered table.
    pub fn append_batch(&self, table: &str, batch: RecordBatch) {
        let mut tables = self.tables.write().unwrap();
        tables.entry(table.to_string()).or_default().push(batch);
    }

    /// Execute a SQL query and return RecordBatch results.
    ///
    /// Full pipeline: Parse → LogicalPlan → Optimize → PhysicalPlan → Execute
    ///
    /// # Example
    ///
    /// ```rust
    /// # use dbx_core::Database;
    /// # fn main() -> dbx_core::DbxResult<()> {
    /// let db = Database::open_in_memory()?;
    /// // Register table data first, then:
    /// // let batches = db.execute_sql("SELECT * FROM users WHERE age > 18")?;
    /// # Ok(())
    /// # }
    /// ```
    pub fn execute_sql(&self, sql: &str) -> DbxResult<Vec<RecordBatch>> {
        let t0 = std::time::Instant::now();

        // Step 0: Intercept CREATE VIEW / DROP VIEW (before parser)
        let sql_trimmed = sql.trim();

        if sql_trimmed.starts_with_ignore_ascii_case("CREATE VIEW") {
            return self.handle_create_view(sql_trimmed);
        }
        if sql_trimmed.starts_with_ignore_ascii_case("DROP VIEW") {
            return self.handle_drop_view(sql_trimmed);
        }

        // Materialized View 인터셀트 (CREATE/DROP/REFRESH)
        if sql_trimmed.starts_with_ignore_ascii_case("CREATE MATERIALIZED VIEW") {
            return self.handle_create_materialized_view(sql_trimmed);
        }
        if sql_trimmed.starts_with_ignore_ascii_case("DROP MATERIALIZED VIEW") {
            return self.handle_drop_materialized_view(sql_trimmed);
        }
        if sql_trimmed.starts_with_ignore_ascii_case("REFRESH MATERIALIZED VIEW") {
            return self.handle_refresh_materialized_view(sql_trimmed);
        }

        // SELECT 시 Materialized View 캐시 히트 확인
        if sql_trimmed.starts_with_ignore_ascii_case("SELECT")
            && let Some(cached) = self.try_matview_cache(sql_trimmed)
        {
            return Ok(cached);
        }

        // Step 0b: Expand views in FROM clauses
        let expanded = self.view_registry.expand(sql_trimmed);
        let sql = expanded.as_str();

        // Step 1: Parse SQL → AST
        let statements = self.sql_parser.parse(sql)?;
        if statements.is_empty() {
            return Ok(vec![]);
        }

        // Step 2: Logical Plan
        let planner = LogicalPlanner::new();
        let logical_plan = planner.plan(&statements[0])?;

        // Step 3: Optimize
        let optimized = self.sql_optimizer.optimize(logical_plan)?;

        // Step 4: Physical Plan
        let physical_planner = PhysicalPlanner::new(Arc::clone(&self.table_schemas));
        let physical_plan = physical_planner.plan(&optimized)?;

        // Step 5: Execute
        let result = self.execute_physical_plan(&physical_plan);

        // ── Metrics ─────────────────────────────────────────────────
        let elapsed_us = t0.elapsed().as_micros() as u64;
        self.metrics.inc_sql_queries();
        self.metrics.query_latency_us.observe(elapsed_us);

        result
    }

    /// CREATE VIEW handler
    fn handle_create_view(&self, sql: &str) -> DbxResult<Vec<RecordBatch>> {
        // Syntax: CREATE VIEW <name> AS <select_sql>
        let upper = sql.to_uppercase();
        let after_view = &sql[upper.find("VIEW").unwrap() + 4..]
            .trim_start()
            .to_owned();
        let as_pos = after_view
            .to_uppercase()
            .find(" AS ")
            .ok_or_else(|| DbxError::SqlParse {
                message: "CREATE VIEW requires AS".to_string(),
                sql: sql.to_string(),
            })?;
        let view_name = after_view[..as_pos].trim();
        let view_sql = after_view[as_pos + 4..].trim();
        self.view_registry.create(view_name, view_sql)?;
        self.one_row_affected_batch()
    }

    /// DROP VIEW handler
    fn handle_drop_view(&self, sql: &str) -> DbxResult<Vec<RecordBatch>> {
        // Syntax: DROP VIEW [IF EXISTS] <name>
        let upper = sql.to_uppercase();
        let after_view = sql[upper.find("VIEW").unwrap() + 4..]
            .trim_start()
            .to_owned();
        let (if_exists, name_part) = if after_view.starts_with_ignore_ascii_case("IF EXISTS") {
            (true, after_view[9..].trim_start().to_string())
        } else {
            (false, after_view.clone())
        };
        let view_name = name_part.trim();

        if if_exists && !self.view_registry.exists(view_name) {
            return self.one_row_affected_batch();
        }

        self.view_registry.drop(view_name)?;
        self.one_row_affected_batch()
    }

    // ────────────────────────────────────────────────────────────────────────
    // Materialized View Handlers
    // ────────────────────────────────────────────────────────────────────────

    /// CREATE MATERIALIZED VIEW <name> [REFRESH EVERY <secs>] AS <select_sql>
    fn handle_create_materialized_view(&self, sql: &str) -> DbxResult<Vec<RecordBatch>> {
        let upper = sql.to_uppercase();

        // 뷰 이름 / SQL 추출 (VIEW 키워드 이후)
        let after_view = sql[upper.find("VIEW").unwrap() + 4..]
            .trim_start()
            .to_owned();
        let after_upper = after_view.to_uppercase();

        // "REFRESH EVERY <n> AS ..." 패턴 감지 (선택적)
        let (refresh_interval_secs, body) = if let Some(re_pos) = after_upper.find("REFRESH EVERY")
        {
            // 이름은 REFRESH EVERY 이전
            let name_part = after_view[..re_pos].trim().to_string();
            let rest = after_view[re_pos + 13..].trim_start().to_string();
            let as_pos = rest
                .to_uppercase()
                .find(" AS ")
                .ok_or_else(|| DbxError::SqlParse {
                    message: "CREATE MATERIALIZED VIEW ... REFRESH EVERY <n> AS <sql>".to_string(),
                    sql: sql.to_string(),
                })?;
            let interval_str = rest[..as_pos].trim();
            let interval_secs: u64 = interval_str.parse().map_err(|_| DbxError::SqlParse {
                message: format!(
                    "REFRESH EVERY requires integer seconds, got '{}'",
                    interval_str
                ),
                sql: sql.to_string(),
            })?;
            let view_sql = rest[as_pos + 4..].trim().to_string();
            (Some(interval_secs), (name_part, view_sql))
        } else {
            let as_pos = after_upper.find(" AS ").ok_or_else(|| DbxError::SqlParse {
                message: "CREATE MATERIALIZED VIEW requires AS".to_string(),
                sql: sql.to_string(),
            })?;
            let name = after_view[..as_pos].trim().to_string();
            let view_sql = after_view[as_pos + 4..].trim().to_string();
            (None, (name, view_sql))
        };

        self.mat_view_registry
            .create(&body.0, &body.1, refresh_interval_secs)?;
        self.one_row_affected_batch()
    }

    /// DROP MATERIALIZED VIEW <name>
    fn handle_drop_materialized_view(&self, sql: &str) -> DbxResult<Vec<RecordBatch>> {
        let upper = sql.to_uppercase();
        let name = sql[upper.find("VIEW").unwrap() + 4..].trim();
        self.mat_view_registry.remove(name)?;
        self.one_row_affected_batch()
    }

    /// REFRESH MATERIALIZED VIEW <name>
    fn handle_refresh_materialized_view(&self, sql: &str) -> DbxResult<Vec<RecordBatch>> {
        let upper = sql.to_uppercase();
        let name = sql[upper.find("VIEW").unwrap() + 4..].trim().to_lowercase();
        let view_sql = self.mat_view_registry.get_sql(&name).ok_or_else(|| {
            DbxError::InvalidArguments(format!("'{}' 구체화된 뷰를 찾을 수 없음", name))
        })?;
        // 뷰 쿼리 실행 후 캐시 저장
        let batches = self.execute_sql(&view_sql)?;
        self.mat_view_registry.set_cache(&name, batches)?;
        self.one_row_affected_batch()
    }

    /// SELECT 쿼리에서 Materialized View 캐시 히트 확인
    ///
    /// "SELECT ... FROM <mv_name>" 패턴에서 FROM 다음 토큰이 등록된 MV이고
    /// 캐시가 fresh하면 캐시를 반환합니다.
    fn try_matview_cache(&self, sql: &str) -> Option<Vec<RecordBatch>> {
        let upper = sql.to_uppercase();
        let from_pos = upper.find(" FROM ")?;
        let after_from = sql[from_pos + 6..].trim();
        // 첫 번째 토큰 (공백/세미콜론/닫는 괄호 기준)
        let name = after_from
            .split(|c: char| c.is_whitespace() || c == ';' || c == ')')
            .next()?;
        if self.mat_view_registry.is_fresh(name) {
            self.mat_view_registry.get_cache(name)
        } else {
            None
        }
    }

    /// Helper: single-row `rows_affected = 1` result batch
    fn one_row_affected_batch(&self) -> DbxResult<Vec<RecordBatch>> {
        use arrow::array::Int64Array;
        use arrow::datatypes::{DataType, Field, Schema};
        let schema = Arc::new(Schema::new(vec![Field::new(
            "rows_affected",
            DataType::Int64,
            false,
        )]));
        let array = Int64Array::from(vec![1_i64]);
        let batch = RecordBatch::try_new(schema, vec![Arc::new(array)]).map_err(DbxError::from)?;
        Ok(vec![batch])
    }

    /// Execute a physical plan against registered table data.
    fn execute_physical_plan(&self, plan: &PhysicalPlan) -> DbxResult<Vec<RecordBatch>> {
        match plan {
            PhysicalPlan::Insert {
                table,
                columns: _,
                values,
            } => {
                // Track OLTP Workload
                self.workload_analyzer
                    .write()
                    .unwrap()
                    .record(crate::engine::workload_analyzer::QueryPattern::PointQuery);

                // Execute INSERT: convert PhysicalExpr values to bytes and insert into Delta Store
                let mut rows_inserted = 0;

                for row_values in values {
                    // For simplicity, use first column as key, rest as value
                    // TODO: Proper schema-based serialization
                    if row_values.is_empty() {
                        continue;
                    }

                    // Extract key from first value
                    let key = match &row_values[0] {
                        PhysicalExpr::Literal(scalar) => {
                            use crate::storage::columnar::ScalarValue;
                            match scalar {
                                ScalarValue::Utf8(s) => s.as_bytes().to_vec(),
                                ScalarValue::Int32(i) => i.to_le_bytes().to_vec(),
                                ScalarValue::Int64(i) => i.to_le_bytes().to_vec(),
                                ScalarValue::Float64(f) => f.to_le_bytes().to_vec(),
                                ScalarValue::Boolean(b) => vec![if *b { 1 } else { 0 }],
                                // The following cases are not expected here, but are added as per instruction
                                // They would typically be handled in a `build_operator` method or similar
                                // where DDL/DML plans are distinguished from query plans.
                                // For now, we'll place them here as a placeholder for the instruction.
                                _ => {
                                    return Err(DbxError::NotImplemented(
                                        "Non-literal key in INSERT".to_string(),
                                    ));
                                }
                            }
                        }
                        _ => {
                            return Err(DbxError::NotImplemented(
                                "Non-literal key in INSERT".to_string(),
                            ));
                        }
                    };

                    // Get or infer schema for Arrow IPC serialization
                    let schema = {
                        let schemas = self.table_schemas.read().unwrap();
                        schemas.get(table.as_str()).cloned()
                    }
                    .unwrap_or_else(|| {
                        // No registered schema: infer from row values
                        Arc::new(
                            infer_schema_from_values(row_values)
                                .expect("Failed to infer schema from values"),
                        )
                    });

                    // Always use Arrow IPC serialization
                    let value_bytes = serialize_to_arrow_ipc(&schema, row_values)?;

                    // ════════════════════════════════════════════
                    // Phase 6: Sharding Scatter Write
                    // ════════════════════════════════════════════
                    self.scatter_gather.scatter_write(
                        &key,
                        &value_bytes,
                        |shard_id, sub_key, sub_val| {
                            let shard_table = format!("{}__shard_{}", table, shard_id);
                            // Write to the specific shard virtual table (which will then be partitioned by crud.rs if PartitionMap exists)
                            let _ = self.insert(&shard_table, sub_key, sub_val);
                        },
                    );

                    rows_inserted += 1;
                }

                // Return result batch indicating success
                use arrow::array::{Int64Array, RecordBatch};
                use arrow::datatypes::{DataType, Field, Schema};

                let schema = Arc::new(Schema::new(vec![Field::new(
                    "rows_inserted",
                    DataType::Int64,
                    false,
                )]));
                let array = Int64Array::from(vec![rows_inserted as i64]);
                let batch =
                    RecordBatch::try_new(schema, vec![Arc::new(array)]).map_err(DbxError::from)?;

                Ok(vec![batch])
            }
            PhysicalPlan::Update {
                table,
                assignments,
                filter,
            } => {
                // Track OLTP Workload
                self.workload_analyzer
                    .write()
                    .unwrap()
                    .record(crate::engine::workload_analyzer::QueryPattern::PointQuery);

                // Execute UPDATE: scan (across partitions), evaluate filter, update matching records
                let target_tables = self.get_tables_to_scan(table, filter.as_ref());
                let mut all_records = Vec::new();
                for target_table in &target_tables {
                    all_records.extend(self.scan(target_table)?);
                }

                let mut rows_updated = 0_i64;

                // Build column name → index mapping from schema
                let column_index_map = {
                    let schemas = self.table_schemas.read().unwrap();
                    schemas.get(table.as_str()).map(|schema| {
                        schema
                            .fields()
                            .iter()
                            .enumerate()
                            .map(|(i, field)| (field.name().clone(), i))
                            .collect::<std::collections::HashMap<String, usize>>()
                    })
                };

                for (key, value_bytes) in all_records {
                    // Reconstruct full record: prepend key (col 0) to value
                    // INSERT stores row_values[1..] as JSON, so key must be re-added
                    let full_record = reconstruct_full_record(&key, &value_bytes);
                    let should_update = evaluate_filter_for_record(filter.as_ref(), &full_record)?;

                    if should_update {
                        let mut current_values: Vec<serde_json::Value> =
                            serde_json::from_slice(&full_record).unwrap_or_else(|_| vec![]);

                        // Apply assignments using schema-based column mapping
                        for (column_name, expr) in assignments.iter() {
                            let target_idx = column_index_map
                                .as_ref()
                                .and_then(|map| map.get(column_name).copied())
                                .unwrap_or_else(|| {
                                    // Fallback: linear search by position
                                    assignments
                                        .iter()
                                        .position(|(n, _)| n == column_name)
                                        .unwrap_or(0)
                                });

                            if let PhysicalExpr::Literal(scalar) = expr {
                                use crate::storage::columnar::ScalarValue;
                                let new_value = match scalar {
                                    ScalarValue::Utf8(s) => serde_json::Value::String(s.clone()),
                                    ScalarValue::Int32(v) => serde_json::Value::Number((*v).into()),
                                    ScalarValue::Int64(v) => serde_json::Value::Number((*v).into()),
                                    ScalarValue::Float64(f) => serde_json::Number::from_f64(*f)
                                        .map(serde_json::Value::Number)
                                        .unwrap_or(serde_json::Value::Null),
                                    ScalarValue::Boolean(b) => serde_json::Value::Bool(*b),
                                    ScalarValue::Binary(b) => {
                                        // Encode binary as base64 string
                                        serde_json::Value::String(base64::Engine::encode(
                                            &base64::engine::general_purpose::STANDARD,
                                            b,
                                        ))
                                    }
                                    ScalarValue::Null => serde_json::Value::Null,
                                };

                                if target_idx < current_values.len() {
                                    current_values[target_idx] = new_value;
                                }
                            }
                        }

                        // Remove key (col 0) before serializing back to storage
                        let storage_values = if current_values.len() > 1 {
                            &current_values[1..]
                        } else {
                            &current_values[..]
                        };
                        let new_value_bytes = serde_json::to_vec(storage_values)
                            .map_err(|e| DbxError::Serialization(e.to_string()))?;

                        self.insert(table, &key, &new_value_bytes)?;
                        rows_updated += 1;
                    }
                }

                // Return result batch
                use arrow::array::{Int64Array, RecordBatch};
                use arrow::datatypes::{DataType, Field, Schema};

                let schema = Arc::new(Schema::new(vec![Field::new(
                    "rows_updated",
                    DataType::Int64,
                    false,
                )]));
                let array = Int64Array::from(vec![rows_updated]);
                let batch =
                    RecordBatch::try_new(schema, vec![Arc::new(array)]).map_err(DbxError::from)?;

                Ok(vec![batch])
            }
            PhysicalPlan::Delete { table, filter } => {
                // Track OLTP Workload
                self.workload_analyzer
                    .write()
                    .unwrap()
                    .record(crate::engine::workload_analyzer::QueryPattern::PointQuery);

                // Execute DELETE: scan table (across partitions), evaluate filter, delete matching records

                // Step 1: Scan all records from the target partition(s)
                let target_tables = self.get_tables_to_scan(table, filter.as_ref());
                let mut all_records = Vec::new();
                for target_table in &target_tables {
                    all_records.extend(self.scan(target_table)?);
                }

                let mut rows_deleted = 0_i64;

                // Step 2: Process each record
                for (key, value_bytes) in all_records {
                    // Reconstruct full record: prepend key (col 0) to value
                    let full_record = reconstruct_full_record(&key, &value_bytes);
                    let should_delete = evaluate_filter_for_record(filter.as_ref(), &full_record)?;

                    if should_delete {
                        // Delete from Delta Store
                        self.delete(table, &key)?;
                        rows_deleted += 1;
                    }
                }

                // Return result batch
                use arrow::array::{Int64Array, RecordBatch};
                use arrow::datatypes::{DataType, Field, Schema};

                let schema = Arc::new(Schema::new(vec![Field::new(
                    "rows_deleted",
                    DataType::Int64,
                    false,
                )]));
                let array = Int64Array::from(vec![rows_deleted]);
                let batch =
                    RecordBatch::try_new(schema, vec![Arc::new(array)]).map_err(DbxError::from)?;

                Ok(vec![batch])
            }
            PhysicalPlan::DropTable { table, if_exists } => {
                // DROP TABLE implementation
                use arrow::array::{Int64Array, RecordBatch};
                use arrow::datatypes::{DataType, Field, Schema};

                // Check if table exists
                let exists = self.table_schemas.read().unwrap().contains_key(table);

                if !exists && !if_exists {
                    return Err(DbxError::TableNotFound(table.clone()));
                }

                if exists {
                    // Remove table schema from memory
                    self.table_schemas.write().unwrap().remove(table);

                    // Delete schema from persistent storage
                    self.wos_for_metadata().delete_schema_metadata(table)?;

                    // Note: Data deletion from Delta Store/WOS would require
                    // scanning and deleting all keys with table prefix
                    // For now, we just remove the schema metadata
                }

                // Return success
                let schema = Arc::new(Schema::new(vec![Field::new(
                    "rows_affected",
                    DataType::Int64,
                    false,
                )]));
                let array = Int64Array::from(vec![1]);
                let batch =
                    RecordBatch::try_new(schema, vec![Arc::new(array)]).map_err(DbxError::from)?;

                Ok(vec![batch])
            }
            PhysicalPlan::CreateTable {
                table,
                columns,
                if_not_exists,
                policy,
            } => {
                // CREATE TABLE implementation
                use arrow::array::{Int64Array, RecordBatch};
                use arrow::datatypes::{DataType, Field, Schema};

                // Check if table already exists
                let exists = self.table_schemas.read().unwrap().contains_key(table);

                if exists && !if_not_exists {
                    return Err(DbxError::Schema(format!(
                        "Table '{}' already exists",
                        table
                    )));
                }

                if !exists {
                    // Create Arrow schema from column definitions
                    let fields: Vec<Field> = columns
                        .iter()
                        .map(|(name, type_str)| {
                            let data_type = match type_str.to_uppercase().as_str() {
                                "INT" | "INTEGER" => DataType::Int64,
                                "TEXT" | "STRING" | "VARCHAR" => DataType::Utf8,
                                "FLOAT" | "DOUBLE" => DataType::Float64,
                                "BOOL" | "BOOLEAN" => DataType::Boolean,
                                _ => DataType::Utf8, // Default to string
                            };
                            Field::new(name, data_type, true)
                        })
                        .collect();

                    let mut schema = Schema::new(fields);

                    // IF Policy exists, embed as JSON into Arrow schema metadata
                    if let Some(p) = policy
                        && let Ok(json) = p.to_json()
                    {
                        let mut map = std::collections::HashMap::new();
                        map.insert("dbx_table_policy".to_string(), json);
                        schema = schema.with_metadata(map);
                    }

                    let schema = Arc::new(schema);

                    // Store schema in memory
                    self.table_schemas
                        .write()
                        .unwrap()
                        .insert(table.clone(), schema.clone());

                    // Persist schema to storage
                    self.wos_for_metadata()
                        .save_schema_metadata(table, &schema)?;
                }

                // Return success
                let schema = Arc::new(Schema::new(vec![Field::new(
                    "rows_affected",
                    DataType::Int64,
                    false,
                )]));
                let array = Int64Array::from(vec![1]);
                let batch =
                    RecordBatch::try_new(schema, vec![Arc::new(array)]).map_err(DbxError::from)?;

                Ok(vec![batch])
            }
            PhysicalPlan::CreateIndex {
                table,
                index_name,
                columns,
                if_not_exists,
            } => {
                use arrow::array::{Int64Array, RecordBatch};
                use arrow::datatypes::{DataType, Field, Schema};

                // Validate: target table must exist
                {
                    let schemas = self.table_schemas.read().unwrap();
                    if !schemas.contains_key(table.as_str()) {
                        return Err(DbxError::Schema(format!(
                            "Table '{}' does not exist",
                            table
                        )));
                    }
                }

                let column = columns.first().ok_or_else(|| {
                    DbxError::Schema("CREATE INDEX requires at least one column".to_string())
                })?;

                let exists = self.index.has_index(table, column);

                if exists && !if_not_exists {
                    return Err(DbxError::IndexAlreadyExists {
                        table: table.clone(),
                        column: column.clone(),
                    });
                }

                if !exists {
                    self.index.create_index(table, column)?;

                    // Register index_name → (table, column) mapping in memory
                    self.index_registry
                        .write()
                        .unwrap()
                        .insert(index_name.clone(), (table.clone(), column.clone()));

                    // Persist index metadata to storage
                    self.wos_for_metadata()
                        .save_index_metadata(index_name, table, column)?;
                }

                let schema = Arc::new(Schema::new(vec![Field::new(
                    "rows_affected",
                    DataType::Int64,
                    false,
                )]));
                let array = Int64Array::from(vec![1]);
                let batch =
                    RecordBatch::try_new(schema, vec![Arc::new(array)]).map_err(DbxError::from)?;

                Ok(vec![batch])
            }
            PhysicalPlan::DropIndex {
                table,
                index_name,
                if_exists,
            } => {
                use arrow::array::{Int64Array, RecordBatch};
                use arrow::datatypes::{DataType, Field, Schema};

                // Resolve index_name → actual column via registry
                let resolved_column = {
                    let registry = self.index_registry.read().unwrap();
                    registry
                        .get(index_name.as_str())
                        .map(|(_, col)| col.clone())
                };

                // Fallback: if not in registry, try index_name as column name
                let column = resolved_column.as_deref().unwrap_or(index_name.as_str());

                let exists = self.index.has_index(table, column);

                if !exists && !if_exists {
                    return Err(DbxError::IndexNotFound {
                        table: table.clone(),
                        column: column.to_string(),
                    });
                }

                if exists {
                    self.index.drop_index(table, column)?;

                    // Remove from registry in memory
                    self.index_registry
                        .write()
                        .unwrap()
                        .remove(index_name.as_str());

                    // Delete index metadata from storage
                    self.wos_for_metadata().delete_index_metadata(index_name)?;
                }

                let schema = Arc::new(Schema::new(vec![Field::new(
                    "rows_affected",
                    DataType::Int64,
                    false,
                )]));
                let array = Int64Array::from(vec![1]);
                let batch =
                    RecordBatch::try_new(schema, vec![Arc::new(array)]).map_err(DbxError::from)?;

                Ok(vec![batch])
            }
            PhysicalPlan::AlterTable { table, operation } => {
                // ALTER TABLE implementation
                use crate::sql::planner::types::AlterTableOperation;
                use arrow::array::{Int64Array, RecordBatch};
                use arrow::datatypes::{DataType, Field, Schema};

                match operation {
                    AlterTableOperation::AddColumn {
                        column_name,
                        data_type,
                    } => {
                        // Get current schema
                        let mut schemas = self.table_schemas.write().unwrap();
                        let current_schema = schemas.get(table).ok_or_else(|| {
                            DbxError::Schema(format!("Table '{}' not found", table))
                        })?;

                        // Convert data type string to Arrow DataType
                        let arrow_type = match data_type.to_uppercase().as_str() {
                            "INT" | "INTEGER" => DataType::Int64,
                            "TEXT" | "VARCHAR" | "STRING" => DataType::Utf8,
                            "FLOAT" | "DOUBLE" | "REAL" => DataType::Float64,
                            "BOOL" | "BOOLEAN" => DataType::Boolean,
                            _ => DataType::Utf8, // Default to string
                        };

                        // Create new field
                        let new_field = Field::new(column_name, arrow_type, true);

                        // Create new schema with added column
                        let mut fields: Vec<Field> = current_schema
                            .fields()
                            .iter()
                            .map(|f| f.as_ref().clone())
                            .collect();
                        fields.push(new_field);
                        let new_schema = Arc::new(Schema::new(fields));

                        // Update schema in memory
                        schemas.insert(table.clone(), new_schema.clone());

                        // Persist updated schema
                        drop(schemas); // Release lock before calling wos
                        self.wos_for_metadata()
                            .save_schema_metadata(table, &new_schema)?;
                    }
                    AlterTableOperation::DropColumn { column_name } => {
                        // Get current schema
                        let mut schemas = self.table_schemas.write().unwrap();
                        let current_schema = schemas.get(table).ok_or_else(|| {
                            DbxError::Schema(format!("Table '{}' not found", table))
                        })?;

                        // Find the column to drop
                        let fields: Vec<Field> = current_schema
                            .fields()
                            .iter()
                            .filter(|f| f.name() != column_name)
                            .map(|f| f.as_ref().clone())
                            .collect();

                        // Check if column was found
                        if fields.len() == current_schema.fields().len() {
                            return Err(DbxError::Schema(format!(
                                "Column '{}' not found in table '{}'",
                                column_name, table
                            )));
                        }

                        // Create new schema without the dropped column
                        let new_schema = Arc::new(Schema::new(fields));

                        // Update schema in memory
                        schemas.insert(table.clone(), new_schema.clone());

                        // Persist updated schema
                        drop(schemas); // Release lock
                        self.wos_for_metadata()
                            .save_schema_metadata(table, &new_schema)?;
                    }
                    AlterTableOperation::RenameColumn { old_name, new_name } => {
                        // Get current schema
                        let mut schemas = self.table_schemas.write().unwrap();
                        let current_schema = schemas.get(table).ok_or_else(|| {
                            DbxError::Schema(format!("Table '{}' not found", table))
                        })?;

                        // Find and rename the column
                        let mut found = false;
                        let fields: Vec<Field> = current_schema
                            .fields()
                            .iter()
                            .map(|f| {
                                if f.name() == old_name {
                                    found = true;
                                    Field::new(new_name, f.data_type().clone(), f.is_nullable())
                                } else {
                                    f.as_ref().clone()
                                }
                            })
                            .collect();

                        // Check if column was found
                        if !found {
                            return Err(DbxError::Schema(format!(
                                "Column '{}' not found in table '{}'",
                                old_name, table
                            )));
                        }

                        // Create new schema with renamed column
                        let new_schema = Arc::new(Schema::new(fields));

                        // Update schema in memory
                        schemas.insert(table.clone(), new_schema.clone());

                        // Persist updated schema
                        drop(schemas); // Release lock
                        self.wos_for_metadata()
                            .save_schema_metadata(table, &new_schema)?;
                    }
                }

                // Return success
                let schema = Arc::new(Schema::new(vec![Field::new(
                    "rows_affected",
                    DataType::Int64,
                    false,
                )]));
                let array = Int64Array::from(vec![1]);
                let batch =
                    RecordBatch::try_new(schema, vec![Arc::new(array)]).map_err(DbxError::from)?;

                Ok(vec![batch])
            }
            _ => {
                // Original logic for SELECT queries
                self.execute_select_plan(plan)
            }
        }
    }

    /// Execute SELECT query plans (original logic)
    fn execute_select_plan(&self, plan: &PhysicalPlan) -> DbxResult<Vec<RecordBatch>> {
        // Automatic Tier Selection & Loading
        // Ensure all tables involved in the query are synced to Columnar Cache (Tier 2).
        for table in plan.tables() {
            let target_tables = self.get_tables_to_scan(&table, None);
            for t in target_tables {
                if !self.columnar_cache.has_table(&t) {
                    let _ = self.sync_columnar_cache(&t);
                }
            }
        }

        let tables = self.tables.read().unwrap();
        let mut operator = self.build_operator(plan, &tables, &self.columnar_cache)?;
        Self::drain_operator(&mut *operator)
    }

    /// Build an operator tree from a physical plan.
    fn build_operator(
        &self,
        plan: &PhysicalPlan,
        tables: &HashMap<String, Vec<RecordBatch>>,
        columnar_cache: &ColumnarCache,
    ) -> DbxResult<Box<dyn PhysicalOperator>> {
        match plan {
            PhysicalPlan::TableScan {
                table,
                projection,
                filter,
                ros_files,
            } => {
                // Track OLAP Workload
                self.workload_analyzer
                    .write()
                    .unwrap()
                    .record(crate::engine::workload_analyzer::QueryPattern::RangeScan);

                let mut filter_pushed_down = false;
                let target_tables = self.get_tables_to_scan(table, filter.as_ref());

                let mut all_batches = Vec::new();
                let mut base_schema = None;

                for t in target_tables {
                    // Try Columnar Cache first (with projection AND filter pushdown!)
                    let cached_results = if let Some(filter_expr) = filter {
                        let filter_expr_clone = filter_expr.clone();
                        // Use pushdown with filter
                        let result = columnar_cache.get_batches_with_filter(
                            &t,
                            if projection.is_empty() {
                                None
                            } else {
                                Some(projection)
                            },
                            move |batch| {
                                use crate::sql::executor::evaluate_expr;
                                use arrow::array::BooleanArray;

                                let array = evaluate_expr(&filter_expr_clone, batch)?;
                                let boolean_array = array
                                    .as_any()
                                    .downcast_ref::<BooleanArray>()
                                    .ok_or_else(|| DbxError::TypeMismatch {
                                    expected: "BooleanArray".to_string(),
                                    actual: format!("{:?}", array.data_type()),
                                })?;
                                Ok(boolean_array.clone())
                            },
                        )?;
                        if result.is_some() {
                            filter_pushed_down = true;
                        }
                        result
                    } else {
                        columnar_cache.get_batches(
                            &t,
                            if projection.is_empty() {
                                None
                            } else {
                                Some(projection)
                            },
                        )?
                    };

                    let (batches, schema, _) = if let Some(cached_batches) = cached_results {
                        if cached_batches.is_empty() {
                            // Table exists but is empty in cache.
                            // We need the schema to return an empty scan.
                            let schema = {
                                let schemas = self.table_schemas.read().unwrap();
                                schemas
                                    .get(table)
                                    .or_else(|| {
                                        let table_lower = table.to_lowercase();
                                        schemas
                                            .iter()
                                            .find(|(k, _)| k.to_lowercase() == table_lower)
                                            .map(|(_, v)| v)
                                    })
                                    .cloned()
                            }
                            .unwrap_or_else(|| Arc::new(arrow::datatypes::Schema::empty()));
                            (vec![], schema, projection.clone())
                        } else {
                            let schema = cached_batches[0].schema();
                            (cached_batches, schema, vec![])
                        }
                    } else {
                        // Try cache again
                        let cached_after_sync = columnar_cache.get_batches(
                            &t,
                            if projection.is_empty() {
                                None
                            } else {
                                Some(projection)
                            },
                        )?;

                        if let Some(batches) = cached_after_sync {
                            if !batches.is_empty() {
                                let schema = batches[0].schema();
                                (batches, schema, vec![])
                            } else {
                                // Table exists but is empty
                                let schema = {
                                    let schemas = self.table_schemas.read().unwrap();
                                    schemas
                                        .get(table)
                                        .or_else(|| {
                                            let table_lower = table.to_lowercase();
                                            schemas
                                                .iter()
                                                .find(|(k, _)| k.to_lowercase() == table_lower)
                                                .map(|(_, v)| v)
                                        })
                                        .cloned()
                                }
                                .unwrap_or_else(|| Arc::new(arrow::datatypes::Schema::empty()));
                                (vec![], schema, projection.clone())
                            }
                        } else {
                            // Not in cache, check HashMap fallback
                            let batches_opt = tables.get(&t).or_else(|| {
                                let table_lower = t.to_lowercase();
                                tables
                                    .iter()
                                    .find(|(k, _)| k.to_lowercase() == table_lower)
                                    .map(|(_, v)| v)
                            });

                            if let Some(batches) = batches_opt {
                                if batches.is_empty() {
                                    let schema = {
                                        let schemas = self.table_schemas.read().unwrap();
                                        schemas
                                            .get(table)
                                            .or_else(|| {
                                                let table_lower = table.to_lowercase();
                                                schemas
                                                    .iter()
                                                    .find(|(k, _)| k.to_lowercase() == table_lower)
                                                    .map(|(_, v)| v)
                                            })
                                            .cloned()
                                    }
                                    .unwrap_or_else(|| Arc::new(arrow::datatypes::Schema::empty()));
                                    (vec![], schema, projection.clone())
                                } else {
                                    let schema = batches[0].schema();
                                    (batches.clone(), schema, projection.clone())
                                }
                            } else {
                                // Skip missing sub-table if it has no data yet
                                let schema = {
                                    let schemas = self.table_schemas.read().unwrap();
                                    schemas.get(table).cloned()
                                }
                                .unwrap_or_else(|| Arc::new(arrow::datatypes::Schema::empty()));
                                (vec![], schema, projection.clone())
                            }
                        }
                    };

                    if base_schema.is_none() {
                        base_schema = Some(schema);
                    }
                    all_batches.extend(batches);
                }

                // ---------------------- PHASE 3: ROS (Parquet) Union (Optimized) ----------------------
                // [Phase 9] 최적화: 매 쿼리마다 read_dir을 수행하는 대신 PhysicalPlan에 전달된 ros_files 정보를 TableScanOperator가 처리하도록 위임함.
                // --------------------------------------------------------------------------

                // Use the base schema from the loop, or try to look it up if all partitions were empty
                let final_schema = base_schema.unwrap_or_else(|| {
                    let schemas = self.table_schemas.read().unwrap();
                    schemas
                        .get(table)
                        .cloned()
                        .unwrap_or_else(|| Arc::new(arrow::datatypes::Schema::empty()))
                });

                let projection_to_use = if projection.is_empty() {
                    vec![]
                } else {
                    projection.clone()
                };
                let mut scan =
                    TableScanOperator::new(table.clone(), final_schema, projection_to_use);

                // ros_files가 채워져 있으면 파이프라인 티어 스캔,
                // 비어있으면 기존의 all_batches 통째 주입
                if !ros_files.is_empty() {
                    scan.start_tier_scan(all_batches, ros_files.clone());
                } else {
                    scan.set_data(all_batches);
                }

                // Wrap with filter if needed AND NOT pushed down
                if let Some(filter_expr) = filter {
                    if !filter_pushed_down {
                        Ok(Box::new(FilterOperator::new(
                            Box::new(scan),
                            filter_expr.clone(),
                        )))
                    } else {
                        // Filter already applied in scan (via cache)
                        Ok(Box::new(scan))
                    }
                } else {
                    Ok(Box::new(scan))
                }
            }

            PhysicalPlan::Projection {
                input,
                exprs,
                aliases,
            } => {
                let input_op = self.build_operator(input, tables, columnar_cache)?;
                use arrow::datatypes::Field;

                let input_schema = input_op.schema();
                let fields: Vec<Field> = exprs
                    .iter()
                    .enumerate()
                    .map(|(i, expr)| {
                        let data_type = expr.get_type(input_schema);
                        let field_name = if let Some(Some(alias)) = aliases.get(i) {
                            alias.clone()
                        } else {
                            format!("col_{}", i)
                        };
                        Field::new(&field_name, data_type, true)
                    })
                    .collect();

                let output_schema = Arc::new(Schema::new(fields));
                Ok(Box::new(ProjectionOperator::new(
                    input_op,
                    output_schema,
                    exprs.clone(),
                )))
            }

            PhysicalPlan::Limit {
                input,
                count,
                offset,
            } => {
                let input_op = self.build_operator(input, tables, columnar_cache)?;
                Ok(Box::new(LimitOperator::new(input_op, *count, *offset)))
            }

            PhysicalPlan::SortMerge { input, order_by } => {
                let input_op = self.build_operator(input, tables, columnar_cache)?;
                Ok(Box::new(SortOperator::new(input_op, order_by.clone())))
            }

            PhysicalPlan::HashAggregate {
                input,
                group_by,
                aggregates,
                mode,
            } => {
                // Track OLAP Workload
                self.workload_analyzer
                    .write()
                    .unwrap()
                    .record(crate::engine::workload_analyzer::QueryPattern::Aggregation);

                let input_op = self.build_operator(input, tables, columnar_cache)?;
                let input_schema = input_op.schema();
                let mut fields = Vec::new();

                for &idx in group_by {
                    fields.push(input_schema.field(idx).clone());
                }

                for agg in aggregates {
                    use arrow::datatypes::DataType;
                    let data_type = match agg.function {
                        crate::sql::planner::AggregateFunction::Count => DataType::Int64,
                        crate::sql::planner::AggregateFunction::Sum
                        | crate::sql::planner::AggregateFunction::Avg
                        | crate::sql::planner::AggregateFunction::Min
                        | crate::sql::planner::AggregateFunction::Max => DataType::Float64,
                    };
                    let name = agg
                        .alias
                        .clone()
                        .unwrap_or_else(|| format!("agg_{:?}", agg.function));
                    fields.push(arrow::datatypes::Field::new(&name, data_type, true));
                }

                let agg_schema = Arc::new(arrow::datatypes::Schema::new(fields));
                Ok(Box::new(
                    HashAggregateOperator::new(
                        input_op,
                        agg_schema,
                        group_by.clone(),
                        aggregates.clone(),
                        *mode,
                    )
                    .with_gpu(self.gpu_manager.clone()),
                ))
            }

            PhysicalPlan::HashJoin {
                left,
                right,
                on,
                join_type,
            } => {
                // Track OLAP Workload
                self.workload_analyzer
                    .write()
                    .unwrap()
                    .record(crate::engine::workload_analyzer::QueryPattern::Join);

                use arrow::datatypes::Field;

                let left_op = self.build_operator(left, tables, columnar_cache)?;
                let right_op = self.build_operator(right, tables, columnar_cache)?;

                // Build joined schema: left columns + right columns
                let left_schema = left_op.schema();
                let right_schema = right_op.schema();

                let mut joined_fields: Vec<Field> = Vec::new();
                for field in left_schema.fields().iter() {
                    let mut f = field.as_ref().clone();
                    // In RIGHT JOIN, left side can be null
                    if matches!(join_type, crate::sql::planner::JoinType::Right) {
                        f = f.with_nullable(true);
                    }
                    joined_fields.push(f);
                }
                for field in right_schema.fields().iter() {
                    let mut f = field.as_ref().clone();
                    // In LEFT JOIN, right side can be null
                    if matches!(join_type, crate::sql::planner::JoinType::Left) {
                        f = f.with_nullable(true);
                    }
                    joined_fields.push(f);
                }

                let joined_schema = Arc::new(Schema::new(joined_fields));

                Ok(Box::new(HashJoinOperator::new(
                    left_op,
                    right_op,
                    joined_schema,
                    on.clone(),
                    *join_type,
                )))
            }

            PhysicalPlan::Insert { .. } => {
                // INSERT should be handled in execute_physical_plan, not here
                unreachable!("INSERT should not reach build_operator")
            }
            PhysicalPlan::Update { .. } => {
                // UPDATE should be handled in execute_physical_plan, not here
                unreachable!("UPDATE should not reach build_operator")
            }
            PhysicalPlan::Delete { .. } => {
                // DELETE should be handled in execute_physical_plan, not here
                unreachable!("DELETE should not reach build_operator")
            }
            PhysicalPlan::DropTable { .. } => {
                // DROP TABLE should be handled in execute_physical_plan, not here
                unreachable!("DROP TABLE should not reach build_operator")
            }
            PhysicalPlan::CreateTable { .. } => {
                // CREATE TABLE should be handled in execute_physical_plan, not here
                unreachable!("CREATE TABLE should not reach build_operator")
            }
            PhysicalPlan::CreateIndex { .. } => {
                // CREATE INDEX should be handled in execute_physical_plan, not here
                unreachable!("CREATE INDEX should not reach build_operator")
            }
            PhysicalPlan::DropIndex { .. } => {
                // DROP INDEX should be handled in execute_physical_plan, not here
                unreachable!("DROP INDEX should not reach build_operator")
            }
            PhysicalPlan::AlterTable { .. } => {
                // ALTER TABLE should be handled in execute_physical_plan, not here
                unreachable!("ALTER TABLE should not reach build_operator")
            }
            PhysicalPlan::CreateFunction { .. } => {
                unreachable!("CREATE FUNCTION should not reach build_operator")
            }
            PhysicalPlan::CreateTrigger { .. } => {
                unreachable!("CREATE TRIGGER should not reach build_operator")
            }
            PhysicalPlan::CreateJob { .. } => {
                unreachable!("CREATE JOB should not reach build_operator")
            }
            PhysicalPlan::DropFunction { .. } => {
                unreachable!("DROP FUNCTION should not reach build_operator")
            }
            PhysicalPlan::DropTrigger { .. } => {
                unreachable!("DROP TRIGGER should not reach build_operator")
            }
            PhysicalPlan::DropJob { .. } => {
                unreachable!("DROP JOB should not reach build_operator")
            }
            PhysicalPlan::GridExchange { .. } => {
                // GridExchange 플레이스홀더 — DistributedExecutor가 런타임에 교체하므로
                // interface의 build_operator에 도달하면 설계 오류
                unreachable!(
                    "GridExchange placeholder must be replaced by DistributedExecutor before build_operator is called"
                )
            }
            PhysicalPlan::ShuffleWriter { .. } => {
                unreachable!(
                    "ShuffleWriter is a distributed operator and not supported in singleton Database"
                )
            }
        }
    }

    /// Drain all batches from an operator.
    fn drain_operator(op: &mut dyn PhysicalOperator) -> DbxResult<Vec<RecordBatch>> {
        let mut results = Vec::new();
        while let Some(batch) = op.next()? {
            if batch.num_rows() > 0 {
                results.push(batch);
            }
        }
        Ok(results)
    }
}

// ════════════════════════════════════════════
// DatabaseSql Trait Implementation
// ════════════════════════════════════════════

impl crate::traits::DatabaseSql for Database {
    fn execute_sql(&self, sql: &str) -> DbxResult<Vec<arrow::record_batch::RecordBatch>> {
        // Reuse existing implementation
        Database::execute_sql(self, sql)
    }

    fn register_table(&self, name: &str, batches: Vec<arrow::record_batch::RecordBatch>) {
        // Reuse existing implementation
        Database::register_table(self, name, batches)
    }

    fn append_batch(&self, table: &str, batch: arrow::record_batch::RecordBatch) -> DbxResult<()> {
        // Reuse existing implementation
        Database::append_batch(self, table, batch);
        Ok(())
    }
}