fraiseql_core/runtime/

aggregate_parser.rs

//! Aggregate Query Parser
//!
//! Parses GraphQL aggregate queries into `AggregationRequest` for execution.
//!
//! # GraphQL Query Format
//!
//! ```graphql
//! query {
//!   sales_aggregate(
//!     where: { customer_id: { _eq: "uuid-123" } }
//!     groupBy: { category: true, occurred_at_day: true }
//!     having: { revenue_sum_gt: 1000 }
//!     orderBy: { revenue_sum: DESC }
//!     limit: 10
//!   ) {
//!     category
//!     occurred_at_day
//!     count
//!     revenue_sum
//!     revenue_avg
//!   }
//! }
//! ```
//!
//! # Parsed Result
//!
//! ```text
//! // Illustrative output structure only — not executable code.
//! AggregationRequest {
//!     table_name: "tf_sales",
//!     where_clause: Some(...),
//!     group_by: vec![
//!         GroupBySelection::Dimension { path: "category", alias: "category" },
//!         GroupBySelection::TemporalBucket { column: "occurred_at", bucket: Day, alias: "occurred_at_day" },
//!     ],
//!     aggregates: vec![
//!         AggregateSelection::Count { alias: "count" },
//!         AggregateSelection::MeasureAggregate { measure: "revenue", function: Sum, alias: "revenue_sum" },
//!         AggregateSelection::MeasureAggregate { measure: "revenue", function: Avg, alias: "revenue_avg" },
//!     ],
//!     having: vec![...],
//!     order_by: vec![...],
//!     limit: Some(10),
//!     offset: None,
//! }
//! ```

use serde_json::Value;

use crate::{
    backend::where_clause::{WhereClause, WhereOperator},
    compiler::{
        aggregate_types::{AggregateFunction, HavingOperator, TemporalBucket},
        aggregation::{
            AggregateSelection, AggregationRequest, GroupBySelection, HavingCondition,
            OrderByClause, OrderDirection, TreeGrouping,
        },
        fact_table::FactTableMetadata,
    },
    error::{FraiseQLError, Result},
};

/// Aggregate query parser
pub struct AggregateQueryParser;

impl AggregateQueryParser {
    /// Parse a simplified aggregate query into `AggregationRequest`.
    ///
    /// For we'll accept a JSON structure that represents the query:
    /// ```json
    /// {
    ///   "table": "tf_sales",
    ///   "groupBy": {
    ///     "category": true,
    ///     "occurred_at_day": true
    ///   },
    ///   "aggregates": [
    ///     {"count": {}},
    ///     {"revenue_sum": {}}
    ///   ],
    ///   "having": {
    ///     "revenue_sum_gt": 1000
    ///   },
    ///   "orderBy": {
    ///     "revenue_sum": "DESC"
    ///   },
    ///   "limit": 10
    /// }
    /// ```
    ///
    /// # Errors
    ///
    /// Returns error if the query structure is invalid or references non-existent measures.
    ///
    /// The `native_columns` map (column name → PostgreSQL cast string) controls whether
    /// WHERE filters and GROUP BY fields are emitted as direct column references
    /// (`WhereClause::NativeField`, `GroupBySelection::NativeDimension`) rather than
    /// JSONB extraction variants.  Pass an empty map to get the old behaviour.
    pub fn parse(
        query_json: &Value,
        metadata: &FactTableMetadata,
        native_columns: &std::collections::HashMap<String, String>,
    ) -> Result<AggregationRequest> {
        // Extract table name
        let table_name = query_json
            .get("table")
            .and_then(|v| v.as_str())
            .ok_or_else(|| FraiseQLError::Validation {
                message: "Missing 'table' field in aggregate query".to_string(),
                path:    None,
            })?
            .to_string();

        // Parse WHERE clause (if present)
        let where_clause = if let Some(where_obj) = query_json.get("where") {
            Some(Self::parse_where_clause(where_obj, metadata, native_columns)?)
        } else {
            None
        };

        // Parse GROUP BY selections
        let group_by = if let Some(group_by_obj) = query_json.get("groupBy") {
            Self::parse_group_by(group_by_obj, metadata, native_columns)?
        } else {
            vec![]
        };

        // Parse aggregate selections from requested fields
        let aggregates = if let Some(agg_array) = query_json.get("aggregates") {
            Self::parse_aggregates(agg_array, metadata)?
        } else {
            vec![]
        };

        // Parse HAVING conditions
        let having = if let Some(having_obj) = query_json.get("having") {
            Self::parse_having(having_obj, &aggregates, metadata)?
        } else {
            vec![]
        };

        // Parse ORDER BY clauses
        let order_by = if let Some(order_obj) = query_json.get("orderBy") {
            Self::parse_order_by(order_obj)?
        } else {
            vec![]
        };

        // Parse LIMIT
        let limit = query_json
            .get("limit")
            .and_then(|v| v.as_u64())
            .map(|n| u32::try_from(n).unwrap_or(u32::MAX));

        // Parse OFFSET
        let offset = query_json
            .get("offset")
            .and_then(|v| v.as_u64())
            .map(|n| u32::try_from(n).unwrap_or(u32::MAX));

        Ok(AggregationRequest {
            table_name,
            where_clause,
            group_by,
            aggregates,
            having,
            order_by,
            limit,
            offset,
        })
    }

    /// Parse WHERE clause from JSON.
    ///
    /// For aggregate queries, WHERE works on denormalized filter columns, on dimensions a
    /// `native_dimension_mapping` maps to a column (read from that column), and on
    /// top-level JSONB dimension keys.
    /// Expected format: `{ "field_operator": value }`
    /// Example: `{ "customer_id_eq": "123", "occurred_at_gte": "2024-01-01" }`
    ///
    /// When a field name matches an entry in `native_columns`, the clause is emitted
    /// as [`WhereClause::NativeField`] (direct column reference) rather than
    /// [`WhereClause::Field`] (JSONB extraction).
    fn parse_where_clause(
        where_obj: &Value,
        metadata: &FactTableMetadata,
        native_columns: &std::collections::HashMap<String, String>,
    ) -> Result<WhereClause> {
        let Some(obj) = where_obj.as_object() else {
            return Ok(WhereClause::And(vec![]));
        };

        let mut conditions = Vec::new();

        for (key, value) in obj {
            // Parse field_operator format (e.g., "customer_id_eq" -> field="customer_id",
            // operator="eq")
            let (field, operator) = split_where_key(key, metadata, native_columns)?;

            let clause = if let Some(column) = mapped_column(metadata, field) {
                // A dimension mapped to a native column is read from the column, as its
                // `groupBy` is (#1231), cast as the column's declared filter type is.
                WhereClause::NativeField {
                    column: column.clone(),
                    pg_cast: native_columns.get(column.as_str()).cloned().unwrap_or_default(),
                    operator,
                    value: value.clone(),
                }
            } else if let Some(pg_cast) = native_columns.get(field) {
                WhereClause::NativeField {
                    column: field.to_string(),
                    pg_cast: pg_cast.clone(),
                    operator,
                    value: value.clone(),
                }
            } else {
                WhereClause::Field {
                    // Recase the JSONB key so a camelCase aggregate filter
                    // (`organizationId_eq`) builds `data->>'organization_id'`
                    // rather than a never-matching `organizationId` key (#486).
                    // Only the non-native branch recases — the native lookup
                    // above stays on the surface name (mirror `query_params`).
                    path: vec![crate::utils::to_snake_case(field)],
                    operator,
                    value: value.clone(),
                }
            };
            conditions.push(clause);
        }

        Ok(WhereClause::And(conditions))
    }

    /// Parse GROUP BY selections.
    ///
    /// Supports two formats:
    /// 1. Boolean true: {"category": true} -> regular dimension
    /// 2. Boolean true with suffix: {"`occurred_at_day"`: true} -> temporal bucket
    /// 3. String bucket name: {"`occurred_at"`: "day"} -> temporal bucket
    ///
    /// When a field name matches an entry in `native_columns`, the selection is emitted
    /// as [`GroupBySelection::NativeDimension`] rather than [`GroupBySelection::Dimension`].
    fn parse_group_by(
        group_by_obj: &Value,
        metadata: &FactTableMetadata,
        native_columns: &std::collections::HashMap<String, String>,
    ) -> Result<Vec<GroupBySelection>> {
        let mut selections = Vec::new();

        if let Some(obj) = group_by_obj.as_object() {
            for (key, value) in obj {
                if value.as_bool() == Some(true) {
                    // Format 1 & 2: Boolean true (with or without suffix)
                    // Priority 1: Try calendar dimension first (highest performance)
                    if let Some(calendar_sel) = Self::try_parse_calendar_bucket(key, metadata)? {
                        selections.push(calendar_sel);
                    } else if let Some(bucket_sel) = Self::parse_temporal_bucket(key, metadata)? {
                        // Priority 2: Fall back to DATE_TRUNC if no calendar dimension
                        selections.push(bucket_sel);
                    } else if let Some(mapped_col) = mapped_column(metadata, key) {
                        // Priority 3: Deep JSONB path mapped to a native column
                        selections.push(GroupBySelection::NativeDimension {
                            column:  mapped_col.clone(),
                            pg_cast: native_columns
                                .get(mapped_col.as_str())
                                .cloned()
                                .unwrap_or_default(),
                            alias:   key.clone(),
                        });
                    } else if let Some(pg_cast) = native_columns.get(key.as_str()) {
                        // Priority 4: Native SQL column (filter-derived) — direct reference
                        selections.push(GroupBySelection::NativeDimension {
                            column:  key.clone(),
                            pg_cast: pg_cast.clone(),
                            alias:   key.clone(),
                        });
                    } else {
                        // Priority 5: Regular JSONB dimension. Unlike priorities 1-4
                        // (each gated by a lookup against server-defined calendar/
                        // temporal/native metadata), this fallback echoes the raw
                        // GraphQL key verbatim as the SELECT alias (`… AS {key}`),
                        // which is interpolated into SQL without quoting. Reject keys
                        // that are not safe identifiers so an attacker cannot inject an
                        // extra SELECT column (H1) — e.g. `a, (SELECT …) AS leak`. This
                        // fires at parse time, independent of the compiler's dimension
                        // allowlist (which is skipped when `dimensions.paths` is empty).
                        validate_dimension_key(key)?;
                        selections.push(GroupBySelection::Dimension {
                            // The `path` is the JSONB extraction key → recase to
                            // snake_case so a camelCase dimension (`machineStatus`)
                            // groups by `data->>'machine_status'` (#486). The `alias`
                            // is the result/response column name, consumed verbatim by
                            // `AggregationProjector`, so it keeps the camel surface name
                            // (the #418/#410 "output key is the alias" rule).
                            path:  crate::utils::to_snake_case(key),
                            alias: key.clone(),
                        });
                    }
                } else if value.as_str() == Some("depth") {
                    // Format 4: tree depth of an ltree path {"path": "depth"} (#1498)
                    selections.push(Self::tree_level(key, TreeGrouping::Depth, metadata)?);
                } else if let Some(level) = value.as_object().and_then(|o| o.get("level")) {
                    // Format 5: roll a path up to a tree level {"path": {"level": 2}} (#1498)
                    let level = level
                        .as_u64()
                        .filter(|n| *n >= 1)
                        .and_then(|n| u32::try_from(n).ok())
                        .ok_or_else(|| FraiseQLError::Validation {
                            message: format!(
                                "groupBy '{key}': level must be a positive integer, got {level}"
                            ),
                            path:    None,
                        })?;
                    selections.push(Self::tree_level(key, TreeGrouping::Level(level), metadata)?);
                } else if let Some(bucket_str) = value.as_str() {
                    // Format 3: String bucket name {"occurred_at": "day"}
                    let bucket = TemporalBucket::from_str(bucket_str)?;

                    // Priority 1: Try calendar dimension first
                    if let Some(calendar_sel) =
                        Self::try_find_calendar_bucket(key, bucket, metadata)
                    {
                        selections.push(calendar_sel);
                    } else {
                        // Priority 2: Fall back to DATE_TRUNC
                        // Verify this column exists in denormalized_filters
                        let column_exists =
                            metadata.denormalized_filters.iter().any(|f| f.name == *key);

                        if !column_exists {
                            return Err(FraiseQLError::Validation {
                                message: format!(
                                    "Temporal bucketing column '{}' not found in denormalized filters",
                                    key
                                ),
                                path:    None,
                            });
                        }

                        selections.push(GroupBySelection::TemporalBucket {
                            column: key.clone(),
                            bucket,
                            alias: key.clone(),
                        });
                    }
                }
            }
        }

        Ok(selections)
    }

    /// A tree grouping of the denormalized filter column `key` (#1498). The column's
    /// type is checked when the request is planned.
    fn tree_level(
        key: &str,
        grouping: TreeGrouping,
        metadata: &FactTableMetadata,
    ) -> Result<GroupBySelection> {
        if !metadata.denormalized_filters.iter().any(|f| f.name == key) {
            return Err(FraiseQLError::Validation {
                message: format!(
                    "groupBy '{key}': a tree level or depth needs an ltree denormalized filter \
                     column, and '{key}' is not a denormalized filter of '{}'",
                    metadata.table_name
                ),
                path:    None,
            });
        }
        let alias = match grouping {
            TreeGrouping::Depth => format!("{key}_depth"),
            TreeGrouping::Level(_) => key.to_string(),
        };
        Ok(GroupBySelection::TreeLevel {
            column: key.to_string(),
            grouping,
            alias,
        })
    }

    /// Parse temporal bucket if the key matches pattern
    fn parse_temporal_bucket(
        key: &str,
        metadata: &FactTableMetadata,
    ) -> Result<Option<GroupBySelection>> {
        // Check for temporal bucket patterns: column_day, column_week, etc.
        for filter_col in &metadata.denormalized_filters {
            for bucket in &[
                ("_second", TemporalBucket::Second),
                ("_minute", TemporalBucket::Minute),
                ("_hour", TemporalBucket::Hour),
                ("_day", TemporalBucket::Day),
                ("_week", TemporalBucket::Week),
                ("_month", TemporalBucket::Month),
                ("_quarter", TemporalBucket::Quarter),
                ("_year", TemporalBucket::Year),
            ] {
                let expected_key = format!("{}{}", filter_col.name, bucket.0);
                if key == expected_key {
                    return Ok(Some(GroupBySelection::TemporalBucket {
                        column: filter_col.name.clone(),
                        bucket: bucket.1,
                        alias:  key.to_string(),
                    }));
                }
            }
        }

        Ok(None)
    }

    /// Try to parse calendar dimension from key pattern (e.g., "`occurred_at_day`")
    ///
    /// Checks if the key matches a calendar dimension pattern and returns
    /// a `CalendarDimension` selection if available, otherwise None.
    fn try_parse_calendar_bucket(
        key: &str,
        metadata: &FactTableMetadata,
    ) -> Result<Option<GroupBySelection>> {
        for calendar_dim in &metadata.calendar_dimensions {
            // Check all temporal bucket suffixes
            for (suffix, bucket_type) in &[
                ("_second", TemporalBucket::Second),
                ("_minute", TemporalBucket::Minute),
                ("_hour", TemporalBucket::Hour),
                ("_day", TemporalBucket::Day),
                ("_week", TemporalBucket::Week),
                ("_month", TemporalBucket::Month),
                ("_quarter", TemporalBucket::Quarter),
                ("_year", TemporalBucket::Year),
            ] {
                let expected_key = format!("{}{}", calendar_dim.source_column, suffix);
                if key == expected_key {
                    // Find matching calendar bucket
                    if let Some((gran, bucket)) =
                        Self::find_calendar_bucket(calendar_dim, *bucket_type)
                    {
                        return Ok(Some(GroupBySelection::CalendarDimension {
                            source_column:   calendar_dim.source_column.clone(),
                            calendar_column: gran.column_name.clone(),
                            json_key:        bucket.json_key.clone(),
                            bucket:          bucket.bucket_type,
                            alias:           key.to_string(),
                        }));
                    }
                }
            }
        }
        Ok(None)
    }

    /// Try to find calendar bucket for explicit temporal request
    ///
    /// Used when user provides explicit bucket like {"`occurred_at"`: "day"}
    fn try_find_calendar_bucket(
        column: &str,
        bucket: TemporalBucket,
        metadata: &FactTableMetadata,
    ) -> Option<GroupBySelection> {
        for calendar_dim in &metadata.calendar_dimensions {
            if calendar_dim.source_column == column {
                if let Some((gran, cal_bucket)) = Self::find_calendar_bucket(calendar_dim, bucket) {
                    return Some(GroupBySelection::CalendarDimension {
                        source_column:   calendar_dim.source_column.clone(),
                        calendar_column: gran.column_name.clone(),
                        json_key:        cal_bucket.json_key.clone(),
                        bucket:          cal_bucket.bucket_type,
                        alias:           column.to_string(),
                    });
                }
            }
        }
        None
    }

    /// Find calendar bucket in available granularities
    ///
    /// Searches through calendar dimension granularities to find a matching bucket type.
    /// Returns the granularity and bucket if found.
    fn find_calendar_bucket(
        calendar_dim: &crate::compiler::fact_table::CalendarDimension,
        bucket: TemporalBucket,
    ) -> Option<(
        &crate::compiler::fact_table::CalendarGranularity,
        &crate::compiler::fact_table::CalendarBucket,
    )> {
        for granularity in &calendar_dim.granularities {
            for cal_bucket in &granularity.buckets {
                if cal_bucket.bucket_type == bucket {
                    return Some((granularity, cal_bucket));
                }
            }
        }
        None
    }

    /// Parse aggregate selections
    fn parse_aggregates(
        agg_array: &Value,
        metadata: &FactTableMetadata,
    ) -> Result<Vec<AggregateSelection>> {
        let mut aggregates = Vec::new();

        if let Some(arr) = agg_array.as_array() {
            for item in arr {
                if let Some(obj) = item.as_object() {
                    // Each object should have one key (the aggregate name)
                    for (agg_name, _value) in obj {
                        aggregates.push(Self::parse_aggregate_selection(agg_name, metadata)?);
                    }
                }
            }
        }

        Ok(aggregates)
    }

    /// Parse a single aggregate selection
    fn parse_aggregate_selection(
        agg_name: &str,
        metadata: &FactTableMetadata,
    ) -> Result<AggregateSelection> {
        // Handle COUNT
        if agg_name == "count" {
            return Ok(AggregateSelection::Count {
                alias: "count".to_string(),
            });
        }

        // Handle COUNT_DISTINCT: supports both "count_distinct" (defaults to first dimension)
        // and "field_count_distinct" pattern (e.g., "product_id_count_distinct")
        if agg_name == "count_distinct" {
            // Default to first dimension path, or "id" if none available
            let default_field = Self::extract_dimension_paths(metadata)
                .first()
                .cloned()
                .unwrap_or_else(|| "id".to_string());
            return Ok(AggregateSelection::CountDistinct {
                field: default_field,
                alias: "count_distinct".to_string(),
            });
        }

        // Handle field_count_distinct pattern (e.g., "customer_id_count_distinct")
        if let Some(stripped) = agg_name.strip_suffix("_count_distinct") {
            // Check if the stripped part matches a dimension path
            let dimension_paths = Self::extract_dimension_paths(metadata);
            if dimension_paths.iter().any(|p| p == stripped) {
                return Ok(AggregateSelection::CountDistinct {
                    field: stripped.to_string(),
                    alias: agg_name.to_string(),
                });
            }
            // Also allow count distinct on measures
            if metadata.measures.iter().any(|m| m.name == stripped) {
                return Ok(AggregateSelection::CountDistinct {
                    field: stripped.to_string(),
                    alias: agg_name.to_string(),
                });
            }
            // If no match found, return error with helpful message
            return Err(FraiseQLError::Validation {
                message: format!(
                    "COUNT DISTINCT field '{}' not found in dimensions or measures. Available: {:?}",
                    stripped, dimension_paths
                ),
                path:    None,
            });
        }

        // Handle boolean aggregates (BOOL_AND, BOOL_OR)
        // e.g., "is_active_bool_and", "has_discount_bool_or"
        for dimension_path in Self::extract_dimension_paths(metadata) {
            if let Some(stripped) = agg_name.strip_suffix("_bool_and") {
                if stripped == dimension_path {
                    return Ok(AggregateSelection::BoolAggregate {
                        field:    dimension_path,
                        function: crate::compiler::aggregate_types::BoolAggregateFunction::And,
                        alias:    agg_name.to_string(),
                    });
                }
            }
            if let Some(stripped) = agg_name.strip_suffix("_bool_or") {
                if stripped == dimension_path {
                    return Ok(AggregateSelection::BoolAggregate {
                        field:    dimension_path,
                        function: crate::compiler::aggregate_types::BoolAggregateFunction::Or,
                        alias:    agg_name.to_string(),
                    });
                }
            }
        }

        // Handle measure aggregates: revenue_sum, revenue_avg, etc.
        for measure in &metadata.measures {
            for func in &[
                ("_sum", AggregateFunction::Sum),
                ("_avg", AggregateFunction::Avg),
                ("_min", AggregateFunction::Min),
                ("_max", AggregateFunction::Max),
                ("_stddev", AggregateFunction::Stddev),
                ("_variance", AggregateFunction::Variance),
                // Advanced aggregates
                ("_array_agg", AggregateFunction::ArrayAgg),
                ("_json_agg", AggregateFunction::JsonAgg),
                ("_jsonb_agg", AggregateFunction::JsonbAgg),
                ("_string_agg", AggregateFunction::StringAgg),
            ] {
                let expected_name = format!("{}{}", measure.name, func.0);
                if agg_name == expected_name {
                    return Ok(AggregateSelection::MeasureAggregate {
                        measure:  measure.name.clone(),
                        function: func.1,
                        alias:    agg_name.to_string(),
                    });
                }
            }
        }

        // Handle native measure aggregates: measures.volume_sum, measures.cost_avg, etc.
        // These are JSONB measure paths mapped to flat SQL columns via native_measures.
        for jsonb_path in metadata.native_measures.keys() {
            for func in &[
                ("_sum", AggregateFunction::Sum),
                ("_avg", AggregateFunction::Avg),
                ("_min", AggregateFunction::Min),
                ("_max", AggregateFunction::Max),
                ("_stddev", AggregateFunction::Stddev),
                ("_variance", AggregateFunction::Variance),
                ("_array_agg", AggregateFunction::ArrayAgg),
                ("_json_agg", AggregateFunction::JsonAgg),
                ("_jsonb_agg", AggregateFunction::JsonbAgg),
                ("_string_agg", AggregateFunction::StringAgg),
            ] {
                let expected_name = format!("{}{}", jsonb_path, func.0);
                if agg_name == expected_name {
                    // Store the JSONB path as the measure name; the planner resolves
                    // it to the native column via FactTableMetadata.native_measures.
                    return Ok(AggregateSelection::MeasureAggregate {
                        measure:  jsonb_path.clone(),
                        function: func.1,
                        alias:    agg_name.to_string(),
                    });
                }
            }
        }

        // Check for dimension-level advanced aggregates
        // e.g., "product_id_array_agg", "product_name_string_agg"
        for dimension_path in Self::extract_dimension_paths(metadata) {
            for func in &[
                ("_array_agg", AggregateFunction::ArrayAgg),
                ("_json_agg", AggregateFunction::JsonAgg),
                ("_jsonb_agg", AggregateFunction::JsonbAgg),
                ("_string_agg", AggregateFunction::StringAgg),
            ] {
                let expected_name = format!("{}{}", dimension_path, func.0);
                if agg_name == expected_name {
                    // For dimension aggregates, store the path as the "measure"
                    return Ok(AggregateSelection::MeasureAggregate {
                        measure:  dimension_path,
                        function: func.1,
                        alias:    agg_name.to_string(),
                    });
                }
            }
        }

        Err(FraiseQLError::Validation {
            message: format!("Unknown aggregate selection: {agg_name}"),
            path:    None,
        })
    }

    /// Extract dimension paths from metadata for advanced aggregate parsing
    fn extract_dimension_paths(metadata: &FactTableMetadata) -> Vec<String> {
        let mut paths = Vec::new();

        // Add dimension paths from JSONB column
        for dim_path in &metadata.dimensions.paths {
            paths.push(dim_path.name.clone());
        }

        // Add denormalized filter columns (these can also be aggregated)
        for filter in &metadata.denormalized_filters {
            paths.push(filter.name.clone());
        }

        paths
    }

    /// Parse HAVING conditions
    fn parse_having(
        having_obj: &Value,
        aggregates: &[AggregateSelection],
        _metadata: &FactTableMetadata,
    ) -> Result<Vec<HavingCondition>> {
        let mut conditions = Vec::new();

        if let Some(obj) = having_obj.as_object() {
            for (key, value) in obj {
                // Parse condition: revenue_sum_gt: 1000
                if let Some((agg_name, operator)) = Self::parse_having_key(key) {
                    // Find the aggregate
                    let aggregate = aggregates
                        .iter()
                        .find(|a| a.alias() == agg_name)
                        .ok_or_else(|| FraiseQLError::Validation {
                            message: format!(
                                "HAVING condition references non-selected aggregate: {agg_name}"
                            ),
                            path:    None,
                        })?
                        .clone();

                    conditions.push(HavingCondition {
                        aggregate,
                        operator,
                        value: value.clone(),
                    });
                }
            }
        }

        Ok(conditions)
    }

    /// Parse HAVING key to extract aggregate name and operator
    fn parse_having_key(key: &str) -> Option<(&str, HavingOperator)> {
        for (suffix, op) in &[
            ("_gt", HavingOperator::Gt),
            ("_gte", HavingOperator::Gte),
            ("_lt", HavingOperator::Lt),
            ("_lte", HavingOperator::Lte),
            ("_eq", HavingOperator::Eq),
            ("_neq", HavingOperator::Neq),
        ] {
            if let Some(agg_name) = key.strip_suffix(suffix) {
                return Some((agg_name, *op));
            }
        }
        None
    }

    /// Parse ORDER BY clauses
    fn parse_order_by(order_obj: &Value) -> Result<Vec<OrderByClause>> {
        let mut clauses = Vec::new();

        if let Some(obj) = order_obj.as_object() {
            for (field, value) in obj {
                #[allow(clippy::match_same_arms)]
                // Reason: explicit ASC arm documents accepted values; wildcard handles malformed
                // input
                let direction = match value.as_str() {
                    Some("ASC" | "asc") => OrderDirection::Asc,
                    Some("DESC" | "desc") => OrderDirection::Desc,
                    _ => OrderDirection::Asc, // Default to ASC
                };

                clauses.push(OrderByClause::new(field.clone(), direction));
            }
        }

        Ok(clauses)
    }
}

/// Validate that a regular-JSONB-dimension GraphQL key is a safe SQL identifier
/// (`[_A-Za-z][_0-9A-Za-z]*`).
///
/// Regular dimensions echo this key verbatim as the SELECT-list alias
/// (`… AS {key}`), interpolated into SQL with no quoting; a key containing
/// commas, parentheses, or a subquery would inject an extra SELECT column (H1).
/// Calendar/temporal/native dimensions are gated by lookups against server-
/// defined metadata, so only this fallback path needs the charset guard.
/// Rejecting (rather than quoting) keeps the response key — which equals the SQL
/// column name — unchanged for legitimate identifiers.
///
/// # Errors
///
/// Returns [`FraiseQLError::Validation`] if `key` is empty or contains a
/// character outside `[_A-Za-z][_0-9A-Za-z]*`.
fn validate_dimension_key(key: &str) -> Result<()> {
    let mut chars = key.chars();
    let first_ok = chars.next().is_some_and(|c| c.is_ascii_alphabetic() || c == '_');
    let rest_ok = chars.all(|c| c.is_ascii_alphanumeric() || c == '_');
    if first_ok && rest_ok {
        Ok(())
    } else {
        Err(FraiseQLError::Validation {
            message: format!(
                "groupBy dimension '{key}' contains invalid characters; \
                 only [_A-Za-z][_0-9A-Za-z]* is allowed"
            ),
            path:    None,
        })
    }
}

/// Split a fact-table `where` key, `<field>_<operator>`, against the operator table.
///
/// Shared by the aggregate and window parsers. Splitting at the last underscore could not
/// reach an operator spelled with one (`descendant_of`, `is_not_null`, `depth_eq`); the
/// aggregate parser refused such keys and the window parser dropped them, widening the row
/// set (#1460, #1499). Every split whose suffix names an operator is a candidate. When
/// several are (`path_depth_eq` reads as `path` + `depth_eq` or as `path_depth` + `eq`),
/// the one whose field the fact table declares (a denormalized filter, a native column or a
/// dimension path) wins, and a key that stays ambiguous is refused rather than guessed.
///
/// # Errors
///
/// `FraiseQLError::Validation` naming the key when no split names an operator, or when the
/// key stays ambiguous.
/// The native column a `native_dimension_mapping` maps the dimension `key` to (#1231).
///
/// Keys compare in `snake_case`, segment by segment: a mapping declared `itemCategory` and
/// one declared `item_category` both serve a request for either spelling. Looking the raw
/// request key up instead made a mapping in the other casing never fire, silently.
pub(super) fn mapped_column<'m>(metadata: &'m FactTableMetadata, key: &str) -> Option<&'m String> {
    let wanted = crate::compiler::fact_table::dimension_key(key);
    metadata
        .native_dimension_mapping
        .iter()
        .find(|(declared, _)| crate::compiler::fact_table::dimension_key(declared) == wanted)
        .map(|(_, column)| column)
}

pub(super) fn split_where_key<'k>(
    key: &'k str,
    metadata: &FactTableMetadata,
    native_columns: &std::collections::HashMap<String, String>,
) -> Result<(&'k str, WhereOperator)> {
    let is_declared = |field: &str| {
        let snake = crate::utils::to_snake_case(field);
        native_columns.contains_key(field)
            || metadata.denormalized_filters.iter().any(|f| f.name == field || f.name == snake)
            || metadata.dimensions.paths.iter().any(|p| p.name == field || p.name == snake)
    };
    let candidates: Vec<(&str, &str, WhereOperator)> = key
        .match_indices('_')
        .filter_map(|(i, _)| {
            let (field, suffix) = (&key[..i], &key[i + 1..]);
            let operator = WhereOperator::from_str(suffix).ok()?;
            (!field.is_empty()).then_some((field, suffix, operator))
        })
        .collect();
    let declared: Vec<_> = candidates.iter().filter(|(f, ..)| is_declared(f)).collect();
    let chosen = match (candidates.as_slice(), declared.as_slice()) {
        ([], _) => {
            return Err(FraiseQLError::validation(format!(
                "`where` key `{key}` is not `<field>_<operator>` with a supported operator"
            )));
        },
        ([only], _) => only,
        (_, [only]) => *only,
        (many, _) => {
            let readings: Vec<String> =
                many.iter().map(|(f, op, _)| format!("`{f}` `{op}`")).collect();
            return Err(FraiseQLError::validation(format!(
                "`where` key `{key}` is ambiguous: it reads as {}; declare the field as a \
                 filter column or a dimension path",
                readings.join(" or ")
            )));
        },
    };
    let (field, _, operator) = chosen;
    Ok((field, operator.clone()))
}

#[cfg(test)]
mod alias_injection_tests {
    #![allow(clippy::unwrap_used, clippy::panic)] // Reason: test code, panics are acceptable
    use std::collections::HashMap;

    use super::*;

    fn empty_metadata() -> FactTableMetadata {
        // `dimensions.paths` is empty on purpose: the compiler's dimension
        // allowlist is skipped in that state (the realistic empty-table-at-
        // compile-time default), so parse-time validation is the live guard.
        serde_json::from_value(serde_json::json!({
            "table_name": "tf_sales",
            "measures": [],
            "dimensions": { "name": "dimensions", "paths": [] },
            "denormalized_filters": []
        }))
        .expect("valid empty fact-table metadata")
    }

    #[test]
    fn validate_dimension_key_accepts_plain_identifiers() {
        assert!(validate_dimension_key("category").is_ok());
        assert!(validate_dimension_key("occurred_at_day").is_ok());
        assert!(validate_dimension_key("_private").is_ok());
    }

    #[test]
    fn validate_dimension_key_rejects_injection_shapes() {
        assert!(validate_dimension_key("a, (SELECT 1) AS x").is_err());
        assert!(validate_dimension_key("x; DROP TABLE t").is_err());
        assert!(validate_dimension_key("1leading_digit").is_err());
        assert!(validate_dimension_key("dotted.path").is_err());
        assert!(validate_dimension_key("").is_err());
    }

    #[test]
    fn parse_rejects_hostile_group_by_key_with_empty_allowlist() {
        // The H1 exploit: an alias that appends a scalar-subquery column. With
        // `dimensions.paths` empty the compiler allowlist never fires, so parse-
        // time validation must reject this on its own.
        let query = serde_json::json!({
            "table": "tf_sales",
            "groupBy": {
                "a, (SELECT string_agg(rolpassword, ',') FROM pg_authid) AS leak": true
            }
        });
        let err = AggregateQueryParser::parse(&query, &empty_metadata(), &HashMap::new())
            .expect_err("hostile groupBy key must be rejected at parse time");
        assert!(matches!(err, FraiseQLError::Validation { .. }), "got {err:?}");
    }

    #[test]
    fn parse_accepts_plain_group_by_key() {
        let query = serde_json::json!({
            "table": "tf_sales",
            "groupBy": { "category": true }
        });
        let req = AggregateQueryParser::parse(&query, &empty_metadata(), &HashMap::new())
            .expect("plain groupBy key must parse");
        assert_eq!(req.group_by.len(), 1);
    }

    #[test]
    fn camel_group_by_dimension_recases_path_but_keeps_alias() {
        // #486: a camelCase groupBy dimension must extract the snake_case JSONB key
        // (`data->>'machine_status'`) so grouping actually buckets rows, while the
        // alias stays the camel surface name because `AggregationProjector` uses it
        // verbatim as the GraphQL response key (#418/#410 "output key is the alias").
        let query = serde_json::json!({
            "table": "tf_sales",
            "groupBy": { "machineStatus": true }
        });
        let req = AggregateQueryParser::parse(&query, &empty_metadata(), &HashMap::new())
            .expect("camel groupBy key must parse");
        match req.group_by.as_slice() {
            [GroupBySelection::Dimension { path, alias }] => {
                assert_eq!(path, "machine_status", "path is the snake JSONB key");
                assert_eq!(alias, "machineStatus", "alias is the camel response key");
            },
            other => panic!("expected a single Dimension, got {other:?}"),
        }
    }
}