snowflake-ml-python 1.22.0__py3-none-any.whl → 1.24.0__py3-none-any.whl

snowflake/ml/feature_store/tile_sql_generator.py

@@ -0,0 +1,1079 @@
+"""SQL generators for tile-based aggregations.
+
+This module provides SQL generation for:
+1. TilingSqlGenerator: Creates the DT query that computes partial aggregations (tiles)
+2. MergingSqlGenerator: Creates the CTEs for merging tiles during dataset generation
+"""
+
+from __future__ import annotations
+
+from snowflake.ml.feature_store.aggregation import (
+    AggregationSpec,
+    AggregationType,
+    interval_to_seconds,
+    parse_interval,
+)
+
+# Maximum number of elements to store in array columns to avoid 128MB limit
+# Assuming ~1KB per value, 100,000 values ≈ 100MB (leaving buffer)
+_MAX_ARRAY_ELEMENTS = 100000
+
+
+class TilingSqlGenerator:
+    """Generates SQL for creating tile Dynamic Tables.
+
+    The tiling query:
+    1. Computes TIME_SLICE to bucket rows into tiles
+    2. Computes partial aggregations per (join_keys, tile_start)
+    3. For simple aggregations: stores SUM/COUNT as scalars
+    4. For list aggregations: stores pre-sorted arrays (ARRAY_AGG with ORDER BY)
+    """
+
+    def __init__(
+        self,
+        source_query: str,
+        join_keys: list[str],
+        timestamp_col: str,
+        feature_granularity: str,
+        features: list[AggregationSpec],
+    ) -> None:
+        """Initialize the TilingSqlGenerator.
+
+        Args:
+            source_query: The source query providing raw event data.
+            join_keys: List of column names used as join keys (from entities).
+            timestamp_col: The timestamp column name.
+            feature_granularity: The tile interval (e.g., "1h", "1d").
+            features: List of aggregation specifications.
+        """
+        self._source_query = source_query
+        self._join_keys = join_keys
+        self._timestamp_col = timestamp_col
+        self._feature_granularity = feature_granularity
+        self._features = features
+
+        # Parse interval for SQL generation
+        self._interval_value, self._interval_unit = parse_interval(feature_granularity)
+
+        # Track if we have any lifetime features (need cumulative columns)
+        self._has_lifetime_features = any(f.is_lifetime() for f in features)
+
+    def generate(self) -> str:
+        """Generate the complete tiling SQL query.
+
+        Returns:
+            SQL query for creating the tile Dynamic Table.
+        """
+        tile_columns = self._generate_tile_columns()
+        # Join keys and timestamp_col are already properly formatted by SqlIdentifier
+        join_keys_str = ", ".join(self._join_keys)
+        ts_col = self._timestamp_col
+
+        if not self._has_lifetime_features:
+            # Simple case: no lifetime features, just partial aggregations
+            query = f"""
+SELECT
+    {join_keys_str},
+    TIME_SLICE({ts_col}, {self._interval_value}, '{self._interval_unit}', 'START') AS TILE_START,
+    {', '.join(tile_columns)}
+FROM ({self._source_query})
+GROUP BY {join_keys_str}, TILE_START
+"""
+        else:
+            # With lifetime features: need cumulative columns via window functions
+            # Structure: SELECT *, cumulative_columns FROM (SELECT partial_columns GROUP BY)
+            cumulative_columns = self._generate_cumulative_columns()
+
+            query = f"""
+SELECT
+    base.*,
+    {', '.join(cumulative_columns)}
+FROM (
+    SELECT
+        {join_keys_str},
+        TIME_SLICE({ts_col}, {self._interval_value}, '{self._interval_unit}', 'START') AS TILE_START,
+        {', '.join(tile_columns)}
+    FROM ({self._source_query})
+    GROUP BY {join_keys_str}, TILE_START
+) base
+"""
+        return query.strip()
+
+    def _generate_tile_columns(self) -> list[str]:
+        """Generate the tile column expressions for all features.
+
+        All simple aggregations share base partial columns:
+        - _PARTIAL_SUM_{col}: SUM(col) - used by SUM, AVG, STD, VAR
+        - _PARTIAL_COUNT_{col}: COUNT(col) - used by COUNT, AVG, STD, VAR
+        - _PARTIAL_SUM_SQ_{col}: SUM(col*col) - used by STD, VAR
+
+        This allows maximum column reuse. For example, SUM(amount) + AVG(amount)
+        only creates 2 columns, not 3.
+
+        Returns:
+            List of SQL column expressions for the tile table.
+        """
+        # Track unique tile columns by their full name to avoid duplicates
+        seen_columns: set[str] = set()
+        columns = []
+        ts_col = self._timestamp_col
+
+        for spec in self._features:
+            src_col = spec.source_column
+
+            if spec.function == AggregationType.SUM:
+                # SUM needs _PARTIAL_SUM
+                col_name = spec.get_tile_column_name("SUM")
+                if col_name not in seen_columns:
+                    columns.append(f"SUM({src_col}) AS {col_name}")
+                    seen_columns.add(col_name)
+
+            elif spec.function == AggregationType.COUNT:
+                # COUNT needs _PARTIAL_COUNT
+                col_name = spec.get_tile_column_name("COUNT")
+                if col_name not in seen_columns:
+                    columns.append(f"COUNT({src_col}) AS {col_name}")
+                    seen_columns.add(col_name)
+
+            elif spec.function == AggregationType.AVG:
+                # AVG needs _PARTIAL_SUM and _PARTIAL_COUNT
+                sum_col = spec.get_tile_column_name("SUM")
+                count_col = spec.get_tile_column_name("COUNT")
+                if sum_col not in seen_columns:
+                    columns.append(f"SUM({src_col}) AS {sum_col}")
+                    seen_columns.add(sum_col)
+                if count_col not in seen_columns:
+                    columns.append(f"COUNT({src_col}) AS {count_col}")
+                    seen_columns.add(count_col)
+
+            elif spec.function == AggregationType.MIN:
+                # MIN needs _PARTIAL_MIN
+                col_name = spec.get_tile_column_name("MIN")
+                if col_name not in seen_columns:
+                    columns.append(f"MIN({src_col}) AS {col_name}")
+                    seen_columns.add(col_name)
+
+            elif spec.function == AggregationType.MAX:
+                # MAX needs _PARTIAL_MAX
+                col_name = spec.get_tile_column_name("MAX")
+                if col_name not in seen_columns:
+                    columns.append(f"MAX({src_col}) AS {col_name}")
+                    seen_columns.add(col_name)
+
+            elif spec.function in (AggregationType.STD, AggregationType.VAR):
+                # STD/VAR need _PARTIAL_SUM, _PARTIAL_COUNT, and _PARTIAL_SUM_SQ
+                sum_col = spec.get_tile_column_name("SUM")
+                count_col = spec.get_tile_column_name("COUNT")
+                sum_sq_col = spec.get_tile_column_name("SUM_SQ")
+                if sum_col not in seen_columns:
+                    columns.append(f"SUM({src_col}) AS {sum_col}")
+                    seen_columns.add(sum_col)
+                if count_col not in seen_columns:
+                    columns.append(f"COUNT({src_col}) AS {count_col}")
+                    seen_columns.add(count_col)
+                if sum_sq_col not in seen_columns:
+                    columns.append(f"SUM({src_col} * {src_col}) AS {sum_sq_col}")
+                    seen_columns.add(sum_sq_col)
+
+            elif spec.function == AggregationType.APPROX_COUNT_DISTINCT:
+                # APPROX_COUNT_DISTINCT uses HLL (HyperLogLog) state
+                col_name = spec.get_tile_column_name("HLL")
+                if col_name not in seen_columns:
+                    columns.append(f"HLL_EXPORT(HLL_ACCUMULATE({src_col})) AS {col_name}")
+                    seen_columns.add(col_name)
+
+            elif spec.function == AggregationType.APPROX_PERCENTILE:
+                # APPROX_PERCENTILE uses T-Digest state
+                col_name = spec.get_tile_column_name("TDIGEST")
+                if col_name not in seen_columns:
+                    columns.append(f"APPROX_PERCENTILE_ACCUMULATE({src_col}) AS {col_name}")
+                    seen_columns.add(col_name)
+
+            elif spec.function == AggregationType.LAST_N:
+                col_name = spec.get_tile_column_name("LAST")
+                if col_name not in seen_columns:
+                    columns.append(
+                        f"ARRAY_SLICE("
+                        f"ARRAY_AGG({src_col}) WITHIN GROUP (ORDER BY {ts_col} DESC), "
+                        f"0, {_MAX_ARRAY_ELEMENTS}) AS {col_name}"
+                    )
+                    seen_columns.add(col_name)
+
+            elif spec.function == AggregationType.LAST_DISTINCT_N:
+                # Uses same tile column as LAST_N (dedup happens at merge time)
+                col_name = spec.get_tile_column_name("LAST")
+                if col_name not in seen_columns:
+                    columns.append(
+                        f"ARRAY_SLICE("
+                        f"ARRAY_AGG({src_col}) WITHIN GROUP (ORDER BY {ts_col} DESC), "
+                        f"0, {_MAX_ARRAY_ELEMENTS}) AS {col_name}"
+                    )
+                    seen_columns.add(col_name)
+
+            elif spec.function == AggregationType.FIRST_N:
+                col_name = spec.get_tile_column_name("FIRST")
+                if col_name not in seen_columns:
+                    columns.append(
+                        f"ARRAY_SLICE("
+                        f"ARRAY_AGG({src_col}) WITHIN GROUP (ORDER BY {ts_col} ASC), "
+                        f"0, {_MAX_ARRAY_ELEMENTS}) AS {col_name}"
+                    )
+                    seen_columns.add(col_name)
+
+            elif spec.function == AggregationType.FIRST_DISTINCT_N:
+                # Uses same tile column as FIRST_N (dedup happens at merge time)
+                col_name = spec.get_tile_column_name("FIRST")
+                if col_name not in seen_columns:
+                    columns.append(
+                        f"ARRAY_SLICE("
+                        f"ARRAY_AGG({src_col}) WITHIN GROUP (ORDER BY {ts_col} ASC), "
+                        f"0, {_MAX_ARRAY_ELEMENTS}) AS {col_name}"
+                    )
+                    seen_columns.add(col_name)
+
+        return columns
+
+    def _generate_cumulative_columns(self) -> list[str]:
+        """Generate cumulative column expressions for lifetime aggregations.
+
+        These columns use window functions to compute running totals per entity.
+        They are computed over the partial columns from the inner GROUP BY.
+
+        Returns:
+            List of SQL column expressions for cumulative columns.
+        """
+        seen_columns: set[str] = set()
+        columns = []
+        join_keys_str = ", ".join(self._join_keys)
+
+        # Only process lifetime features
+        lifetime_features = [f for f in self._features if f.is_lifetime()]
+
+        for spec in lifetime_features:
+            if spec.function == AggregationType.SUM:
+                # Cumulative SUM
+                partial_col = spec.get_tile_column_name("SUM")
+                cum_col = spec.get_cumulative_column_name("SUM")
+                if cum_col not in seen_columns:
+                    columns.append(
+                        f"SUM({partial_col}) OVER ("
+                        f"PARTITION BY {join_keys_str} "
+                        f"ORDER BY TILE_START "
+                        f"ROWS BETWEEN UNBOUNDED PRECEDING AND CURRENT ROW"
+                        f") AS {cum_col}"
+                    )
+                    seen_columns.add(cum_col)
+
+            elif spec.function == AggregationType.COUNT:
+                # Cumulative COUNT
+                partial_col = spec.get_tile_column_name("COUNT")
+                cum_col = spec.get_cumulative_column_name("COUNT")
+                if cum_col not in seen_columns:
+                    columns.append(
+                        f"SUM({partial_col}) OVER ("
+                        f"PARTITION BY {join_keys_str} "
+                        f"ORDER BY TILE_START "
+                        f"ROWS BETWEEN UNBOUNDED PRECEDING AND CURRENT ROW"
+                        f") AS {cum_col}"
+                    )
+                    seen_columns.add(cum_col)
+
+            elif spec.function == AggregationType.AVG:
+                # Cumulative AVG needs cumulative SUM and COUNT
+                partial_sum = spec.get_tile_column_name("SUM")
+                partial_count = spec.get_tile_column_name("COUNT")
+                cum_sum = spec.get_cumulative_column_name("SUM")
+                cum_count = spec.get_cumulative_column_name("COUNT")
+
+                if cum_sum not in seen_columns:
+                    columns.append(
+                        f"SUM({partial_sum}) OVER ("
+                        f"PARTITION BY {join_keys_str} "
+                        f"ORDER BY TILE_START "
+                        f"ROWS BETWEEN UNBOUNDED PRECEDING AND CURRENT ROW"
+                        f") AS {cum_sum}"
+                    )
+                    seen_columns.add(cum_sum)
+                if cum_count not in seen_columns:
+                    columns.append(
+                        f"SUM({partial_count}) OVER ("
+                        f"PARTITION BY {join_keys_str} "
+                        f"ORDER BY TILE_START "
+                        f"ROWS BETWEEN UNBOUNDED PRECEDING AND CURRENT ROW"
+                        f") AS {cum_count}"
+                    )
+                    seen_columns.add(cum_count)
+
+            elif spec.function == AggregationType.MIN:
+                # Cumulative MIN (running minimum)
+                partial_col = spec.get_tile_column_name("MIN")
+                cum_col = spec.get_cumulative_column_name("MIN")
+                if cum_col not in seen_columns:
+                    columns.append(
+                        f"MIN({partial_col}) OVER ("
+                        f"PARTITION BY {join_keys_str} "
+                        f"ORDER BY TILE_START "
+                        f"ROWS BETWEEN UNBOUNDED PRECEDING AND CURRENT ROW"
+                        f") AS {cum_col}"
+                    )
+                    seen_columns.add(cum_col)
+
+            elif spec.function == AggregationType.MAX:
+                # Cumulative MAX (running maximum)
+                partial_col = spec.get_tile_column_name("MAX")
+                cum_col = spec.get_cumulative_column_name("MAX")
+                if cum_col not in seen_columns:
+                    columns.append(
+                        f"MAX({partial_col}) OVER ("
+                        f"PARTITION BY {join_keys_str} "
+                        f"ORDER BY TILE_START "
+                        f"ROWS BETWEEN UNBOUNDED PRECEDING AND CURRENT ROW"
+                        f") AS {cum_col}"
+                    )
+                    seen_columns.add(cum_col)
+
+            elif spec.function in (AggregationType.STD, AggregationType.VAR):
+                # Cumulative STD/VAR needs cumulative SUM, COUNT, and SUM_SQ
+                partial_sum = spec.get_tile_column_name("SUM")
+                partial_count = spec.get_tile_column_name("COUNT")
+                partial_sum_sq = spec.get_tile_column_name("SUM_SQ")
+                cum_sum = spec.get_cumulative_column_name("SUM")
+                cum_count = spec.get_cumulative_column_name("COUNT")
+                cum_sum_sq = spec.get_cumulative_column_name("SUM_SQ")
+
+                if cum_sum not in seen_columns:
+                    columns.append(
+                        f"SUM({partial_sum}) OVER ("
+                        f"PARTITION BY {join_keys_str} "
+                        f"ORDER BY TILE_START "
+                        f"ROWS BETWEEN UNBOUNDED PRECEDING AND CURRENT ROW"
+                        f") AS {cum_sum}"
+                    )
+                    seen_columns.add(cum_sum)
+                if cum_count not in seen_columns:
+                    columns.append(
+                        f"SUM({partial_count}) OVER ("
+                        f"PARTITION BY {join_keys_str} "
+                        f"ORDER BY TILE_START "
+                        f"ROWS BETWEEN UNBOUNDED PRECEDING AND CURRENT ROW"
+                        f") AS {cum_count}"
+                    )
+                    seen_columns.add(cum_count)
+                if cum_sum_sq not in seen_columns:
+                    columns.append(
+                        f"SUM({partial_sum_sq}) OVER ("
+                        f"PARTITION BY {join_keys_str} "
+                        f"ORDER BY TILE_START "
+                        f"ROWS BETWEEN UNBOUNDED PRECEDING AND CURRENT ROW"
+                        f") AS {cum_sum_sq}"
+                    )
+                    seen_columns.add(cum_sum_sq)
+
+            # Note: APPROX_COUNT_DISTINCT (HLL_COMBINE) and APPROX_PERCENTILE (APPROX_PERCENTILE_COMBINE)
+            # do NOT support cumulative window frames in Snowflake.
+            # These will be handled at merge time by aggregating all tiles.
+
+            # Note: FIRST_N, FIRST_DISTINCT_N, LAST_N, LAST_DISTINCT_N lifetime
+            # are handled at merge time by scanning tiles, not via cumulative columns
+
+        return columns
+
+
+class MergingSqlGenerator:
+    """Generates CTEs for merging tiles during dataset generation.
+
+    The merging process:
+    1. TILES_JOINED_FVi: Join tiles with spine, filtering by window and complete tiles only
+    2. SIMPLE_MERGED_FVi: Aggregate simple features (SUM, COUNT, AVG)
+    3. LIST_MERGED_FVi: Flatten and aggregate list features (LAST_N, etc.)
+    4. FVi: Combine simple and list results
+    """
+
+    def __init__(
+        self,
+        tile_table: str,
+        join_keys: list[str],
+        timestamp_col: str,
+        feature_granularity: str,
+        features: list[AggregationSpec],
+        spine_timestamp_col: str,
+        fv_index: int,
+    ) -> None:
+        """Initialize the MergingSqlGenerator.
+
+        Args:
+            tile_table: Fully qualified name of the tile table.
+            join_keys: List of join key column names.
+            timestamp_col: The timestamp column from the feature view.
+            feature_granularity: The tile interval.
+            features: List of aggregation specifications.
+            spine_timestamp_col: The timestamp column from the spine.
+            fv_index: Index of this feature view (for CTE naming).
+        """
+        self._tile_table = tile_table
+        self._join_keys = join_keys
+        self._timestamp_col = timestamp_col
+        self._feature_granularity = feature_granularity
+        self._features = features
+        self._spine_timestamp_col = spine_timestamp_col
+        self._fv_index = fv_index
+
+        # Separate lifetime from non-lifetime features
+        self._lifetime_features = [f for f in features if f.is_lifetime()]
+        self._non_lifetime_features = [f for f in features if not f.is_lifetime()]
+
+        # Separate non-lifetime features by type (simple vs list)
+        self._simple_features = [f for f in self._non_lifetime_features if f.function.is_simple()]
+        self._list_features = [f for f in self._non_lifetime_features if f.function.is_list()]
+
+        # Lifetime features are all simple (validation ensures only SUM, COUNT, AVG, MIN, MAX, STD, VAR)
+        self._lifetime_simple_features = self._lifetime_features
+
+        # Parse interval
+        self._interval_value, self._interval_unit = parse_interval(feature_granularity)
+        self._interval_seconds = interval_to_seconds(feature_granularity)
+
+        # Calculate max window in tiles for filtering (only for non-lifetime features)
+        if self._non_lifetime_features:
+            # Max tiles needed is the max of (window + offset) across all non-lifetime features
+            max_lookback_seconds = max(
+                f.get_window_seconds() + f.get_offset_seconds() for f in self._non_lifetime_features
+            )
+            self._max_tiles_needed = (max_lookback_seconds + self._interval_seconds - 1) // self._interval_seconds
+        else:
+            self._max_tiles_needed = 0
+
+    def generate_all_ctes(self) -> list[tuple[str, str]]:
+        """Generate all CTEs needed for this feature view.
+
+        The optimization flow:
+        1. SPINE_BOUNDARY: Add truncated tile boundary to spine
+        2. UNIQUE_BOUNDS: Get distinct (entity, boundary) pairs
+        3. TILES_JOINED: Join tiles to unique boundaries (for non-lifetime features)
+        4. SIMPLE_MERGED: Aggregate simple non-lifetime features per boundary
+        5. LIST_MERGED: Aggregate list non-lifetime features per boundary
+        6. LIFETIME_MERGED: ASOF JOIN for lifetime features (O(1) per boundary)
+        7. LIFETIME_LIST_MERGED: Scan tiles for lifetime list features
+        8. FV: Join back to spine to expand results
+
+        Returns:
+            List of (cte_name, cte_body) tuples.
+        """
+        ctes = []
+
+        # CTE 1: Spine with tile boundary (for join-back)
+        ctes.append(self._generate_spine_boundary_cte())
+
+        # CTE 2: Unique boundaries (optimization - reduce aggregation work)
+        ctes.append(self._generate_unique_boundaries_cte())
+
+        # CTE 3: Join tiles with unique boundaries (only if we have non-lifetime features)
+        if self._non_lifetime_features:
+            ctes.append(self._generate_tiles_joined_cte())
+
+        # CTE 4: Simple aggregations (if any non-lifetime)
+        if self._simple_features:
+            ctes.append(self._generate_simple_merged_cte())
+
+        # CTE 5: List aggregations (if any non-lifetime)
+        if self._list_features:
+            ctes.append(self._generate_list_merged_cte())
+
+        # CTE 6: Lifetime simple aggregations (using ASOF JOIN on cumulative columns)
+        if self._lifetime_simple_features:
+            ctes.append(self._generate_lifetime_merged_cte())
+
+        # CTE 7: Combine all results and join back to spine
+        ctes.append(self._generate_combined_cte())
+
+        return ctes
+
+    def _generate_spine_boundary_cte(self) -> tuple[str, str]:
+        """Generate CTE that adds tile boundary to deduplicated spine.
+
+        The tile boundary is the truncated timestamp that determines which
+        complete tiles are visible. All spine rows with the same boundary
+        will have identical feature values.
+
+        Returns:
+            Tuple of (cte_name, cte_body).
+        """
+        cte_name = f"SPINE_BOUNDARY_FV{self._fv_index}"
+
+        # Quote column names to preserve case-sensitivity from spine dataframe
+        # The spine_timestamp_col is passed as-is from the user (e.g., "query_ts")
+        quoted_join_keys = [f'"{k}"' for k in self._join_keys]
+        quoted_spine_ts = f'"{self._spine_timestamp_col}"'
+
+        # Select all columns plus the tile boundary
+        all_cols = quoted_join_keys + [quoted_spine_ts]
+        select_cols = ", ".join(all_cols)
+
+        # DATE_TRUNC to tile granularity gives us the boundary
+        # All timestamps in the same granularity window see the same complete tiles
+        cte_body = f"""
+    SELECT DISTINCT {select_cols},
+           DATE_TRUNC('{self._interval_unit.lower()}', {quoted_spine_ts}) AS TILE_BOUNDARY
+    FROM SPINE
+"""
+        return cte_name, cte_body.strip()
+
+    def _generate_unique_boundaries_cte(self) -> tuple[str, str]:
+        """Generate CTE with unique (entity, boundary) pairs.
+
+        This is the key optimization: instead of computing features for each
+        spine row, we compute once per unique boundary and join back.
+
+        Returns:
+            Tuple of (cte_name, cte_body).
+        """
+        cte_name = f"UNIQUE_BOUNDS_FV{self._fv_index}"
+
+        # Quote column names to handle case-sensitivity
+        quoted_join_keys = [f'"{k}"' for k in self._join_keys]
+        keys_str = ", ".join(quoted_join_keys)
+
+        cte_body = f"""
+    SELECT DISTINCT {keys_str}, TILE_BOUNDARY
+    FROM SPINE_BOUNDARY_FV{self._fv_index}
+"""
+        return cte_name, cte_body.strip()
+
+    def _generate_tiles_joined_cte(self) -> tuple[str, str]:
+        """Generate the CTE that joins tiles with unique boundaries.
+
+        This joins tiles to UNIQUE_BOUNDS (not full spine), which is much smaller
+        when there are many spine rows per tile boundary.
+
+        Returns:
+            Tuple of (cte_name, cte_body) for the tiles joined CTE.
+        """
+        cte_name = f"TILES_JOINED_FV{self._fv_index}"
+
+        # Quote column names for spine columns (case-sensitive)
+        quoted_join_keys = [f'"{k}"' for k in self._join_keys]
+
+        # Tile table column names match the join keys (already SqlIdentifier-formatted)
+        tile_join_keys = list(self._join_keys)
+        tile_keys_str = ", ".join(tile_join_keys)
+
+        # Join conditions: quoted spine columns to uppercase tile columns
+        join_conditions = [f"UB.{qk} = TILES.{tk}" for qk, tk in zip(quoted_join_keys, tile_join_keys)]
+
+        # Get all tile columns we need (deduplicated)
+        tile_columns_set: set[str] = set()
+        for spec in self._features:
+            if spec.function == AggregationType.SUM:
+                tile_columns_set.add(spec.get_tile_column_name("SUM"))
+            elif spec.function == AggregationType.COUNT:
+                tile_columns_set.add(spec.get_tile_column_name("COUNT"))
+            elif spec.function == AggregationType.AVG:
+                tile_columns_set.add(spec.get_tile_column_name("SUM"))
+                tile_columns_set.add(spec.get_tile_column_name("COUNT"))
+            elif spec.function == AggregationType.MIN:
+                tile_columns_set.add(spec.get_tile_column_name("MIN"))
+            elif spec.function == AggregationType.MAX:
+                tile_columns_set.add(spec.get_tile_column_name("MAX"))
+            elif spec.function in (AggregationType.STD, AggregationType.VAR):
+                tile_columns_set.add(spec.get_tile_column_name("SUM"))
+                tile_columns_set.add(spec.get_tile_column_name("COUNT"))
+                tile_columns_set.add(spec.get_tile_column_name("SUM_SQ"))
+            elif spec.function == AggregationType.APPROX_COUNT_DISTINCT:
+                tile_columns_set.add(spec.get_tile_column_name("HLL"))
+            elif spec.function == AggregationType.APPROX_PERCENTILE:
+                tile_columns_set.add(spec.get_tile_column_name("TDIGEST"))
+            elif spec.function in (AggregationType.LAST_N, AggregationType.LAST_DISTINCT_N):
+                tile_columns_set.add(spec.get_tile_column_name("LAST"))
+            elif spec.function in (AggregationType.FIRST_N, AggregationType.FIRST_DISTINCT_N):
+                tile_columns_set.add(spec.get_tile_column_name("FIRST"))
+        tile_columns = sorted(tile_columns_set)  # Sort for deterministic output
+
+        tile_columns_str = ", ".join(f"TILES.{col}" for col in tile_columns)
+
+        # Window filter: only include tiles within the max window and complete tiles
+        # Complete tiles: tile_end <= tile_boundary (not spine timestamp)
+        # tile_end = DATEADD(interval_unit, interval_value, tile_start)
+        cte_body = f"""
+    SELECT
+        UB.*,
+        TILES.TILE_START,
+        {tile_columns_str}
+    FROM UNIQUE_BOUNDS_FV{self._fv_index} UB
+    LEFT JOIN (
+        SELECT {tile_keys_str}, TILE_START, {', '.join(tile_columns)}
+        FROM {self._tile_table}
+    ) TILES
+    ON {' AND '.join(join_conditions)}
+    -- Window filter: tiles within max window from tile boundary
+    AND TILES.TILE_START >= DATEADD(
+        {self._interval_unit}, -{self._max_tiles_needed * self._interval_value}, UB.TILE_BOUNDARY
+    )
+    -- Complete tiles only: tile_end <= tile_boundary
+    AND DATEADD({self._interval_unit}, {self._interval_value}, TILES.TILE_START) <= UB.TILE_BOUNDARY
+"""
+        return cte_name, cte_body.strip()
+
+    def _get_tile_filter_condition(self, spec: AggregationSpec) -> str:
+        """Generate the CASE WHEN condition for filtering tiles by window and offset.
+
+        For a feature with window W and offset O, we want tiles where:
+        - TILE_START >= TILE_BOUNDARY - W - O (start of window)
+        - TILE_START < TILE_BOUNDARY - O (end of window, before offset)
+
+        Args:
+            spec: The aggregation specification.
+
+        Returns:
+            SQL condition string for use in CASE WHEN.
+        """
+        window_tiles = (spec.get_window_seconds() + self._interval_seconds - 1) // self._interval_seconds
+        offset_tiles = spec.get_offset_seconds() // self._interval_seconds
+
+        if offset_tiles == 0:
+            # No offset: just filter by window start
+            return (
+                f"TILE_START >= DATEADD({self._interval_unit}, "
+                f"-{window_tiles * self._interval_value}, TILE_BOUNDARY)"
+            )
+        else:
+            # With offset: filter by both window start and end (shifted by offset)
+            window_start_tiles = window_tiles + offset_tiles
+            return (
+                f"TILE_START >= DATEADD({self._interval_unit}, "
+                f"-{window_start_tiles * self._interval_value}, TILE_BOUNDARY) "
+                f"AND TILE_START < DATEADD({self._interval_unit}, "
+                f"-{offset_tiles * self._interval_value}, TILE_BOUNDARY)"
+            )
+
+    def _get_list_tile_filter_condition(self, spec: AggregationSpec) -> str:
+        """Generate filter condition for list aggregations (with table prefix).
+
+        Similar to _get_tile_filter_condition but uses t. prefix for table references.
+
+        Args:
+            spec: The aggregation specification.
+
+        Returns:
+            SQL condition string for use in WHERE clause.
+        """
+        window_tiles = (spec.get_window_seconds() + self._interval_seconds - 1) // self._interval_seconds
+        offset_tiles = spec.get_offset_seconds() // self._interval_seconds
+
+        if offset_tiles == 0:
+            # No offset: just filter by window start
+            return (
+                f"t.TILE_START >= DATEADD({self._interval_unit}, "
+                f"-{window_tiles * self._interval_value}, t.TILE_BOUNDARY)"
+            )
+        else:
+            # With offset: filter by both window start and end (shifted by offset)
+            window_start_tiles = window_tiles + offset_tiles
+            return (
+                f"t.TILE_START >= DATEADD({self._interval_unit}, "
+                f"-{window_start_tiles * self._interval_value}, t.TILE_BOUNDARY) "
+                f"AND t.TILE_START < DATEADD({self._interval_unit}, "
+                f"-{offset_tiles * self._interval_value}, t.TILE_BOUNDARY)"
+            )
+
+    def _generate_simple_merged_cte(self) -> tuple[str, str]:
+        """Generate the CTE for simple aggregations (SUM, COUNT, AVG).
+
+        Groups by entity keys + TILE_BOUNDARY (not spine timestamp) for efficiency.
+
+        Returns:
+            Tuple of (cte_name, cte_body) for the simple merged CTE.
+        """
+        cte_name = f"SIMPLE_MERGED_FV{self._fv_index}"
+
+        # Quote column names for case-sensitivity (from UNIQUE_BOUNDS which inherits from spine)
+        quoted_join_keys = [f'"{k}"' for k in self._join_keys]
+        # Group by entity keys + TILE_BOUNDARY (optimization)
+        group_by_cols = quoted_join_keys + ["TILE_BOUNDARY"]
+        group_by_str = ", ".join(group_by_cols)
+
+        agg_columns = []
+        for spec in self._simple_features:
+            output_col = spec.get_sql_column_name()
+            tile_filter = self._get_tile_filter_condition(spec)
+
+            if spec.function == AggregationType.SUM:
+                col_name = spec.get_tile_column_name("SUM")
+                agg_columns.append(f"SUM(CASE WHEN {tile_filter} " f"THEN {col_name} ELSE 0 END) AS {output_col}")
+
+            elif spec.function == AggregationType.COUNT:
+                col_name = spec.get_tile_column_name("COUNT")
+                agg_columns.append(f"SUM(CASE WHEN {tile_filter} " f"THEN {col_name} ELSE 0 END) AS {output_col}")
+
+            elif spec.function == AggregationType.MIN:
+                col_name = spec.get_tile_column_name("MIN")
+                agg_columns.append(f"MIN(CASE WHEN {tile_filter} " f"THEN {col_name} ELSE NULL END) AS {output_col}")
+
+            elif spec.function == AggregationType.MAX:
+                col_name = spec.get_tile_column_name("MAX")
+                agg_columns.append(f"MAX(CASE WHEN {tile_filter} " f"THEN {col_name} ELSE NULL END) AS {output_col}")
+
+            elif spec.function == AggregationType.AVG:
+                # AVG = SUM(partial_sums) / SUM(partial_counts)
+                sum_col = spec.get_tile_column_name("SUM")
+                count_col = spec.get_tile_column_name("COUNT")
+                agg_columns.append(
+                    f"CASE WHEN SUM(CASE WHEN {tile_filter} "
+                    f"THEN {count_col} ELSE 0 END) > 0 "
+                    f"THEN SUM(CASE WHEN {tile_filter} "
+                    f"THEN {sum_col} ELSE 0 END) / "
+                    f"SUM(CASE WHEN {tile_filter} "
+                    f"THEN {count_col} ELSE 0 END) "
+                    f"ELSE NULL END AS {output_col}"
+                )
+
+            elif spec.function == AggregationType.VAR:
+                # VAR = (SUM_SQ / COUNT) - (SUM / COUNT)^2
+                # Using parallel variance formula
+                # GREATEST(0, ...) clamps to non-negative to handle floating-point errors
+                sum_col = spec.get_tile_column_name("SUM")
+                count_col = spec.get_tile_column_name("COUNT")
+                sum_sq_col = spec.get_tile_column_name("SUM_SQ")
+                agg_columns.append(
+                    f"CASE WHEN SUM(CASE WHEN {tile_filter} "
+                    f"THEN {count_col} ELSE 0 END) > 0 "
+                    f"THEN GREATEST(0, ("
+                    f"SUM(CASE WHEN {tile_filter} "
+                    f"THEN {sum_sq_col} ELSE 0 END) / "
+                    f"SUM(CASE WHEN {tile_filter} "
+                    f"THEN {count_col} ELSE 0 END)"
+                    f") - POWER("
+                    f"SUM(CASE WHEN {tile_filter} "
+                    f"THEN {sum_col} ELSE 0 END) / "
+                    f"SUM(CASE WHEN {tile_filter} "
+                    f"THEN {count_col} ELSE 0 END), 2)) "
+                    f"ELSE NULL END AS {output_col}"
+                )
+
+            elif spec.function == AggregationType.STD:
+                # STD = SQRT(VAR) = SQRT((SUM_SQ / COUNT) - (SUM / COUNT)^2)
+                # GREATEST(0, ...) clamps variance to non-negative to handle floating-point errors
+                # that can cause sqrt of tiny negative numbers like -4.54747e-13
+                sum_col = spec.get_tile_column_name("SUM")
+                count_col = spec.get_tile_column_name("COUNT")
+                sum_sq_col = spec.get_tile_column_name("SUM_SQ")
+                agg_columns.append(
+                    f"CASE WHEN SUM(CASE WHEN {tile_filter} "
+                    f"THEN {count_col} ELSE 0 END) > 0 "
+                    f"THEN SQRT(GREATEST(0, "
+                    f"SUM(CASE WHEN {tile_filter} "
+                    f"THEN {sum_sq_col} ELSE 0 END) / "
+                    f"SUM(CASE WHEN {tile_filter} "
+                    f"THEN {count_col} ELSE 0 END)"
+                    f" - POWER("
+                    f"SUM(CASE WHEN {tile_filter} "
+                    f"THEN {sum_col} ELSE 0 END) / "
+                    f"SUM(CASE WHEN {tile_filter} "
+                    f"THEN {count_col} ELSE 0 END), 2))) "
+                    f"ELSE NULL END AS {output_col}"
+                )
+
+            elif spec.function == AggregationType.APPROX_COUNT_DISTINCT:
+                # Combine HLL states and estimate count
+                # HLL_ESTIMATE(HLL_COMBINE(HLL_IMPORT(state))) gives the approximate count
+                col_name = spec.get_tile_column_name("HLL")
+                agg_columns.append(
+                    f"HLL_ESTIMATE(HLL_COMBINE("
+                    f"CASE WHEN {tile_filter} THEN HLL_IMPORT({col_name}) ELSE NULL END"
+                    f")) AS {output_col}"
+                )
+
+            elif spec.function == AggregationType.APPROX_PERCENTILE:
+                # Combine T-Digest states and estimate percentile
+                # APPROX_PERCENTILE_ESTIMATE(APPROX_PERCENTILE_COMBINE(state), percentile)
+                col_name = spec.get_tile_column_name("TDIGEST")
+                percentile = spec.params.get("percentile", 0.5)
+                agg_columns.append(
+                    f"APPROX_PERCENTILE_ESTIMATE(APPROX_PERCENTILE_COMBINE("
+                    f"CASE WHEN {tile_filter} THEN {col_name} ELSE NULL END"
+                    f"), {percentile}) AS {output_col}"
+                )
+
+        cte_body = f"""
+    SELECT
+        {group_by_str},
+        {', '.join(agg_columns)}
+    FROM TILES_JOINED_FV{self._fv_index}
+    GROUP BY {group_by_str}
+"""
+        return cte_name, cte_body.strip()
+
+    def _generate_list_merged_cte(self) -> tuple[str, str]:
+        """Generate the CTE for list aggregations using LATERAL FLATTEN."""
+        cte_name = f"LIST_MERGED_FV{self._fv_index}"
+
+        # Generate a more efficient single-pass CTE
+        # Each list feature gets its own lateral flatten in the FROM clause
+        cte_body = self._generate_list_cte_body()
+
+        return cte_name, cte_body.strip()
+
+    def _generate_lifetime_merged_cte(self) -> tuple[str, str]:
+        """Generate the CTE for lifetime simple aggregations using ASOF JOIN.
+
+        Uses ASOF JOIN on cumulative columns for O(1) lookup per boundary.
+        This is much faster than aggregating all tiles from the beginning of time.
+
+        Returns:
+            Tuple of (cte_name, cte_body) for the lifetime merged CTE.
+        """
+        cte_name = f"LIFETIME_MERGED_FV{self._fv_index}"
+
+        # Quote column names for case-sensitivity
+        quoted_join_keys = [f'"{k}"' for k in self._join_keys]
+        group_by_cols = quoted_join_keys + ["TILE_BOUNDARY"]
+
+        # Tile table column names match the join keys (already SqlIdentifier-formatted)
+        tile_join_keys = list(self._join_keys)
+
+        # Build ASOF JOIN condition: match the most recent tile before the boundary
+        asof_match = "UB.TILE_BOUNDARY > TILES.TILE_START"
+        join_conditions = [f"UB.{qk} = TILES.{tk}" for qk, tk in zip(quoted_join_keys, tile_join_keys)]
+        join_conditions_str = " AND ".join(join_conditions)
+
+        # Build select columns for lifetime features
+        select_cols = []
+        for spec in self._lifetime_simple_features:
+            output_col = spec.get_sql_column_name()
+
+            if spec.function == AggregationType.SUM:
+                cum_col = spec.get_cumulative_column_name("SUM")
+                select_cols.append(f"TILES.{cum_col} AS {output_col}")
+
+            elif spec.function == AggregationType.COUNT:
+                cum_col = spec.get_cumulative_column_name("COUNT")
+                select_cols.append(f"TILES.{cum_col} AS {output_col}")
+
+            elif spec.function == AggregationType.AVG:
+                cum_sum = spec.get_cumulative_column_name("SUM")
+                cum_count = spec.get_cumulative_column_name("COUNT")
+                select_cols.append(
+                    f"CASE WHEN TILES.{cum_count} > 0 "
+                    f"THEN TILES.{cum_sum} / TILES.{cum_count} "
+                    f"ELSE NULL END AS {output_col}"
+                )
+
+            elif spec.function == AggregationType.MIN:
+                cum_col = spec.get_cumulative_column_name("MIN")
+                select_cols.append(f"TILES.{cum_col} AS {output_col}")
+
+            elif spec.function == AggregationType.MAX:
+                cum_col = spec.get_cumulative_column_name("MAX")
+                select_cols.append(f"TILES.{cum_col} AS {output_col}")
+
+            elif spec.function == AggregationType.VAR:
+                cum_sum = spec.get_cumulative_column_name("SUM")
+                cum_count = spec.get_cumulative_column_name("COUNT")
+                cum_sum_sq = spec.get_cumulative_column_name("SUM_SQ")
+                select_cols.append(
+                    f"CASE WHEN TILES.{cum_count} > 0 "
+                    f"THEN GREATEST(0, "
+                    f"TILES.{cum_sum_sq} / TILES.{cum_count} "
+                    f"- POWER(TILES.{cum_sum} / TILES.{cum_count}, 2)) "
+                    f"ELSE NULL END AS {output_col}"
+                )
+
+            elif spec.function == AggregationType.STD:
+                cum_sum = spec.get_cumulative_column_name("SUM")
+                cum_count = spec.get_cumulative_column_name("COUNT")
+                cum_sum_sq = spec.get_cumulative_column_name("SUM_SQ")
+                select_cols.append(
+                    f"CASE WHEN TILES.{cum_count} > 0 "
+                    f"THEN SQRT(GREATEST(0, "
+                    f"TILES.{cum_sum_sq} / TILES.{cum_count} "
+                    f"- POWER(TILES.{cum_sum} / TILES.{cum_count}, 2))) "
+                    f"ELSE NULL END AS {output_col}"
+                )
+
+            elif spec.function == AggregationType.APPROX_COUNT_DISTINCT:
+                cum_col = spec.get_cumulative_column_name("HLL")
+                select_cols.append(f"HLL_ESTIMATE(HLL_IMPORT(TILES.{cum_col})) AS {output_col}")
+
+            elif spec.function == AggregationType.APPROX_PERCENTILE:
+                cum_col = spec.get_cumulative_column_name("TDIGEST")
+                percentile = spec.params.get("percentile", 0.5)
+                select_cols.append(f"APPROX_PERCENTILE_ESTIMATE(TILES.{cum_col}, {percentile}) AS {output_col}")
+
+        select_cols_str = ", ".join(select_cols)
+
+        # Qualify group by columns with UB alias to avoid ambiguity in ASOF JOIN
+        qualified_group_cols = [f"UB.{col}" for col in group_by_cols]
+        qualified_group_str = ", ".join(qualified_group_cols)
+
+        cte_body = f"""
+    SELECT
+        {qualified_group_str},
+        {select_cols_str}
+    FROM UNIQUE_BOUNDS_FV{self._fv_index} UB
+    ASOF JOIN {self._tile_table} TILES
+      MATCH_CONDITION ({asof_match})
+      ON {join_conditions_str}
+"""
+        return cte_name, cte_body.strip()
+
+    def _generate_list_cte_body(self) -> str:
+        """Generate the body of the list merged CTE.
+
+        Groups by entity keys + TILE_BOUNDARY (not spine timestamp) for efficiency.
+
+        Returns:
+            SQL string for the list merged CTE body.
+        """
+        # Quote column names for case-sensitivity (from UNIQUE_BOUNDS which inherits from spine)
+        quoted_join_keys = [f'"{k}"' for k in self._join_keys]
+        # Group by entity keys + TILE_BOUNDARY (optimization)
+        group_by_cols = quoted_join_keys + ["TILE_BOUNDARY"]
+        group_by_str = ", ".join(group_by_cols)
+        select_group_cols = ", ".join(f"t.{col}" for col in group_by_cols)
+
+        if not self._list_features:
+            return f"SELECT {group_by_str} FROM TILES_JOINED_FV{self._fv_index} WHERE 1=0"
+
+        # Build subqueries for each list feature
+        feature_subqueries = []
+
+        for _idx, spec in enumerate(self._list_features):
+            # Determine the tile column name based on aggregation type
+            if spec.function in (AggregationType.LAST_N, AggregationType.LAST_DISTINCT_N):
+                col_name = spec.get_tile_column_name("LAST")
+                order_clause = "t.TILE_START DESC, flat.INDEX ASC"
+            else:  # FIRST_N, FIRST_DISTINCT_N
+                col_name = spec.get_tile_column_name("FIRST")
+                order_clause = "t.TILE_START ASC, flat.INDEX ASC"
+
+            output_col = spec.get_sql_column_name()
+            n_value = spec.params["n"]
+
+            # Calculate tile filter condition for window and offset
+            tile_filter = self._get_list_tile_filter_condition(spec)
+
+            is_distinct = spec.function in (AggregationType.LAST_DISTINCT_N, AggregationType.FIRST_DISTINCT_N)
+
+            if is_distinct:
+                # For distinct, use QUALIFY to keep first occurrence of each value
+                # Note: Inner query uses t. prefix, outer query uses bare column names
+                subquery = f"""
+    (SELECT {group_by_str},
+            ARRAY_AGG(val) WITHIN GROUP (ORDER BY rn) AS {output_col}
+     FROM (
+         SELECT {select_group_cols}, flat.VALUE AS val,
+                ROW_NUMBER() OVER (PARTITION BY {select_group_cols} ORDER BY {order_clause}) AS rn,
+                ROW_NUMBER() OVER (PARTITION BY {select_group_cols}, flat.VALUE ORDER BY {order_clause}) AS dup_rn
+         FROM TILES_JOINED_FV{self._fv_index} t,
+         LATERAL FLATTEN(INPUT => t.{col_name}) flat
+         WHERE {tile_filter}
+         AND flat.VALUE IS NOT NULL
+     ) ranked
+     WHERE dup_rn = 1 AND rn <= {n_value}
+     GROUP BY {group_by_str}
+    )"""
+            else:
+                # Non-distinct: straightforward flatten and aggregate
+                subquery = f"""
+    (SELECT {select_group_cols},
+            ARRAY_SLICE(
+                ARRAY_AGG(flat.VALUE) WITHIN GROUP (ORDER BY {order_clause}),
+                0, {n_value}
+            ) AS {output_col}
+     FROM TILES_JOINED_FV{self._fv_index} t,
+     LATERAL FLATTEN(INPUT => t.{col_name}) flat
+     WHERE {tile_filter}
+     AND flat.VALUE IS NOT NULL
+     GROUP BY {select_group_cols}
+    )"""
+
+            feature_subqueries.append((output_col, subquery))
+
+        # Combine all subqueries with JOINs
+        if len(feature_subqueries) == 1:
+            return feature_subqueries[0][1]
+
+        # Multiple list features: join them together
+        first_name, first_query = feature_subqueries[0]
+        result = "SELECT sq0.*, "
+        for i, (name, _) in enumerate(feature_subqueries[1:], 1):
+            result += f"sq{i}.{name}"
+            if i < len(feature_subqueries) - 1:
+                result += ", "
+
+        result += f"\n    FROM {first_query} sq0"
+
+        for i, (_name, query) in enumerate(feature_subqueries[1:], 1):
+            join_cond = " AND ".join(f"sq0.{col} = sq{i}.{col}" for col in group_by_cols)
+            result += f"\n    LEFT JOIN {query} sq{i}\n    ON {join_cond}"
+
+        return result
+
+    def _generate_combined_cte(self) -> tuple[str, str]:
+        """Generate the final CTE that combines and expands results.
+
+        This joins the merged results (grouped by entity + TILE_BOUNDARY) back to
+        the original spine (SPINE_BOUNDARY) to expand to per-spine-row output.
+
+        Returns:
+            Tuple of (cte_name, cte_body) for the combined CTE.
+        """
+        cte_name = f"FV{self._fv_index:03d}"
+
+        # Quote column names for case-sensitivity
+        quoted_join_keys = [f'"{k}"' for k in self._join_keys]
+        quoted_spine_ts = f'"{self._spine_timestamp_col}"'
+        spine_output_cols = quoted_join_keys + [quoted_spine_ts]
+        # Merged results are grouped by entity + TILE_BOUNDARY
+        boundary_group_cols = quoted_join_keys + ["TILE_BOUNDARY"]
+
+        # Select columns from spine (entity keys + original timestamp)
+        select_cols = [f"s.{col}" for col in spine_output_cols]
+
+        # Add simple feature columns (non-lifetime)
+        for spec in self._simple_features:
+            select_cols.append(f"simple.{spec.get_sql_column_name()}")
+
+        # Add list feature columns (non-lifetime)
+        for spec in self._list_features:
+            select_cols.append(f"list_agg.{spec.get_sql_column_name()}")
+
+        # Add lifetime simple feature columns
+        for spec in self._lifetime_simple_features:
+            select_cols.append(f"lifetime.{spec.get_sql_column_name()}")
+
+        # Build FROM clause with all necessary JOINs
+        from_clause = f"SPINE_BOUNDARY_FV{self._fv_index} s"
+        joins = []
+
+        # Join condition template
+        def make_join_cond(alias: str) -> str:
+            return " AND ".join(f"s.{col} = {alias}.{col}" for col in boundary_group_cols)
+
+        # Add JOINs for each CTE that has features
+        if self._simple_features:
+            joins.append(f"LEFT JOIN SIMPLE_MERGED_FV{self._fv_index} simple ON {make_join_cond('simple')}")
+
+        if self._list_features:
+            joins.append(f"LEFT JOIN LIST_MERGED_FV{self._fv_index} list_agg ON {make_join_cond('list_agg')}")
+
+        if self._lifetime_simple_features:
+            joins.append(f"LEFT JOIN LIFETIME_MERGED_FV{self._fv_index} lifetime ON {make_join_cond('lifetime')}")
+
+        if joins:
+            cte_body = f"""
+    SELECT {', '.join(select_cols)}
+    FROM {from_clause}
+    {chr(10).join('    ' + j for j in joins)}
+"""
+        else:
+            # No features (shouldn't happen)
+            cte_body = f"""
+    SELECT DISTINCT {', '.join(spine_output_cols)}
+    FROM SPINE
+"""
+
+        return cte_name, cte_body.strip()
+
+    def get_output_columns(self) -> list[str]:
+        """Get the list of output column names from this feature view."""
+        return [spec.output_column for spec in self._features]