xlr8 0.1.2__py3-none-any.whl

xlr8/storage/reader.py

@@ -0,0 +1,1276 @@
+"""
+Parquet file reader for cache-aware loading.
+
+This module reads Parquet files written by the Rust backend and converts them
+back into DataFrames with proper value decoding and type reconstruction.
+
+DATA FLOW
+=========
+
+STEP 1: DISCOVER RUST-WRITTEN FILES
+------------------------------------
+The Rust backend (rust_backend.fetch_chunks_bson) writes Parquet files with
+timestamp-based naming derived from actual document data:
+
+    cache_dir/.cache/abc123def/
+        ts_1704067200_1704070800_part_0000.parquet
+        ts_1704070801_1704074400_part_0000.parquet
+        ts_1704074401_1704078000_part_0000.parquet
+        ts_1704078001_1704081600_part_0000.parquet
+        ...
+
+Filename format: ts_{min_sec}_{max_sec}_part_{counter:04}.parquet
+- min_sec: Unix timestamp (seconds) of earliest document in file
+- max_sec: Unix timestamp (seconds) of latest document in file
+- counter: Per-worker sequential counter (0000, 0001, 0002, ...)
+  Only increments if same worker writes multiple files with identical timestamps
+
+How timestamps ensure uniqueness:
+- Each chunk/bracket targets different time ranges
+- Multiple workers process non-overlapping time ranges
+- Natural file separation by actual data timestamps
+- Counter only needed if worker flushes multiple batches with identical ranges
+
+Fallback format (no timestamps): part_{counter:04}.parquet
+Used when time_field is None or documents lack timestamps
+
+
+STEP 2: READ & CONCATENATE
+---------------------------
+Pandas: Read all files sequentially, concatenate into single DataFrame
+Polars: Read all files in parallel (native multi-file support)
+
+Both engines use PyArrow under the hood for efficient Parquet parsing.
+
+
+STEP 3: DECODE TYPES.ANY STRUCT VALUES
+---------------------------------------
+Types.Any fields are encoded as Arrow structs by Rust backend:
+
+    Parquet stores:
+    {
+        "value": {
+            "float_value": 42.5,
+            "int_value": null,
+            "string_value": null,
+            "bool_value": null,
+            ...
+        }
+    }
+
+    After decoding (coalesce first non-null field):
+    {"value": 42.5}
+
+This decoding happens in Rust via decode_any_struct_arrow() for maximum
+performance.
+
+
+STEP 4: FLATTEN NESTED STRUCTS
+-------------------------------
+Convert nested struct columns to dotted field names:
+
+    Before: {"metadata": {"device_id": "123...", "sensor_id": "456..."}}
+    After:  {"metadata.device_id": "123...", "metadata.sensor_id": "456..."}
+
+
+STEP 5: RECONSTRUCT OBJECTIDS
+------------------------------
+Convert string-encoded ObjectIds back to bson.ObjectId instances:
+
+    "507f1f77bcf86cd799439011" -> ObjectId("507f1f77bcf86cd799439011")
+
+
+OUTPUT: DataFrame ( or Polars to stream pyarrow.Table )
+-----------------
+    timestamp          metadata.device_id    value
+ 0  2024-01-15 12:00   64a1b2c3...           42.5
+ 1  2024-01-15 12:01   64a1b2c3...           43.1
+ 2  2024-01-15 12:02   64a1b2c3...           "active"
+
+"""
+
+import logging
+from datetime import datetime
+from pathlib import Path
+from typing import Any, Dict, Generator, Iterator, List, Literal, Optional, Tuple, Union
+
+import duckdb
+import pandas as pd
+import polars as pl
+import pyarrow as pa
+import pyarrow.parquet as pq
+from bson import ObjectId
+
+from xlr8.constants import DEFAULT_BATCH_SIZE
+
+logger = logging.getLogger(__name__)
+
+
+class ParquetReader:
+    """
+    Reads Parquet files from cache directory.
+
+    Provides streaming and batch reading of documents from Parquet files.
+    Supports reading all files in a cache directory or specific partitions.
+
+    Example:
+        >>> reader = ParquetReader(cache_dir=".cache/abc123def")
+        >>>
+        >>> # Stream all documents
+        >>> for doc in reader.iter_documents():
+        ...     print(doc)
+        >>>
+        >>> # Or load to DataFrame
+        >>> df = reader.to_dataframe()
+    """
+
+    def __init__(self, cache_dir: Union[str, Path]):
+        """
+        Initialize reader for cache directory.
+
+        Args:
+            cache_dir: Directory containing parquet files
+        """
+        self.cache_dir = Path(cache_dir)
+
+        if not self.cache_dir.exists():
+            raise FileNotFoundError(f"Cache directory not found: {cache_dir}")
+
+        # Find all parquet files (may be empty if query returned no results)
+        self.parquet_files = sorted(self.cache_dir.glob("*.parquet"))
+
+    def iter_documents(
+        self,
+        batch_size: int = DEFAULT_BATCH_SIZE,
+    ) -> Iterator[Dict[str, Any]]:
+        """
+        Stream documents from all parquet files.
+
+        Reads in batches to avoid loading entire dataset into memory.
+
+        Args:
+            batch_size: Number of rows to read per batch
+
+        Yields:
+            Document dictionaries
+
+        Example:
+            >>> for doc in reader.iter_documents(batch_size=5000):
+            ...     process(doc)
+        """
+        for parquet_file in self.parquet_files:
+            # Read in batches
+            parquet_file_obj = pq.ParquetFile(parquet_file)
+
+            for batch in parquet_file_obj.iter_batches(batch_size=batch_size):
+                # Convert Arrow batch to pandas then to dicts
+                df_batch = batch.to_pandas()
+
+                for _, row in df_batch.iterrows():
+                    yield row.to_dict()
+
+    def _is_any_type(self, field_type: Any) -> bool:
+        """Check if field_type is an Any type (supports both class and instance)."""
+        from xlr8.schema.types import Any as AnyType
+
+        # Support both Types.Any (class) and Types.Any() (instance)
+        if isinstance(field_type, AnyType):
+            return True
+        if isinstance(field_type, type) and issubclass(field_type, AnyType):
+            return True
+        return False
+
+    def _decode_struct_values(self, df: pd.DataFrame, schema: Any) -> pd.DataFrame:
+        """
+        Decode struct-encoded Any-typed columns back to actual values.
+
+        For columns marked as Any type in schema, extracts the actual value
+        from the struct bitmap representation (float_value, int_value, etc.).
+
+        Uses Rust Arrow-native decoding for maximum performance (~40x faster).
+
+        Note: This is a fallback path. The fast path decodes directly from Arrow
+        before to_pandas() conversion, avoiding dict overhead entirely.
+        """
+        if not hasattr(schema, "fields"):
+            return df
+
+        # Import Rust Arrow-native decoder (required)
+        from xlr8.rust_backend import decode_any_struct_arrow
+
+        # Find Any-typed fields in schema
+        for field_name, field_type in schema.fields.items():
+            if self._is_any_type(field_type) and field_name in df.columns:
+                # Column contains struct-encoded values (dicts)
+                col = df[field_name]
+
+                if len(col) == 0:
+                    continue
+
+                # Check if it's a struct (dict) column - skip if already decoded
+                first_val = col.iloc[0]
+                if not isinstance(first_val, dict):
+                    # Already decoded in fast path - skip
+                    continue
+
+                # Build struct type dynamically based on the dict keys
+                sample_dict = first_val
+                struct_fields = []
+                field_type_map = {
+                    "float_value": pa.float64(),
+                    "int32_value": pa.int32(),
+                    "int64_value": pa.int64(),
+                    "string_value": pa.string(),
+                    "objectid_value": pa.string(),
+                    "decimal128_value": pa.string(),
+                    "regex_value": pa.string(),
+                    "binary_value": pa.string(),
+                    "document_value": pa.string(),
+                    "array_value": pa.string(),
+                    "bool_value": pa.bool_(),
+                    "datetime_value": pa.timestamp("ms"),  # Use ms for new schema
+                    "null_value": pa.bool_(),
+                }
+
+                for key in sample_dict.keys():
+                    if key in field_type_map:
+                        struct_fields.append((key, field_type_map[key]))
+
+                any_struct_type = pa.struct(struct_fields)
+
+                # Convert to PyArrow array - this is a single pass over the data
+                arrow_array = pa.array(col.tolist(), type=any_struct_type)
+
+                # Decode in Rust - direct memory access to Arrow memory
+                decoded_values = decode_any_struct_arrow(arrow_array)
+                df[field_name] = decoded_values
+
+        return df
+
+    def _flatten_struct_columns(self, df: pd.DataFrame) -> pd.DataFrame:
+        """
+        Flatten nested struct columns into separate columns.
+
+        Example:
+            metadata: {'sensor_id': '...', 'device_id': '...'}
+            -> metadata.sensor_id: '...', metadata.device_id: '...'
+
+        """
+        if df.empty:
+            return df
+
+        struct_cols = []
+        for col in df.columns:
+            # Check if column contains dicts (structs)
+            if len(df) > 0 and isinstance(df[col].iloc[0], dict):
+                struct_cols.append(col)
+
+        for col in struct_cols:
+            # FAST PATH: Extract struct fields directly using list comprehension
+            # This is ~5x faster than pd.json_normalize() for large datasets
+            col_values = df[col].tolist()
+
+            # Detect subcolumns from first non-null row
+            first_val = col_values[0] if col_values else {}
+            subcolumns = list(first_val.keys()) if isinstance(first_val, dict) else []
+
+            # Build new columns efficiently
+            new_cols = {}
+            for subcol in subcolumns:
+                new_col_name = f"{col}.{subcol}"
+                new_cols[new_col_name] = [
+                    row.get(subcol) if isinstance(row, dict) else None
+                    for row in col_values
+                ]
+
+            # Drop original struct column
+            df = df.drop(columns=[col])
+
+            # Add flattened columns
+            for new_col_name, values in new_cols.items():
+                df[new_col_name] = values
+
+        return df
+
+    def _reconstruct_objectids(self, df: pd.DataFrame, schema: Any) -> pd.DataFrame:
+        """
+        Reconstruct ObjectId columns from string representation.
+
+        Converts string ObjectIds back to bson.ObjectId instances.
+        """
+        from xlr8.schema.types import ObjectId as ObjectIdType
+
+        # Find all ObjectId fields in schema (including nested ones)
+        objectid_fields = []
+
+        if hasattr(schema, "fields"):
+            for field_name, field_type in schema.fields.items():
+                if isinstance(field_type, ObjectIdType):
+                    objectid_fields.append(field_name)
+                elif hasattr(field_type, "fields"):
+                    # Nested struct with ObjectId fields
+                    for nested_name, nested_type in field_type.fields.items():
+                        if isinstance(nested_type, ObjectIdType):
+                            objectid_fields.append(f"{field_name}.{nested_name}")
+
+        # Convert string columns back to ObjectId
+        for field in objectid_fields:
+            if field in df.columns:
+                df[field] = df[field].apply(
+                    lambda x: ObjectId(x) if x and pd.notna(x) else x
+                )
+
+        return df
+
+    def _decode_struct_values_polars(
+        self,
+        df: "pl.DataFrame",
+        schema: Any,
+        any_type_strategy: Literal["float", "string", "keep_struct"] = "float",
+    ) -> "pl.DataFrame":
+        """
+        Decode struct-encoded Any-typed columns back to actual values (Polars).
+
+        Args:
+            df: Polars DataFrame
+            schema: Schema with field type info
+            any_type_strategy: How to decode:
+                - "float": Coalesce to Float64, prioritize numeric (default)
+                - "string": Convert everything to string (lossless)
+                - "keep_struct": Keep raw struct, don't decode
+        """
+        if not hasattr(schema, "fields"):
+            return df
+
+        # Find Any-typed fields in schema
+        for field_name, field_type in schema.fields.items():
+            if self._is_any_type(field_type) and field_name in df.columns:
+                # Check if column is a struct
+                col_dtype = df.schema[field_name]
+                if str(col_dtype).startswith("Struct"):
+                    # Strategy: keep_struct - don't decode at all
+                    if any_type_strategy == "keep_struct":
+                        continue
+
+                    try:
+                        # Get field names from the struct
+                        struct_fields = (
+                            col_dtype.fields if hasattr(col_dtype, "fields") else []
+                        )  # type: ignore[attr-defined]
+                        field_names = (
+                            [f.name for f in struct_fields] if struct_fields else []
+                        )
+
+                        if any_type_strategy == "string":
+                            # Convert ALL value types to string
+                            coalesce_exprs = []
+
+                            # String first (already string)
+                            if "string_value" in field_names:
+                                coalesce_exprs.append(
+                                    pl.col(field_name).struct.field("string_value")
+                                )
+
+                            # Float to string
+                            if "float_value" in field_names:
+                                coalesce_exprs.append(
+                                    pl.col(field_name)
+                                    .struct.field("float_value")
+                                    .cast(pl.Utf8)
+                                )
+
+                            # Int to string
+                            for int_name in ["int64_value", "int32_value"]:
+                                if int_name in field_names:
+                                    coalesce_exprs.append(
+                                        pl.col(field_name)
+                                        .struct.field(int_name)
+                                        .cast(pl.Utf8)
+                                    )
+
+                            # Bool to string
+                            if "bool_value" in field_names:
+                                coalesce_exprs.append(
+                                    pl.col(field_name)
+                                    .struct.field("bool_value")
+                                    .cast(pl.Utf8)
+                                )
+
+                            # ObjectId, decimal, etc. (already strings)
+                            for str_field in [
+                                "objectid_value",
+                                "decimal128_value",
+                                "regex_value",
+                                "binary_value",
+                                "document_value",
+                                "array_value",
+                            ]:
+                                if str_field in field_names:
+                                    coalesce_exprs.append(
+                                        pl.col(field_name).struct.field(str_field)
+                                    )
+
+                            if coalesce_exprs:
+                                df = df.with_columns(
+                                    pl.coalesce(coalesce_exprs).alias(field_name)
+                                )
+
+                        else:  # "float" strategy (default)
+                            # Coalesce to Float64, prioritize numeric
+                            coalesce_exprs = []
+
+                            # Try float first (highest precision)
+                            if "float_value" in field_names:
+                                coalesce_exprs.append(
+                                    pl.col(field_name).struct.field("float_value")
+                                )
+
+                            # Try various int types, cast to float
+                            for int_name in ["int64_value", "int32_value"]:
+                                if int_name in field_names:
+                                    coalesce_exprs.append(
+                                        pl.col(field_name)
+                                        .struct.field(int_name)
+                                        .cast(pl.Float64)
+                                    )
+
+                            # Try bool (as 0.0/1.0)
+                            if "bool_value" in field_names:
+                                coalesce_exprs.append(
+                                    pl.col(field_name)
+                                    .struct.field("bool_value")
+                                    .cast(pl.Float64)
+                                )
+
+                            if coalesce_exprs:
+                                if len(coalesce_exprs) == 1:
+                                    df = df.with_columns(
+                                        coalesce_exprs[0].alias(field_name)
+                                    )
+                                else:
+                                    df = df.with_columns(
+                                        pl.coalesce(coalesce_exprs).alias(field_name)
+                                    )
+                            else:
+                                logger.warning(
+                                    "Could not decode struct column '%s': "
+                                    "no numeric fields in %s",
+                                    field_name,
+                                    field_names,
+                                )
+                    except (AttributeError, KeyError, ValueError) as e:
+                        logger.warning("Error decoding struct '%s': %s", field_name, e)
+
+        return df
+
+    def _process_dataframe(
+        self,
+        df: Union[pd.DataFrame, "pl.DataFrame"],
+        engine: Literal["pandas", "polars"],
+        schema: Optional[Any] = None,
+        coerce: Literal["raise", "error"] = "raise",
+        any_type_strategy: Literal["float", "string", "keep_struct"] = "float",
+    ) -> Union[pd.DataFrame, "pl.DataFrame"]:
+        """
+        Process DataFrame: decode struct values, flatten structs and
+        reconstruct ObjectIds.
+
+        Args:
+            df: DataFrame to process
+            engine: "pandas" or "polars"
+            schema: Schema for ObjectId reconstruction
+            coerce: Error handling mode ("raise" or "error")
+            any_type_strategy: How to decode Any() structs in Polars
+                (float/string/keep_struct)
+
+        Returns:
+            Processed DataFrame
+        """
+        if engine == "pandas":
+            # First, decode Any-typed struct columns back to actual values
+            if schema is not None:
+                try:
+                    df = self._decode_struct_values(df, schema)  # type: ignore[arg-type]
+                except (AttributeError, KeyError, ValueError, TypeError) as e:
+                    if coerce == "error":
+                        logger.error("Error decoding struct values: %s", e)
+                    else:
+                        raise
+
+            # Flatten struct columns (e.g., metadata -> metadata.sensor_id)
+            df = self._flatten_struct_columns(df)  # type: ignore[arg-type]
+
+            # Reconstruct ObjectIds from strings
+            if schema is not None:
+                try:
+                    df = self._reconstruct_objectids(df, schema)
+                except (AttributeError, KeyError, ValueError, TypeError) as e:
+                    if coerce == "error":
+                        logger.error("Error reconstructing ObjectIds: %s", e)
+                    else:
+                        raise
+
+            return df
+        elif engine == "polars":
+            # Polars: decode Any-typed struct columns and keep dotted column names
+            if schema is not None:
+                try:
+                    df = self._decode_struct_values_polars(
+                        df, schema, any_type_strategy
+                    )  # type: ignore[arg-type]
+                except (AttributeError, KeyError, ValueError, TypeError) as e:
+                    if coerce == "error":
+                        logger.error("Error decoding struct values (polars): %s", e)
+                    else:
+                        raise
+            return df
+
+    def to_dataframe(
+        self,
+        engine: str = "pandas",
+        schema: Optional[Any] = None,
+        time_field: Optional[str] = None,
+        start_date: Optional[datetime] = None,
+        end_date: Optional[datetime] = None,
+        coerce: Literal["raise", "error"] = "raise",
+        any_type_strategy: Literal["float", "string", "keep_struct"] = "float",
+    ) -> Union[pd.DataFrame, "pl.DataFrame"]:
+        """
+        Load all parquet files into a DataFrame.
+
+        Args:
+            engine: "pandas" or "polars"
+            schema: Schema for ObjectId reconstruction and struct flattening (required)
+            time_field: Name of time field for date filtering (from schema.time_field)
+            start_date: Filter data from this date (inclusive, tz-aware datetime)
+            end_date: Filter data until this date (exclusive, tz-aware datetime)
+            coerce: Error handling mode:
+                    - "raise": Raise exceptions on schema validation errors (default)
+                    - "error": Log errors and store None for invalid values
+            any_type_strategy: How to decode Types.Any() struct columns in Polars:
+                    - "float": Coalesce to Float64, prioritize numeric (default)
+                    - "string": Convert everything to string (lossless)
+                    - "keep_struct": Keep raw struct, don't decode
+
+        Returns:
+            DataFrame with all documents (structs flattened, ObjectIds reconstructed)
+
+        Example:
+            >>> df = reader.to_dataframe(
+            ...     schema=schema,
+            ...     time_field="timestamp",
+            ...     start_date=datetime(2024, 6, 1, tzinfo=timezone.utc),
+            ...     end_date=datetime(2024, 6, 15, tzinfo=timezone.utc),
+            ... )
+        """
+        # Build PyArrow filter for date range (predicate pushdown)
+        # Convert datetime to PyArrow timestamp[ms] to match Parquet column type (no tz)
+        filters = None
+        if time_field and (start_date or end_date):
+            filter_conditions = []
+            if start_date:
+                # Convert to ms timestamp (no tz) to match Parquet storage
+                start_ts = pa.scalar(start_date, type=pa.timestamp("ms"))
+                filter_conditions.append((time_field, ">=", start_ts))
+            if end_date:
+                end_ts = pa.scalar(end_date, type=pa.timestamp("ms"))
+                filter_conditions.append((time_field, "<", end_ts))
+            if filter_conditions:
+                filters = filter_conditions
+
+        if engine == "polars":
+            # Return empty DataFrame if no parquet files (query returned no results)
+            if not self.parquet_files:
+                return pl.DataFrame()
+
+            # Use scan_parquet for lazy evaluation with predicate pushdown
+            # This only reads the row groups that match the filter conditions
+            lf = pl.scan_parquet(self.parquet_files)
+
+            # Apply date filter with predicate pushdown (reads only matching data)
+            # Convert datetime to naive (no timezone) to match Parquet column dtype
+            if time_field and (start_date or end_date):
+                if start_date:
+                    start_naive = (
+                        start_date.replace(tzinfo=None)
+                        if start_date.tzinfo
+                        else start_date
+                    )
+                    lf = lf.filter(pl.col(time_field) >= start_naive)
+                if end_date:
+                    end_naive = (
+                        end_date.replace(tzinfo=None) if end_date.tzinfo else end_date
+                    )
+                    lf = lf.filter(pl.col(time_field) < end_naive)
+
+            # Collect executes the query with predicate pushdown
+            df = lf.collect()
+
+            return self._process_dataframe(
+                df, engine, schema, coerce, any_type_strategy
+            )
+
+        elif engine == "pandas":
+            # Return empty DataFrame if no parquet files (query returned no results)
+            if not self.parquet_files:
+                return pd.DataFrame()
+
+            # Read all files with optional filter (predicate pushdown)
+            # Use PyArrow to read, then convert to pandas - this allows
+            # struct columns to stay in Arrow format for fast Rust decoding
+            tables = []
+            for parquet_file in self.parquet_files:
+                try:
+                    # Use PyArrow filters for efficient predicate pushdown
+                    table = pq.read_table(parquet_file, filters=filters)
+                    tables.append(table)
+                except Exception as e:
+                    if coerce == "error":
+                        logger.error(f"Error reading {parquet_file}: {e}")
+                        continue
+                    raise
+
+            if not tables:
+                return pd.DataFrame()
+
+            # Concatenate Arrow tables
+            combined_table = pa.concat_tables(tables)
+
+            # FAST PATH: Decode Any-typed struct columns directly in Arrow
+            # This gives us 44x speedup because Rust reads Arrow memory directly
+            # without Python iteration over dicts
+            any_columns_decoded = {}
+            columns_to_drop = []
+            if schema and hasattr(schema, "fields"):
+                from xlr8.rust_backend import decode_any_struct_arrow
+
+                for field_name, field_type in schema.fields.items():
+                    if (
+                        self._is_any_type(field_type)
+                        and field_name in combined_table.column_names
+                    ):
+                        col = combined_table.column(field_name)
+                        if pa.types.is_struct(col.type):
+                            # Decode in Rust - returns Python list of mixed types
+                            combined = col.combine_chunks()
+                            decoded_values = decode_any_struct_arrow(combined)
+                            any_columns_decoded[field_name] = decoded_values
+                            # Mark for removal to avoid slow dict conversion
+                            # in to_pandas()
+                            columns_to_drop.append(field_name)
+
+            # Drop decoded struct columns before pandas conversion
+            # to avoid dict overhead
+            if columns_to_drop:
+                combined_table = combined_table.drop(columns_to_drop)
+
+            # Convert to pandas (non-Any columns go through normal path)
+            df = combined_table.to_pandas()
+
+            # Add back Any columns with decoded values
+            # (bypassing struct->dict->decode path)
+            for field_name, decoded_values in any_columns_decoded.items():
+                df[field_name] = decoded_values
+
+            return self._process_dataframe(df, engine, schema, coerce)
+
+        else:
+            raise ValueError(f"Unknown engine: {engine}. Use 'pandas' or 'polars'")
+
+    def iter_dataframe_batches(
+        self,
+        batch_size: int = 10000,
+        schema: Optional[Any] = None,
+        time_field: Optional[str] = None,
+        start_date: Optional[datetime] = None,
+        end_date: Optional[datetime] = None,
+        coerce: Literal["raise", "error"] = "raise",
+    ) -> Generator[pd.DataFrame, None, None]:
+        """
+        Yield DataFrames in batches without loading all data into memory.
+
+        This is memory-efficient: only batch_size rows are in memory at a time.
+        Uses PyArrow's batch iteration for efficient streaming.
+
+        Use this when NO sorting is needed. For sorted batches, use
+        iter_globally_sorted_batches().
+
+        Args:
+            batch_size: Number of rows per batch (default: 10,000)
+            schema: Schema for struct decoding and ObjectId reconstruction
+            time_field: Name of time field for date filtering
+            start_date: Filter data from this date (inclusive, tz-aware)
+            end_date: Filter data until this date (exclusive, tz-aware)
+            coerce: Error handling mode ("raise" or "error")
+
+        Yields:
+            pd.DataFrame: Batches of processed rows
+
+        Example:
+            >>> for batch_df in reader.iter_dataframe_batches(batch_size=5000):
+            ...     process(batch_df)
+        """
+        import pyarrow.parquet as pq
+
+        batch_count = 0
+        total_rows = 0
+
+        for parquet_file in self.parquet_files:
+            try:
+                # Open parquet file for batch iteration
+                parquet_file_obj = pq.ParquetFile(parquet_file)
+
+                for batch in parquet_file_obj.iter_batches(batch_size=batch_size):
+                    # Convert Arrow batch to pandas
+                    batch_df = batch.to_pandas()
+
+                    # Apply date filter if specified
+                    # (in case predicate pushdown not supported)
+                    if time_field and (start_date or end_date):
+                        if time_field in batch_df.columns:
+                            if start_date:
+                                batch_df = batch_df[batch_df[time_field] >= start_date]
+                            if end_date:
+                                batch_df = batch_df[batch_df[time_field] < end_date]
+
+                    if len(batch_df) == 0:
+                        continue
+
+                    # Process the batch (decode structs, flatten, reconstruct ObjectIds)
+                    processed_df = self._process_dataframe(
+                        batch_df, "pandas", schema, coerce
+                    )
+
+                    batch_count += 1
+                    total_rows += len(processed_df)
+
+                    yield processed_df
+
+            except Exception as e:
+                if coerce == "error":
+                    logger.error(f"Error reading batch from {parquet_file}: {e}")
+                    continue
+                raise
+
+        logger.debug(f"Yielded {batch_count} batches, {total_rows} total rows")
+
+    def get_globally_sorted_dataframe(
+        self,
+        sort_spec: List[Tuple[str, int]],
+        schema: Optional[Any] = None,
+        time_field: Optional[str] = None,
+        start_date: Optional[datetime] = None,
+        end_date: Optional[datetime] = None,
+        coerce: Literal["raise", "error"] = "raise",
+        memory_limit_mb: Optional[int] = None,
+        threads: Optional[int] = None,
+    ) -> pd.DataFrame:
+        """
+        Return entire globally sorted DataFrame using DuckDB K-way merge.
+
+        More efficient than iter_globally_sorted_batches() when you want
+        the full result, as it avoids batch iteration overhead and just
+        fetches all rows at once.
+        for to_dataframe_batches() where streaming is required.
+
+        Args:
+            sort_spec: Sort specification as [(field, direction), ...]
+            schema: Schema for ObjectId reconstruction and advanced sorting
+            time_field: Field for date filtering
+            start_date: Filter data from this date (inclusive, tz-aware)
+            end_date: Filter data until this date (exclusive, tz-aware)
+            coerce: Error handling mode
+            memory_limit_mb: DuckDB memory limit
+            threads: DuckDB thread count
+
+        Returns:
+            pd.DataFrame: Complete sorted DataFrame
+        """
+        if not self.parquet_files:
+            return pd.DataFrame()
+
+        # Expand parent fields to children in schema definition order
+        sort_spec = self._expand_parent_sort_fields(sort_spec, schema)
+
+        # Get list of parquet files
+        file_paths = [str(f) for f in self.parquet_files]
+
+        logger.debug(
+            f"DuckDB K-way merge (full): {len(file_paths)} files, sort_spec={sort_spec}"
+        )
+
+        try:
+            # Create DuckDB connection
+            conn = duckdb.connect(":memory:")
+
+            # Configure DuckDB to use allocated resources
+            if memory_limit_mb:
+                conn.execute(f"SET memory_limit = '{memory_limit_mb}MB'")
+                logger.info(f"DuckDB memory_limit set to {memory_limit_mb} MB")
+
+            if threads:
+                conn.execute(f"SET threads = {threads}")
+                logger.info(f"DuckDB threads set to {threads}")
+
+            # Build ORDER BY with MongoDB type ordering
+            # (same logic as iter_globally_sorted_batches)
+            order_clauses = []
+            for field_name, direction in sort_spec:
+                dir_sql = "ASC" if direction == 1 else "DESC"
+                if schema and schema.has_field(field_name):
+                    field_type = schema.get_field_type(field_name)
+                else:
+                    field_type = None
+                is_any = self._is_any_type(field_type) if field_type else True
+
+                if is_any:
+                    # Complete MongoDB type ordering for Any() fields
+                    type_clause = f"""CASE
+                        WHEN "{field_name}" IS NULL OR "{field_name}".null_value IS TRUE
+                            THEN 0
+                        WHEN "{field_name}".float_value IS NOT NULL
+                            OR "{field_name}".int32_value IS NOT NULL
+                            OR "{field_name}".int64_value IS NOT NULL
+                            OR "{field_name}".decimal128_value IS NOT NULL
+                            THEN 1
+                        WHEN "{field_name}".string_value IS NOT NULL THEN 2
+                        WHEN "{field_name}".document_value IS NOT NULL THEN 3
+                        WHEN "{field_name}".array_value IS NOT NULL THEN 4
+                        WHEN "{field_name}".binary_value IS NOT NULL THEN 5
+                        WHEN "{field_name}".objectid_value IS NOT NULL THEN 6
+                        WHEN "{field_name}".bool_value IS NOT NULL THEN 7
+                        WHEN "{field_name}".datetime_value IS NOT NULL THEN 8
+                        WHEN "{field_name}".regex_value IS NOT NULL THEN 9
+                        ELSE 10
+                    END {dir_sql}"""
+
+                    # Value comparisons for each type
+                    num_clause = (
+                        f'COALESCE("{field_name}".float_value, '
+                        f'CAST("{field_name}".int32_value AS DOUBLE), '
+                        f'CAST("{field_name}".int64_value AS DOUBLE)) {dir_sql}'
+                    )
+                    str_clause = f'"{field_name}".string_value {dir_sql}'
+                    doc_clause = f'"{field_name}".document_value {dir_sql}'
+                    arr_clause = f'"{field_name}".array_value {dir_sql}'
+                    bin_clause = f'"{field_name}".binary_value {dir_sql}'
+                    oid_clause = f'"{field_name}".objectid_value {dir_sql}'
+                    bool_clause = f'"{field_name}".bool_value {dir_sql}'
+                    date_clause = f'"{field_name}".datetime_value {dir_sql}'
+                    regex_clause = f'"{field_name}".regex_value {dir_sql}'
+
+                    order_clauses.extend(
+                        [
+                            type_clause,
+                            num_clause,
+                            str_clause,
+                            doc_clause,
+                            arr_clause,
+                            bin_clause,
+                            oid_clause,
+                            bool_clause,
+                            date_clause,
+                            regex_clause,
+                        ]
+                    )
+                else:
+                    # Simple field - use direct comparison
+                    order_clauses.append(f'"{field_name}" {dir_sql}')
+
+            order_by = ", ".join(order_clauses)
+            files = ", ".join([f"'{f}'" for f in file_paths])
+            query = f"SELECT * FROM read_parquet([{files}]) ORDER BY {order_by}"
+
+            print(f"[DuckDB] K-way merge (full): {len(file_paths)} files")
+
+            # Fetch entire result at once using df()
+            df = conn.execute(query).df()
+
+            # Ensure time field is UTC
+            if time_field and time_field in df.columns:
+                if pd.api.types.is_datetime64_any_dtype(df[time_field]):
+                    if df[time_field].dt.tz is not None:
+                        df[time_field] = df[time_field].dt.tz_convert("UTC")
+                    else:
+                        df[time_field] = df[time_field].dt.tz_localize("UTC")
+
+            # Apply date filtering if needed
+            if time_field and (start_date or end_date):
+                if start_date:
+                    df = df[df[time_field] >= start_date]
+                if end_date:
+                    df = df[df[time_field] < end_date]
+
+            # Process the DataFrame (decode structs, reconstruct ObjectIds)
+            df = self._process_dataframe(df, "pandas", schema, coerce)
+
+            conn.close()
+            print(f"[DuckDB] K-way merge complete: {len(df):,} rows")
+            logger.debug(f"DuckDB K-way merge complete: {len(df):,} rows")
+
+            return df
+
+        except Exception as e:
+            logger.error(f"DuckDB K-way merge failed: {e}")
+            raise
+
+    def _expand_parent_sort_fields(
+        self, sort_spec: List[Tuple[str, int]], schema: Optional[Any]
+    ) -> List[Tuple[str, int]]:
+        """
+        Expand parent field sorts to their child fields in schema definition order.
+
+        When user sorts by a parent field like "metadata" but the schema has
+        flattened fields like "metadata.device_id", expand to all children.
+
+        Args:
+            sort_spec: Original [(field, direction), ...]
+            schema: XLR8 schema with field definitions
+
+        Returns:
+            Expanded sort spec with parent fields replaced by children
+
+        Raises:
+            ValueError: If field not found and no children exist
+        """
+        if schema is None:
+            return sort_spec
+
+        expanded = []
+        # Schema.fields preserves insertion order (Python 3.7+)
+        all_fields = list(schema.fields.keys())
+
+        for field_name, direction in sort_spec:
+            if schema.has_field(field_name):
+                # Field exists directly in schema
+                expanded.append((field_name, direction))
+            else:
+                # Look for child fields with this prefix (in schema order)
+                prefix = f"{field_name}."
+                children = [f for f in all_fields if f.startswith(prefix)]
+
+                if children:
+                    logger.info(
+                        f"Sort field '{field_name}' expanded to children "
+                        f"(schema order): {children}"
+                    )
+                    for child in children:
+                        expanded.append((child, direction))
+                else:
+                    raise ValueError(
+                        f"Sort field '{field_name}' not found in schema "
+                        f"and has no child fields. "
+                        f"Available fields: {sorted(all_fields)[:10]}"
+                        + ("..." if len(all_fields) > 10 else "")
+                    )
+
+        return expanded
+
+    def iter_globally_sorted_batches(
+        self,
+        sort_field: Optional[str] = None,
+        ascending: bool = True,
+        batch_size: int = DEFAULT_BATCH_SIZE,
+        schema: Optional[Any] = None,
+        time_field: Optional[str] = None,
+        start_date: Optional[datetime] = None,
+        end_date: Optional[datetime] = None,
+        coerce: Literal["raise", "error"] = "raise",
+        sort_spec: Optional[List[Tuple[str, int]]] = None,
+        # DuckDB configuration
+        memory_limit_mb: Optional[int] = None,
+        threads: Optional[int] = None,
+    ) -> Generator[pd.DataFrame, None, None]:
+        """
+        Yield globally sorted batches using DuckDB K-way merge.
+
+        This method reads all Parquet files in the cache directory and
+        yields batches in globally sorted order. Uses Rust's K-way merge
+        with MongoDB BSON comparison for 100% compatibility.
+
+        Supports advanced sorting:
+        - Parent fields (e.g., "metadata" expands to all child fields)
+        - Types.Any() with full MongoDB BSON type ordering (Objects, Arrays, Binary)
+
+        RAM Usage:
+            O(K × batch_size) where K = number of files.
+            Already handled by flush_ram_limit_mb.
+
+        Args:
+            sort_field: Field to sort by (use sort_spec for multi-field sorting).
+            ascending: Sort direction (use sort_spec for mixed directions).
+            batch_size: Number of rows per yielded DataFrame (default: 10,000)
+            schema: Schema for ObjectId reconstruction and advanced sorting
+            time_field: Field for date filtering (usually same as sort_field)
+            start_date: Filter data from this date (inclusive, tz-aware)
+            end_date: Filter data until this date (exclusive, tz-aware)
+            coerce: Error handling mode ("raise" or "error")
+            sort_spec: Sort specification as [(field, direction), ...] where
+                      direction is 1 (ASC) or -1 (DESC). Preferred over sort_field.
+
+        Yields:
+            pd.DataFrame: Batches in globally sorted order
+
+        Example:
+            >>> reader = ParquetReader(".cache/abc123def")
+            >>> # Simple sort
+            >>> for batch in reader.iter_globally_sorted_batches(
+            ...     sort_spec=[("timestamp", 1)],
+            ...     schema=schema,
+            ...     batch_size=10_000
+            ... ):
+            ...     process(batch)
+            >>>
+            >>> # Advanced: parent field + Any type
+            >>> for batch in reader.iter_globally_sorted_batches(
+            ...     sort_spec=[("metadata", -1), ("value", 1)],
+            ...     schema=schema,
+            ... ):
+            ...     process(batch)
+        """
+
+        if not self.parquet_files:
+            return
+
+        # Handle backwards compatibility
+        if sort_spec is None and sort_field is not None:
+            direction = 1 if ascending else -1
+            sort_spec = [(sort_field, direction)]
+
+        if sort_spec is None:
+            raise ValueError("sort_spec or sort_field is required")
+
+        # Expand parent fields to children in schema definition order
+        sort_spec = self._expand_parent_sort_fields(sort_spec, schema)
+
+        # Get list of parquet files
+        file_paths = [str(f) for f in self.parquet_files]
+
+        logger.debug(
+            f"DuckDB K-way merge: {len(file_paths)} files, sort_spec={sort_spec}"
+        )
+
+        try:
+            # Create DuckDB connection
+            conn = duckdb.connect(":memory:")
+
+            # Configure DuckDB to use allocated resources
+            if memory_limit_mb:
+                conn.execute(f"SET memory_limit = '{memory_limit_mb}MB'")
+                logger.info(f"DuckDB memory_limit set to {memory_limit_mb} MB")
+
+            if threads:
+                conn.execute(f"SET threads = {threads}")
+                logger.info(f"DuckDB threads set to {threads}")
+
+            # Query DuckDB settings to verify
+            memory_result = conn.execute(
+                "SELECT current_setting('memory_limit')"
+            ).fetchone()
+            actual_memory = memory_result[0] if memory_result else "unknown"
+            threads_result = conn.execute(
+                "SELECT current_setting('threads')"
+            ).fetchone()
+            actual_threads = threads_result[0] if threads_result else "unknown"
+            logger.debug(
+                f"DuckDB configured: memory={actual_memory}, threads={actual_threads}"
+            )
+
+            # Build ORDER BY with MongoDB type ordering
+            order_clauses = []
+            for field_name, direction in sort_spec:
+                dir_sql = "ASC" if direction == 1 else "DESC"
+                # Check if field exists in schema before getting type
+                if schema and schema.has_field(field_name):
+                    field_type = schema.get_field_type(field_name)
+                else:
+                    field_type = None
+                is_any = self._is_any_type(field_type) if field_type else True
+
+                if is_any:
+                    # Complete MongoDB type ordering for Any() fields:
+                    # Reference: https://www.mongodb.com/docs/manual/reference/bson-type-comparison-order/
+                    # 1. MinKey (internal)
+                    # 2. Null
+                    # 3. Numbers (int, long, double, decimal)
+                    # 4. Symbol, String
+                    # 5. Object
+                    # 6. Array
+                    # 7. BinData
+                    # 8. ObjectId
+                    # 9. Boolean
+                    # 10. Date
+                    # 11. Timestamp
+                    # 12. Regular Expression
+                    # 13. MaxKey (internal)
+
+                    # Type priority clause
+                    type_clause = f"""CASE
+                        WHEN "{field_name}" IS NULL OR "{field_name}".null_value IS TRUE
+                            THEN 0
+                        WHEN "{field_name}".float_value IS NOT NULL
+                            OR "{field_name}".int32_value IS NOT NULL
+                            OR "{field_name}".int64_value IS NOT NULL
+                            OR "{field_name}".decimal128_value IS NOT NULL
+                            THEN 1
+                        WHEN "{field_name}".string_value IS NOT NULL THEN 2
+                        WHEN "{field_name}".document_value IS NOT NULL THEN 3
+                        WHEN "{field_name}".array_value IS NOT NULL THEN 4
+                        WHEN "{field_name}".binary_value IS NOT NULL THEN 5
+                        WHEN "{field_name}".objectid_value IS NOT NULL THEN 6
+                        WHEN "{field_name}".bool_value IS NOT NULL THEN 7
+                        WHEN "{field_name}".datetime_value IS NOT NULL THEN 8
+                        WHEN "{field_name}".regex_value IS NOT NULL THEN 9
+                        ELSE 10
+                    END {dir_sql}"""
+
+                    # Value comparisons for each type
+                    num_clause = (
+                        f'COALESCE("{field_name}".float_value, '
+                        f'CAST("{field_name}".int32_value AS DOUBLE), '
+                        f'CAST("{field_name}".int64_value AS DOUBLE)) {dir_sql}'
+                    )
+                    str_clause = f'"{field_name}".string_value {dir_sql}'
+                    # JSON strings compare lexicographically
+                    doc_clause = f'"{field_name}".document_value {dir_sql}'
+                    # JSON arrays compare lexicographically
+                    arr_clause = f'"{field_name}".array_value {dir_sql}'
+                    bin_clause = f'"{field_name}".binary_value {dir_sql}'
+                    oid_clause = f'"{field_name}".objectid_value {dir_sql}'
+                    bool_clause = f'"{field_name}".bool_value {dir_sql}'
+                    date_clause = f'"{field_name}".datetime_value {dir_sql}'
+                    regex_clause = f'"{field_name}".regex_value {dir_sql}'
+
+                    order_clauses.extend(
+                        [
+                            type_clause,
+                            num_clause,
+                            str_clause,
+                            doc_clause,
+                            arr_clause,
+                            bin_clause,
+                            oid_clause,
+                            bool_clause,
+                            date_clause,
+                            regex_clause,
+                        ]
+                    )
+                else:
+                    # Simple field - use direct comparison
+                    order_clauses.append(f'"{field_name}" {dir_sql}')
+
+            order_by = ", ".join(order_clauses)
+            files = ", ".join([f"'{f}'" for f in file_paths])
+            query = f"SELECT * FROM read_parquet([{files}]) ORDER BY {order_by}"
+
+            result = conn.execute(query)
+
+            # Use fetchmany() cursor API - this ACTUALLY streams incrementally
+            # without loading all data into memory (unlike fetch_df_chunk)
+            # NOTE: DuckDB's k-way merge uses internal buffering
+            # separate from batch_size.
+            # batch_size only controls how much we pull at once,
+            # not DuckDB's merge buffer.
+            batch_count = 0
+            total_rows = 0
+            column_names = [desc[0] for desc in result.description]
+
+            print(
+                f"[DuckDB] K-way merge started: {len(file_paths)} files, "
+                f"batch_size={batch_size:,}"
+            )
+
+            while True:
+                # Fetch batch as list of tuples
+                rows = result.fetchmany(batch_size)
+                if not rows:
+                    break
+
+                batch_count += 1
+                total_rows += len(rows)
+
+                # Convert to DataFrame
+                batch_df = pd.DataFrame(rows, columns=column_names)
+                logger.debug(
+                    f"Streamed batch {batch_count}: {len(batch_df)} rows "
+                    f"from DuckDB K-way merge"
+                )
+
+                # Ensure time field is UTC (DuckDB might return naive)
+                if time_field and time_field in batch_df.columns:
+                    if pd.api.types.is_datetime64_any_dtype(batch_df[time_field]):
+                        if batch_df[time_field].dt.tz is not None:
+                            batch_df[time_field] = batch_df[time_field].dt.tz_convert(
+                                "UTC"
+                            )
+                        else:
+                            batch_df[time_field] = batch_df[time_field].dt.tz_localize(
+                                "UTC"
+                            )
+
+                # Apply date filtering if needed
+                if time_field and (start_date or end_date):
+                    if start_date:
+                        batch_df = batch_df[batch_df[time_field] >= start_date]
+                    if end_date:
+                        batch_df = batch_df[batch_df[time_field] < end_date]
+                    if len(batch_df) == 0:
+                        continue
+
+                # Process the batch (decode structs, reconstruct ObjectIds)
+                processed_df = self._process_dataframe(
+                    batch_df, "pandas", schema, coerce
+                )
+                yield processed_df
+
+            conn.close()
+            logger.debug("DuckDB K-way merge complete")
+
+        except Exception as e:
+            if coerce == "error":
+                logger.error(f"Error in globally sorted streaming: {e}")
+                return
+            raise
+
+    def get_statistics(self) -> Dict[str, Any]:
+        """
+        Get statistics about cached data.
+
+        Returns:
+            Dict with file count, total rows, size, schema info
+        """
+        total_rows = 0
+        total_size = 0
+        schema = None
+
+        for parquet_file in self.parquet_files:
+            # File size
+            total_size += parquet_file.stat().st_size
+
+            # Read metadata
+            parquet_meta = pq.read_metadata(parquet_file)
+            total_rows += parquet_meta.num_rows
+
+            # Get schema from first file
+            if schema is None:
+                schema = parquet_meta.schema.to_arrow_schema()
+
+        return {
+            "file_count": len(self.parquet_files),
+            "total_rows": total_rows,
+            "total_size_mb": round(total_size / (1024 * 1024), 2),
+            "schema_fields": [field.name for field in schema] if schema else [],
+            "cache_dir": str(self.cache_dir),
+        }
+
+    def __repr__(self) -> str:
+        stats = self.get_statistics()
+        return (
+            f"ParquetReader(files={stats['file_count']}, "
+            f"rows={stats['total_rows']:,}, "
+            f"size={stats['total_size_mb']:.1f}MB)"
+        )
+
+    def __len__(self) -> int:
+        """Return total number of rows across all files."""
+        return self.get_statistics()["total_rows"]