JingsongLi commented on code in PR #9850: URL: https://github.com/apache/paimon/pull/9850#discussion_r4056456298
########## paimon-python/pypaimon/read/reader/parquet_page_index_reader.py: ########## @@ -0,0 +1,626 @@ +# Licensed to the Apache Software Foundation (ASF) under one +# or more contributor license agreements. See the NOTICE file +# distributed with this work for additional information +# regarding copyright ownership. The ASF licenses this file +# to you under the Apache License, Version 2.0 (the +# "License"); you may not use this file except in compliance +# with the License. You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, +# software distributed under the License is distributed on an +# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY +# KIND, either express or implied. See the License for the +# specific language governing permissions and limitations +# under the License. + +"""Read contiguous Parquet row windows using their OffsetIndex. + +Selected encoded pages are placed in bounded, in-memory Parquet files. PyArrow +still decodes the pages, including dictionary and compression encodings. Source +files are never rewritten. Nested fields retain their complete physical schema +and align their leaf columns at common row boundaries. Files without indexes and +disjoint ranges use the normal reader: per-page seeks can amplify requests in +filesystems that prefetch remote data (including Jindo). +""" + +import base64 +import bisect +import struct + +import pyarrow as pa +import pyarrow.parquet as pq + + +# Bound encoded page data retained by the temporary column files. +_MAX_PAGE_BYTES = 32 * 1024 * 1024 +# Bound all serialized OffsetIndexes and their retained typed PageLocations. +_MAX_INDEX_BYTES = 8 * 1024 * 1024 +_MAX_PAGE_LOCATIONS = 128 * 1024 + + +class _PageIndexBudgetExceeded(Exception): + pass + + +class _Compact: + """Thrift compact values used by Parquet metadata (no generated bindings).""" + + def __init__(self, data): + self.data = memoryview(data) + self.position = 0 + + def take(self, size): + end = self.position + size + if size < 0 or end > len(self.data): + raise ValueError("Truncated Parquet page-index metadata") + result = self.data[self.position:end] + self.position = end + return result + + def unsigned(self): + result = 0 + for shift in range(0, 70, 7): + value = self.take(1)[0] + result |= (value & 127) << shift + if value < 128: + return result + raise ValueError("Invalid Parquet compact integer") + + def value(self, kind, depth=0): + if depth > 64: + raise ValueError("Parquet metadata nesting exceeds 64 levels") + if kind in (1, 2): + return kind == 1 + if kind == 3: + return self.take(1).tobytes() + if kind in (4, 5, 6): + value = self.unsigned() + return (value >> 1) ^ -(value & 1) + if kind == 7: + return self.take(8).tobytes() + if kind == 8: + return self.take(self.unsigned()).tobytes() + if kind in (9, 10): + header = self.take(1)[0] + count, element = header >> 4, header & 15 + if count == 15: + count = self.unsigned() + if count > len(self.data) - self.position: + raise ValueError("Invalid Parquet compact collection size") + return element, [ + self.value(self.take(1)[0] if element in (1, 2) else element, depth + 1) + for _ in range(count) + ] + if kind == 12: + fields = {} + previous = 0 + while True: + header = self.take(1)[0] + if header == 0: + return fields + delta, field_kind = header >> 4, header & 15 + field = previous + delta if delta else self.value(4) + if field in fields: + raise ValueError("Duplicate Parquet compact field") + fields[field] = field_kind, self.value(field_kind, depth + 1) + previous = field + raise ValueError("Unsupported Parquet compact type: {}".format(kind)) + + +class _OffsetIndexDecoder: + """Decode typed PageLocations without materializing generic Thrift trees.""" + + def __init__(self, data, max_locations): + self.parser = _Compact(data) + self.max_locations = max_locations + self.remaining_items = max_locations * 4 + + def decode(self): + locations = None + previous = 0 + seen = set() + while True: + field = self._field(previous) + if field is None: + break + field_id, kind = field + if field_id in seen: + raise ValueError("Duplicate Parquet OffsetIndex field") + seen.add(field_id) + previous = field_id + if field_id == 1: + if kind != 9: + raise ValueError("Invalid Parquet OffsetIndex page locations") + locations = self._locations() + else: + self._skip(kind) + if locations is None or self.parser.position != len(self.parser.data): + raise ValueError("Invalid Parquet OffsetIndex") + return locations + + def _field(self, previous): + header = self.parser.take(1)[0] + if header == 0: + return None + delta, kind = header >> 4, header & 15 + field = previous + delta if delta else self.parser.value(4) + if field <= 0: + raise ValueError("Invalid Parquet compact field") + return field, kind + + def _collection(self): + header = self.parser.take(1)[0] + count, element = header >> 4, header & 15 + if count == 15: + count = self.parser.unsigned() + if count > len(self.parser.data) - self.parser.position: + raise ValueError("Invalid Parquet compact collection size") + return count, element + + def _locations(self): + count, element = self._collection() + if element != 12: + raise ValueError("Invalid Parquet OffsetIndex page locations") + if count > self.max_locations: + raise _PageIndexBudgetExceeded( + "Parquet OffsetIndex exceeds page-location budget") + self._consume(count) + return [self._location() for _ in range(count)] + + def _location(self): + values = [None, None, None] + expected = (6, 5, 6) + previous = 0 + seen = set() + while True: + field = self._field(previous) + if field is None: + break + field_id, kind = field + if field_id in seen: + raise ValueError("Duplicate Parquet PageLocation field") + seen.add(field_id) + previous = field_id + if 1 <= field_id <= 3: + if kind != expected[field_id - 1]: + raise ValueError("Invalid Parquet PageLocation field type") + values[field_id - 1] = self.parser.value(kind) + else: + self._skip(kind) + if any(value is None for value in values): + raise ValueError("Missing Parquet PageLocation field") + offset, size, first_row = values + if offset < 0 or size <= 0 or first_row < 0: + raise ValueError("Invalid Parquet PageLocation") + return offset, size, first_row + + def _consume(self, count): + if count > self.remaining_items: + raise _PageIndexBudgetExceeded( + "Parquet compact metadata exceeds object budget") + self.remaining_items -= count + + def _skip_collection_value(self, kind, depth): + if kind in (1, 2): + actual = self.parser.take(1)[0] + if actual not in (1, 2): + raise ValueError("Invalid Parquet compact boolean") + else: + self._skip(kind, depth) + + def _skip(self, kind, depth=0): + if depth > 64: + raise ValueError("Parquet metadata nesting exceeds 64 levels") + if kind in (1, 2): + return + if kind == 3: + self.parser.take(1) + return + if kind in (4, 5, 6): + self.parser.unsigned() + return + if kind == 7: + self.parser.take(8) + return + if kind == 8: + self.parser.take(self.parser.unsigned()) + return + if kind in (9, 10): + count, element = self._collection() + self._consume(count) + for _ in range(count): + self._skip_collection_value(element, depth + 1) + return + if kind == 11: + count = self.parser.unsigned() + self._consume(count * 2) + if count: + kinds = self.parser.take(1)[0] + key_kind, value_kind = kinds >> 4, kinds & 15 + for _ in range(count): + self._skip_collection_value(key_kind, depth + 1) + self._skip_collection_value(value_kind, depth + 1) + return + if kind == 12: + previous = 0 + seen = set() + while True: + field = self._field(previous) + if field is None: + return + field_id, field_kind = field + if field_id in seen: + raise ValueError("Duplicate Parquet compact field") + seen.add(field_id) + previous = field_id + self._skip(field_kind, depth + 1) + raise ValueError("Unsupported Parquet compact type: {}".format(kind)) + + +def _decode_offset_index(data, max_locations): + return _OffsetIndexDecoder(data, max_locations).decode() + + +def _unsigned(value): + result = bytearray() + while value >= 128: + result.append((value & 127) | 128) + value >>= 7 + result.append(value) + return bytes(result) + + +def _encode(kind, value): + if kind in (1, 2): + return bytes([1 if value else 2]) + if kind in (3, 7): + return value + if kind in (4, 5, 6): + return _unsigned(value * 2 if value >= 0 else -value * 2 - 1) + if kind == 8: + return _unsigned(len(value)) + value + if kind in (9, 10): + element, items = value + size = len(items) + header = bytes([(min(size, 15) << 4) | element]) + if size >= 15: + header += _unsigned(size) + return header + b"".join(_encode(element, item) for item in items) + if kind == 12: + result = bytearray() + previous = 0 + for field, (field_kind, item) in sorted(value.items()): + delta = field - previous + if field_kind in (1, 2): + field_kind = 1 if item else 2 + if 0 < delta < 16: + result.append((delta << 4) | field_kind) + else: + result.append(field_kind) + result.extend(_encode(4, field)) + if field_kind not in (1, 2): + result.extend(_encode(field_kind, item)) + previous = field + result.append(0) + return bytes(result) + raise ValueError("Unsupported Parquet compact type: {}".format(kind)) + + +def _get(fields, field, default=None): + return fields[field][1] if field in fields else default + + +def _read_exact(source, offset, length): + if offset < 4 or length <= 0: + raise ValueError("Invalid Parquet page-index byte range") + data = source.read_at(length, offset) + if len(data) != length: + raise OSError("Truncated Parquet page-index byte range") + return data + + +def _read_index_ranges(source, ranges): + groups = [] + for key, offset, length in sorted(ranges, key=lambda item: item[1]): + if offset < 4 or length <= 0: + raise ValueError("Invalid Parquet page-index byte range") + end = offset + length + if groups and offset < groups[-1][1]: + raise ValueError("Overlapping Parquet page-index byte ranges") + if groups and offset == groups[-1][1]: + groups[-1][1] = end + groups[-1][2].append((key, offset, length)) + else: + groups.append([offset, end, [(key, offset, length)]]) + result = {} + for start, end, members in groups: + data = memoryview(_read_exact(source, start, end - start)) + for key, offset, length in members: + result[key] = data[offset - start:offset - start + length] + return result + + +class ParquetPageIndexReader: + def __init__(self, source, metadata, schema, footer, columns, fields, batch_size): + self.source = source + self.metadata = metadata + self.schema = schema + self.footer = footer + self.columns = columns + self.fields = fields + self.batch_size = batch_size + + @classmethod + def create(cls, source, parquet_file, columns, row_groups, batch_size): + metadata = parquet_file.metadata + schema = parquet_file.schema_arrow + if not columns or len(set(schema.names)) != len(schema): + return None + indices = [schema.get_field_index(name) for name in columns] + if any(index < 0 for index in indices) or len(set(indices)) != len(indices): + return None + if not any(getattr(metadata.row_group(group).column(index), + "has_offset_index", False) + for group in row_groups for index in range(metadata.num_columns)): + return None + output = pa.BufferOutputStream() + metadata.write_metadata_file(output) + serialized = output.getvalue().to_pybytes() + length = struct.unpack("<I", serialized[-8:-4])[0] + footer = _Compact(serialized[-8 - length:-8]).value(12) Review Comment: [P2] Avoid materializing every unselected row group footer tree create serializes and generic-decodes the complete FileMetaData, then retains the whole Python object tree even when row_groups selects only one group. On a valid one-column file with 5,000 row groups, create(..., row_groups=[0]) turned a 574,774-byte footer into 25,377,527 retained bytes, peaked at 26,528,093 bytes, and took 0.742 seconds locally. This is paid on the default-enabled row-range path and scales with irrelevant groups. Please stream or retain only schema plus selected row groups, cache a bounded decoded representation, or reject oversized or over-fragmented metadata before generic materialization and fall back to the ordinary reader. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
