JingsongLi commented on code in PR #8136: URL: https://github.com/apache/paimon/pull/8136#discussion_r3458071671
########## paimon-python/pypaimon/common/predicate_json_parser.py: ########## @@ -0,0 +1,365 @@ +################################################################################ +# 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. +################################################################################ + +import json +import re +from typing import Callable + +import pyarrow as pa +import pyarrow.compute as pc + + +def parse_predicate_to_batch_filter(json_str: str) -> Callable[[pa.RecordBatch], pa.Array]: + data = json.loads(json_str) + return _build_filter(data) + + +def _build_filter(data: dict) -> Callable[[pa.RecordBatch], pa.Array]: + kind = data["kind"] + if kind == "LEAF": + return _build_leaf_filter(data) + elif kind == "COMPOUND": + return _build_compound_filter(data) + raise ValueError(f"Unknown predicate kind: {kind}") + + +def _build_leaf_filter(data: dict) -> Callable: + transform = data["transform"] + function = data["function"] + literals = data.get("literals", []) + + def filter_fn(batch: pa.RecordBatch) -> pa.Array: + value_array = _apply_predicate_transform(transform, batch) + return _apply_leaf_function(function, value_array, literals, len(batch)) + + return filter_fn + + +def _build_compound_filter(data: dict) -> Callable: + function = data["function"] + child_filters = [_build_filter(child) for child in data["children"]] + + def filter_fn(batch: pa.RecordBatch) -> pa.Array: + if function == "AND": + result = child_filters[0](batch) + for cf in child_filters[1:]: + result = pc.and_(result, cf(batch)) + return result + elif function == "OR": + result = child_filters[0](batch) + for cf in child_filters[1:]: + result = pc.or_(result, cf(batch)) + return result + raise ValueError(f"Unknown compound function: {function}") + + return filter_fn + + +def _apply_predicate_transform(transform: dict, batch: pa.RecordBatch) -> pa.Array: + name = transform["name"] + + if name == "FIELD_REF": + return batch.column(transform["fieldRef"]["name"]) + + elif name == "CAST": + col = batch.column(transform["fieldRef"]["name"]) + target_type = _paimon_type_to_arrow(transform["type"]) + return pc.cast(col, target_type) + + elif name == "UPPER": + input_col = _resolve_transform_input(transform["inputs"][0], batch) + return pc.utf8_upper(input_col) + + elif name == "LOWER": + input_col = _resolve_transform_input(transform["inputs"][0], batch) + return pc.utf8_lower(input_col) + + elif name == "CONCAT": + resolved = [_resolve_transform_input(inp, batch) for inp in transform["inputs"]] + if not resolved: + return pa.nulls(len(batch), type=pa.string()) + return pc.binary_join_element_wise(*resolved, "") + + elif name == "CONCAT_WS": + sep = _resolve_transform_input(transform["inputs"][0], batch) + values = [_resolve_transform_input(inp, batch) for inp in transform["inputs"][1:]] + if not values: + return pa.nulls(len(batch), type=pa.string()) + result = pc.binary_join_element_wise(*values, sep, null_handling='skip') + if len(result) < len(batch): + padded = result.to_pylist() + [None] * (len(batch) - len(result)) Review Comment: Padding the compacted Arrow result at the end misaligns rows. drops rows whose value inputs are all null, so for , , it returns ; this code pads that to , while Java should evaluate the rows as . In a row-filter predicate that can allow or deny the wrong rows. Please make the CONCAT_WS transform row-preserving, and add a mixed all-null/non-null regression test. ########## paimon-python/pypaimon/read/reader/auth_masking_reader.py: ########## @@ -0,0 +1,229 @@ +################################################################################ +# 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. +################################################################################ + +import json +from typing import Callable, Dict, List, Optional + +import pyarrow as pa +import pyarrow.compute as pc + +from pypaimon.common.predicate_json_parser import ( + _collect_all_field_refs_from_transform, + _paimon_type_to_arrow, +) +from pypaimon.read.reader.iface.record_batch_reader import RecordBatchReader + + +class RecordReaderToBatchAdapter(RecordBatchReader): + + def __init__(self, inner, schema: pa.Schema, chunk_size: int = 65536, include_row_kind: bool = False): + self._inner = inner + self._schema = schema + self._chunk_size = chunk_size + self._exhausted = False + self._pending_iterator = None + self._include_row_kind = include_row_kind + + def read_arrow_batch(self) -> Optional[pa.RecordBatch]: + if self._exhausted: + return None + row_tuples = [] + row_kinds = [] + while len(row_tuples) < self._chunk_size: + if self._pending_iterator is not None: + row = self._pending_iterator.next() + while row is not None: + row_tuples.append( + row.row_tuple[row.offset:row.offset + row.arity]) + if self._include_row_kind: + row_kinds.append(row.get_row_kind().to_string()) + if len(row_tuples) >= self._chunk_size: + return self._flush(row_tuples, row_kinds) + row = self._pending_iterator.next() + self._pending_iterator = None + + row_iterator = self._inner.read_batch() + if row_iterator is None: + self._exhausted = True + break + self._pending_iterator = row_iterator + + if not row_tuples: + return None + return self._flush(row_tuples, row_kinds) + + def _flush(self, row_tuples, row_kinds=None): + columns_data = list(zip(*row_tuples)) + pydict = { + name: list(col) + for name, col in zip(self._schema.names, columns_data) + } + batch = pa.RecordBatch.from_pydict(pydict, schema=self._schema) + if row_kinds: + row_kind_array = pa.array(row_kinds, type=pa.string()) + row_kind_field = pa.field("_row_kind", pa.string()) + new_schema = pa.schema([row_kind_field] + list(batch.schema)) + columns = [row_kind_array] + [batch.column(i) for i in range(batch.num_columns)] + batch = pa.RecordBatch.from_arrays(columns, schema=new_schema) + return batch + + def close(self): + self._inner.close() + + +class AuthFilterReader(RecordBatchReader): + + def __init__(self, inner_reader: RecordBatchReader, filter_fn: Callable[[pa.RecordBatch], pa.Array]): + self._inner = inner_reader + self._filter_fn = filter_fn + + def read_arrow_batch(self) -> Optional[pa.RecordBatch]: + batch = self._inner.read_arrow_batch() + if batch is None: + return None + mask = self._filter_fn(batch) + return batch.filter(mask) + + def close(self): + self._inner.close() + + +class AuthMaskingReader(RecordBatchReader): + + def __init__(self, inner_reader: RecordBatchReader, masking_rules: Dict[str, str], read_fields: List): + self._inner = inner_reader + self._masking_rules = masking_rules + self._read_fields = read_fields + read_field_names = {f.name for f in read_fields} + # Filter to projected columns, then validate field references. + self._parsed_rules = { + col: json.loads(tj) for col, tj in masking_rules.items() + if col in read_field_names + } + for col_name, transform in self._parsed_rules.items(): + for ref_name in _collect_all_field_refs_from_transform(transform): + if ref_name not in read_field_names: + raise RuntimeError( + f"Column masking refers to field '{ref_name}' which is not present " + f"in output row type. Available fields: {read_field_names}" + ) + + def read_arrow_batch(self) -> Optional[pa.RecordBatch]: + batch = self._inner.read_arrow_batch() + if batch is None: + return None + original_batch = batch + masked_columns = {} + for col_name, transform in self._parsed_rules.items(): + if col_name in original_batch.schema.names: + col_idx = original_batch.schema.get_field_index(col_name) + target_col_type = original_batch.schema.field(col_idx).type + masked_columns[col_idx] = self._apply_transform(transform, original_batch, target_col_type) + for col_idx, masked_array in masked_columns.items(): + original_field = original_batch.schema.field(col_idx) + if masked_array.type != original_field.type: + masked_array = pc.cast(masked_array, original_field.type) + batch = batch.set_column( + col_idx, pa.field(original_field.name, original_field.type, nullable=True), masked_array) + return batch + + def close(self): + self._inner.close() + + def _apply_transform( + self, + transform: dict, + original_batch: pa.RecordBatch, + target_col_type: pa.DataType, + ) -> pa.Array: + name = transform["name"] + + if name == "NULL": + return pa.nulls(len(original_batch), type=target_col_type) + + elif name == "FIELD_REF": + ref_name = transform["fieldRef"]["name"] + return original_batch.column(ref_name) + + elif name == "CAST": + ref_name = transform["fieldRef"]["name"] + source_col = original_batch.column(ref_name) + target_type = _paimon_type_to_arrow(transform["type"]) + return pc.cast(source_col, target_type) + + elif name == "UPPER": + input_col = self._resolve_input(transform["inputs"][0], original_batch) + return pc.utf8_upper(input_col) + + elif name == "LOWER": + input_col = self._resolve_input(transform["inputs"][0], original_batch) + return pc.utf8_lower(input_col) + + elif name == "CONCAT": + return self._apply_concat(transform["inputs"], original_batch) + + elif name == "CONCAT_WS": + return self._apply_concat_ws(transform["inputs"], original_batch) + + raise ValueError(f"Unknown transform type: {name}") + + def _resolve_input(self, inp, original_batch: pa.RecordBatch) -> pa.Array: + if isinstance(inp, dict): + return original_batch.column(inp["name"]) + elif isinstance(inp, str): + return pa.array([inp] * len(original_batch), type=pa.string()) + elif inp is None: + return pa.nulls(len(original_batch), type=pa.string()) + return pa.array([str(inp)] * len(original_batch), type=pa.string()) + + def _apply_concat(self, inputs: list, original_batch: pa.RecordBatch) -> pa.Array: + resolved = [self._resolve_input(inp, original_batch) for inp in inputs] + if not resolved: + return pa.nulls(len(original_batch), type=pa.string()) + return pc.binary_join_element_wise(*resolved, "") + + def _apply_concat_ws(self, inputs: list, original_batch: pa.RecordBatch) -> pa.Array: + if len(inputs) < 2: + return pa.nulls(len(original_batch), type=pa.string()) + sep = self._resolve_input(inputs[0], original_batch) + values = [self._resolve_input(inp, original_batch) for inp in inputs[1:]] + result = pc.binary_join_element_wise(*values, sep, null_handling='skip') + if len(result) < len(original_batch): + padded = result.to_pylist() + [None] * (len(original_batch) - len(result)) Review Comment: This has the same row-alignment problem as the predicate CONCAT_WS path. Arrow returns a compacted array when some rows have all-null value inputs, so appending padding at the end shifts the later rows' masked values. For example + with separator becomes here, but Java would produce . That masks the wrong rows; please preserve original row positions and cover the mixed all-null/non-null case in tests. -- 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]
