MaxGekk commented on code in PR #55326: URL: https://github.com/apache/spark/pull/55326#discussion_r3382213359
########## sql/core/src/main/scala/org/apache/spark/sql/execution/datasources/parquet/types/ops/ParquetTypeOps.scala: ########## @@ -0,0 +1,219 @@ +/* + * 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. + */ + +package org.apache.spark.sql.execution.datasources.parquet.types.ops + +import org.apache.parquet.io.api.RecordConsumer +import org.apache.parquet.schema.Type +import org.apache.parquet.schema.Type.Repetition + +import org.apache.spark.sql.catalyst.expressions.SpecializedGetters +import org.apache.spark.sql.catalyst.util.RebaseDateTime.RebaseSpec +import org.apache.spark.sql.execution.datasources.parquet.{HasParentContainerUpdater, ParentContainerUpdater} +import org.apache.spark.sql.internal.SQLConf +import org.apache.spark.sql.types.{DataType, StructType, TimeType} + +/** + * Optional trait for Parquet storage format integration in the Types Framework. + * + * Implement this trait to enable Parquet read/write support for a framework type. Each framework + * type that supports Parquet provides a concrete implementation and registers it in the companion + * object's apply() method. + * + * The trait covers all Parquet concerns: + * - Schema conversion: Spark DataType <-> Parquet schema type + * - Value write: writing values to Parquet RecordConsumer + * - Row-based read: creating Parquet converters for reading into InternalRow + * - Vectorized read: creating batch updaters for columnar reading + * - Filter pushdown: creating Parquet filter predicates for predicate pushdown + * - Type gates: declaring Parquet support/capability + * - Schema clipping: declaring internal struct schema for column pruning + * + * DISPATCH PATTERN: Framework FIRST at all integration sites. Each Parquet infrastructure + * method wraps itself with: + * {{{ + * ParquetTypeOps(dt).map(_.method(...)).getOrElse(methodDefault(dt, ...)) + * }}} + * The original code is extracted to a *Default method unchanged. When the framework flag is ON, + * the ops handles the type. When OFF, the *Default fallback executes the original code path. + * + * STRUCT-BACKED TYPES: Types stored as Parquet groups should override the + * extended newConverter overload (which provides schemaConverter/convertTz/rebase specs for + * recursive field conversion) and declare parquetStructSchema for column pruning. + * + * @see TimeTypeParquetOps for a reference implementation (primitive Long-backed type) + * @since 4.3.0 + */ +private[parquet] trait ParquetTypeOps extends Serializable { + + /** The DataType this Ops instance handles. */ + def dataType: DataType + + // ==================== Schema Conversion ==================== + + /** + * Converts this Spark DataType to a Parquet schema Type (for the write path). + * + * For primitive types: returns a PrimitiveType with the appropriate annotation. + * For struct-backed types: returns a GroupType with sub-fields and a logical type annotation. + * + * @param fieldName the column/field name in the Parquet schema + * @param repetition REQUIRED, OPTIONAL, or REPEATED + * @return the Parquet Type for this DataType + */ + def convertToParquetType(fieldName: String, repetition: Repetition): Type + + // ==================== Value Write ==================== + + /** + * Creates a value writer that writes values of this type to a Parquet RecordConsumer. + * + * The writer is a closure that captures the RecordConsumer and is called once per row during + * Parquet file writing. The RecordConsumer is instance-specific (one per output file). + * + * IMPORTANT: The RecordConsumer is passed as a lazy supplier (() => RecordConsumer) because + * makeWriter is called during ParquetWriteSupport.init(), when recordConsumer is still null. + * It gets set later in prepareForWrite(). The closure must evaluate the supplier at write + * time, not at creation time. The existing infrastructure closures work because they capture + * `this.recordConsumer` (a var field - Scala closures over vars capture the reference, not + * the value). The supplier lambda achieves the same lazy evaluation for ops code. + * + * For primitive types: directly calls recordConsumer.addLong/addBinary/etc. with any + * necessary conversion (e.g., nanos to micros for TimeType). + * + * @param recordConsumer lazy supplier for the Parquet output stream (null during init) + * @param makeFieldWriter callback into ParquetWriteSupport for recursive field writer creation + * (struct-backed types use this to create writers for sub-fields) + * @return a closure that writes a value from a row at the given ordinal + */ + def makeWriter( + recordConsumer: () => RecordConsumer, + makeFieldWriter: DataType => (SpecializedGetters, Int) => Unit + ): (SpecializedGetters, Int) => Unit + + // ==================== Row-Based Read ==================== + + /** + * Creates a Parquet Converter for reading values of this type (simple version). + * + * Primitive types override this method. The converter typically extends + * ParquetPrimitiveConverter and overrides addLong/addBinary/etc. to perform any necessary + * conversion (e.g., micros to nanos for TimeType). + * + * WARNING: Struct-backed types must override the EXTENDED overload below instead. This simple + * version does not provide the schemaConverter, timezone, or rebase specs needed for recursive + * field conversion. Overriding only this method for a struct-backed type will compile but + * produce incorrect behavior at runtime (missing timezone conversion, calendar rebasing, etc.). + * + * @param parquetType the Parquet schema type for this field + * @param updater the parent container to set converted values into + * @return a Converter that reads Parquet values into the parent container + */ + def newConverter( + parquetType: org.apache.parquet.schema.Type, + updater: ParentContainerUpdater): org.apache.parquet.io.api.Converter + with HasParentContainerUpdater Review Comment: Minor design suggestion: the abstract method here is the *simple* overload, while the extended overload (below) has a default that delegates to it. As the scaladoc warns, struct-backed types must override the extended overload, yet they are still forced to provide an implementation of this simple one that the docs say is wrong for them ("will compile but produce incorrect behavior at runtime"). Consider making the *extended* overload the abstract one (and giving the simple overload a default that throws / delegates), so future struct-backed implementations aren't required to ship a misleading simple impl. Not blocking for this PR since the only reference type (`TimeType`) is primitive and correctly uses the simple version - just something to tidy before a struct-backed type lands. -- 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] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
