cloud-fan commented on code in PR #55326: URL: https://github.com/apache/spark/pull/55326#discussion_r3431342321
########## sql/core/src/main/scala/org/apache/spark/sql/execution/datasources/parquet/types/ops/TimeTypeParquetOps.scala: ########## @@ -0,0 +1,117 @@ +/* + * 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.{Converter, RecordConsumer} +import org.apache.parquet.schema.{LogicalTypeAnnotation, Type, Types} +import org.apache.parquet.schema.LogicalTypeAnnotation.TimeUnit +import org.apache.parquet.schema.PrimitiveType.PrimitiveTypeName.INT64 +import org.apache.parquet.schema.Type.Repetition + +import org.apache.spark.sql.catalyst.expressions.SpecializedGetters +import org.apache.spark.sql.catalyst.util.DateTimeUtils +import org.apache.spark.sql.errors.QueryExecutionErrors +import org.apache.spark.sql.execution.datasources.parquet.{HasParentContainerUpdater, ParentContainerUpdater, ParquetPrimitiveConverter} +import org.apache.spark.sql.internal.SQLConf +import org.apache.spark.sql.types.{DataType, TimeType} + +/** + * Parquet operations for TimeType. + * + * TimeType is a primitive Long-backed type stored in Parquet as INT64 with the + * TIME(isAdjustedToUTC=false, unit=MICROS) logical type annotation. + * + * IMPORTANT - internal vs Parquet representation: + * - Spark internal: nanoseconds since midnight (Long) + * - Parquet storage: microseconds since midnight (INT64) + * - Write path: nanos -> micros (DateTimeUtils.nanosToMicros) + * - Read path: micros -> nanos (DateTimeUtils.microsToNanos) + * + * @param t the TimeType with precision information + * @since 4.3.0 + */ +case class TimeTypeParquetOps(t: TimeType) extends ParquetTypeOps { + + // ==================== Schema Conversion ==================== + + override def convertToParquetType( + fieldName: String, repetition: Repetition, inShredded: Boolean): Type = + Types.primitive(INT64, repetition) + .as(LogicalTypeAnnotation.timeType(false, TimeUnit.MICROS)) + .named(fieldName) + + // ==================== Value Write ==================== + + override def makeWriter( + recordConsumer: () => RecordConsumer, + makeFieldWriter: DataType => (SpecializedGetters, Int) => Unit + ): (SpecializedGetters, Int) => Unit = + // Evaluate the supplier at write time (not creation time) because recordConsumer + // is null during init() and set later in prepareForWrite(). + (row: SpecializedGetters, ordinal: Int) => + recordConsumer().addLong(DateTimeUtils.nanosToMicros(row.getLong(ordinal))) + + // ==================== Row-Based Read ==================== + + override def newConverter( + parquetType: Type, + updater: ParentContainerUpdater): Converter with HasParentContainerUpdater = { + // Framework-first dispatch in ParquetRowConverter routes here whenever the + // requested Spark type is TimeType, regardless of the actual Parquet encoding. + // Without this guard, files whose column is raw INT64, INT64 TIME(NANOS), + // INT64 TIMESTAMP(MICROS), INT32 TIME(MILLIS), etc. would silently decode as + // microsToNanos(value) and produce wrong results. Mirrors the inline guard + // that existed in ParquetRowConverter before the framework dispatch. + TimeTypeParquetOps.requireCompatibleParquetType(t, parquetType) + new ParquetPrimitiveConverter(updater) { + override def addLong(value: Long): Unit = { + this.updater.setLong(DateTimeUtils.microsToNanos(value)) + } + } + } + + // ==================== Vectorized Read ==================== + + override def isBatchReadSupported(sqlConf: SQLConf): Boolean = true Review Comment: `isBatchReadSupported = true` here makes the legacy vectorized path the default for a top-level `TimeType`, and that path bypasses `requireCompatibleParquetType` entirely. `ParquetVectorUpdaterFactory` (line 168) matches `sparkType instanceof TimeType` with no annotation check and returns `LongAsNanosUpdater`, which blindly does `micros→nanos`. So the row-path guard's `isAdjustedToUTC=true` rejection (and its raw-INT64 / TIME(NANOS) protection) only fires when reads go row-based — vectorized reader off, or the column nested in a struct/array/map. Reading `INT64 TIME(MICROS, isAdjustedToUTC=true)` as `TimeType` via an explicit read schema therefore succeeds vectorized (the default) but fails with `FAILED_READ_FILE` row-based. This is decisive evidence for SPARK-57416 rather than a new ask: the guard's tightening is real but not "unconditional" — its observable reach depends on `enableVectorizedReader` and nesting. Worth folding into the SPARK-57416 framing, since the vectorized path is a second lever, not just `requireCompatibleParquetType`. Flagging it as my own late catch from the prior round (c0c7a0d), where I verified 6-site equivalence but missed that the vectorized path is the one site the row guard never sees. -- 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]
