voonhous commented on code in PR #18961: URL: https://github.com/apache/hudi/pull/18961#discussion_r3446051491
########## hudi-spark-datasource/hudi-spark4-common/src/main/java/org/apache/hudi/variant/Spark4VariantShreddingProvider.java: ########## @@ -0,0 +1,672 @@ +/* + * 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.hudi.variant; + +import org.apache.hudi.avro.VariantShreddingProvider; +import org.apache.hudi.common.schema.HoodieSchema; + +import org.apache.avro.Conversions; +import org.apache.avro.LogicalType; +import org.apache.avro.LogicalTypes; +import org.apache.avro.Schema; +import org.apache.avro.generic.GenericData; +import org.apache.avro.generic.GenericFixed; +import org.apache.avro.generic.GenericRecord; +import org.apache.spark.types.variant.ShreddingUtils; +import org.apache.spark.types.variant.Variant; +import org.apache.spark.types.variant.VariantSchema; +import org.apache.spark.types.variant.VariantShreddingWriter; +import org.apache.spark.types.variant.VariantShreddingWriter.ShreddedResult; +import org.apache.spark.types.variant.VariantShreddingWriter.ShreddedResultBuilder; + +import java.math.BigDecimal; +import java.nio.ByteBuffer; +import java.util.ArrayList; +import java.util.IdentityHashMap; +import java.util.List; +import java.util.Map; +import java.util.UUID; + +/** + * Implementation of {@link VariantShreddingProvider} using Spark 4's variant parsing library. + * + * <p>This class bridges the Avro record path and Spark's {@link VariantShreddingWriter} + * to allow {@code HoodieRecordType.AVRO} to write shredded variant types. It converts + * the shredded output into Avro {@link GenericRecord}s that can be written via + * {@link org.apache.hudi.avro.HoodieAvroWriteSupport}.</p> + * + * <p>The shredding logic is delegated to {@link VariantShreddingWriter#castShredded}, + * which handles scalar, object, and array shredding including residual value construction + * for non-matching fields. This class implements the {@link ShreddedResult} and + * {@link ShreddedResultBuilder} interfaces to collect the shredded components into + * Avro GenericRecords.</p> + */ +public class Spark4VariantShreddingProvider implements VariantShreddingProvider { + + private static final String VALUE_FIELD = "value"; + private static final String METADATA_FIELD = "metadata"; + private static final String TYPED_VALUE_FIELD = "typed_value"; + + @Override + public GenericRecord shredVariantRecord( + GenericRecord unshreddedVariant, + Schema shreddedSchema, + HoodieSchema.Variant variantSchema) { + + ByteBuffer valueBuf = (ByteBuffer) unshreddedVariant.get(VALUE_FIELD); + ByteBuffer metadataBuf = (ByteBuffer) unshreddedVariant.get(METADATA_FIELD); + + if (valueBuf == null || metadataBuf == null) { + return null; + } + + byte[] valueBytes = toByteArray(valueBuf); + byte[] metadataBytes = toByteArray(metadataBuf); + + Variant variant = new Variant(valueBytes, metadataBytes); + + // Build VariantSchema from the Avro shredded schema, registering + // Avro schemas at each level for GenericRecord construction. + AvroShreddedResultBuilder builder = new AvroShreddedResultBuilder(); + VariantSchema sparkSchema = buildVariantSchema(shreddedSchema, true, builder); + + // Delegate to Spark's VariantShreddingWriter for the actual shredding logic. + AvroShreddedResult result = (AvroShreddedResult) + VariantShreddingWriter.castShredded(variant, sparkSchema, builder); + + return result.toGenericRecord(); + } + + @Override + public GenericRecord rebuildVariantRecord( + GenericRecord shreddedVariant, + Schema shreddedSchema, + Schema unshreddedSchema) { + + if (shreddedVariant == null) { + return null; + } + ByteBuffer metadataBuf = (ByteBuffer) shreddedVariant.get(METADATA_FIELD); + if (metadataBuf == null) { + return null; + } + + // Reuse the same VariantSchema index assignment as the write path (no builder needed on read). + VariantSchema sparkSchema = buildVariantSchema(shreddedSchema, true, null); + + // Delegate to Spark's reconstruction algorithm (inverse of castShredded). + Variant variant = ShreddingUtils.rebuild(new AvroVariantRow(shreddedVariant, sparkSchema), sparkSchema); + + GenericRecord out = new GenericData.Record(unshreddedSchema); + out.put(METADATA_FIELD, ByteBuffer.wrap(variant.getMetadata())); + out.put(VALUE_FIELD, ByteBuffer.wrap(variant.getValue())); + return out; + } + + /** + * Builds a {@link VariantSchema} from an Avro {@link Schema} representing a + * shredded variant structure ({@code value}, {@code metadata}, {@code typed_value}). + * + * <p>This method also registers the Avro schema mapping in the builder so that + * {@link AvroShreddedResultBuilder#createEmpty} can create results with the + * correct Avro schema at each nesting level.</p> + */ + private VariantSchema buildVariantSchema(Schema avroSchema, boolean isTopLevel, + AvroShreddedResultBuilder builder) { + Schema.Field valueField = avroSchema.getField(VALUE_FIELD); + Schema.Field metadataField = avroSchema.getField(METADATA_FIELD); + Schema.Field typedValueField = avroSchema.getField(TYPED_VALUE_FIELD); + + int idx = 0; + int variantIdx = valueField != null ? idx++ : -1; + int topLevelMetadataIdx; + if (metadataField != null && isTopLevel) { + topLevelMetadataIdx = idx++; + } else { + topLevelMetadataIdx = -1; + if (metadataField != null) { + idx++; + } + } + int typedIdx = typedValueField != null ? idx++ : -1; + int numFields = idx; + + VariantSchema.ScalarType scalarSchema = null; + VariantSchema.ObjectField[] objectSchema = null; + VariantSchema arraySchema = null; + + if (typedValueField != null) { + Schema tvSchema = unwrapNullable(typedValueField.schema()); + + switch (tvSchema.getType()) { + case RECORD: + // Object shredding: each field has a nested {value, typed_value} sub-struct + List<VariantSchema.ObjectField> fields = new ArrayList<>(); + for (Schema.Field field : tvSchema.getFields()) { + Schema fieldSchema = unwrapNullable(field.schema()); + VariantSchema subSchema = buildVariantSchema(fieldSchema, false, builder); + fields.add(new VariantSchema.ObjectField(field.name(), subSchema)); + } + objectSchema = fields.toArray(new VariantSchema.ObjectField[0]); + break; + + case ARRAY: + // Array shredding: elements follow the shredding schema + Schema elementSchema = unwrapNullable(tvSchema.getElementType()); + arraySchema = buildVariantSchema(elementSchema, false, builder); + break; + + default: + // Scalar shredding + scalarSchema = avroTypeToScalarType(tvSchema); + break; + } + } + + VariantSchema result = new VariantSchema( + typedIdx, variantIdx, topLevelMetadataIdx, numFields, + scalarSchema, objectSchema, arraySchema); + + // The read (rebuild) path passes a null builder: it needs the VariantSchema indices but no + // Avro-schema registration (registration only feeds write-side result construction). + if (builder != null) { + builder.registerSchema(result, avroSchema); + } + + return result; + } + + /** + * Maps an Avro {@link Schema} type (potentially with logical type annotations) + * to a {@link VariantSchema.ScalarType}. + */ + private VariantSchema.ScalarType avroTypeToScalarType(Schema schema) { + LogicalType logicalType = schema.getLogicalType(); + + // Check logical types first + if (logicalType != null) { + if (logicalType instanceof LogicalTypes.Decimal) { + LogicalTypes.Decimal decimal = (LogicalTypes.Decimal) logicalType; + return new VariantSchema.DecimalType(decimal.getPrecision(), decimal.getScale()); + } + String name = logicalType.getName(); + if ("date".equals(name)) { + return new VariantSchema.DateType(); + } + if ("timestamp-micros".equals(name)) { + return new VariantSchema.TimestampType(); + } Review Comment: Good catch -- it's currently unreachable but a real latent trap, so I hardened it. The Variant binary spec stores timestamps in microseconds, so a `timestamp-millis` typed_value can't represent a variant timestamp; mapping it to the micros `TimestampType` would scale the value 1000x on rebuild. Made millis decline instead, so the value stays in the residual unshredded binary: ```java // The Variant binary spec stores timestamps in microseconds, so a millisecond-precision // typed_value cannot represent a variant timestamp. Decline to shred it as a scalar (the value // stays in the residual unshredded binary) rather than mapping it to the micros TimestampType, // which would silently scale the value by 1000x. if ("timestamp-millis".equals(name) || "local-timestamp-millis".equals(name)) { return null; } ``` In practice this never fires (Hudi writes typed_value via `createTimestampMicros`, the read path only reads Hudi-shredded files, and the Spark inferrer produces micros from micros-only variant samples), so behavior is unchanged -- this just removes the silent-corruption path. `avroTypeToScalarType` is shared write/read code and the concern is sharpest on the rebuild path, so this landed in the base PR #18938; it'll flow into this PR on the next rebase. -- 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]
