voonhous commented on code in PR #18432: URL: https://github.com/apache/hudi/pull/18432#discussion_r3022275190
########## hudi-spark-datasource/hudi-spark-common/src/main/scala/org/apache/spark/sql/hudi/analysis/VectorDistanceUtils.scala: ########## @@ -0,0 +1,159 @@ +/* + * 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.hudi.analysis + +import org.apache.spark.sql.catalyst.plans.logical.HoodieVectorSearchTableValuedFunction.DistanceMetric +import org.apache.spark.sql.expressions.UserDefinedFunction +import org.apache.spark.sql.functions.udf +import org.apache.spark.sql.hudi.command.exception.HoodieAnalysisException +import org.apache.spark.sql.types.{ByteType, DataType, DoubleType, FloatType} + +/** + * Vector distance utilities: raw distance functions and Spark UDF factories. + * + * Raw functions operate on Array[Double] and can be used outside of Spark + * (e.g. for verification, index building, or non-Spark distance computation). + * + * UDF factories produce typed Spark UDFs for Float, Double, and Byte corpus columns. + */ +object VectorDistanceUtils { + + /** + * Cosine distance: 1 - (a . b) / (||a|| * ||b||). + * Returns 0.0 for identical vectors, 1.0 for orthogonal, 2.0 for opposite. + * Returns 1.0 if either vector is zero (convention: maximal distance). + */ + def cosineDistance(a: Array[Double], b: Array[Double]): Double = { + require(a.length == b.length, s"Vector dimension mismatch: ${a.length} vs ${b.length}") + var dot = 0.0; var normA = 0.0; var normB = 0.0; var i = 0 + while (i < a.length) { + dot += a(i) * b(i); normA += a(i) * a(i); normB += b(i) * b(i); i += 1 + } + val denom = math.sqrt(normA) * math.sqrt(normB) + if (denom == 0.0) 1.0 else 1.0 - (dot / denom) Review Comment: When vectors are nearly identical, `dot / denom` can exceed 1.0 due to floating-point rounding, making `1.0 - x` slightly negative. The result returns [0, 2], but this breaks this assumption, and could cause confusing negative `_hudi_distance` values. We can use: `math.max(0.0, 1.0 - (dot / denom))` or clamp to `[0.0, 2.0]`. ########## hudi-spark-datasource/hudi-spark-common/src/main/scala/org/apache/spark/sql/hudi/analysis/VectorDistanceUtils.scala: ########## @@ -0,0 +1,159 @@ +/* + * 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.hudi.analysis + +import org.apache.spark.sql.catalyst.plans.logical.HoodieVectorSearchTableValuedFunction.DistanceMetric +import org.apache.spark.sql.expressions.UserDefinedFunction +import org.apache.spark.sql.functions.udf +import org.apache.spark.sql.hudi.command.exception.HoodieAnalysisException +import org.apache.spark.sql.types.{ByteType, DataType, DoubleType, FloatType} + +/** + * Vector distance utilities: raw distance functions and Spark UDF factories. + * + * Raw functions operate on Array[Double] and can be used outside of Spark + * (e.g. for verification, index building, or non-Spark distance computation). + * + * UDF factories produce typed Spark UDFs for Float, Double, and Byte corpus columns. + */ +object VectorDistanceUtils { + + /** + * Cosine distance: 1 - (a . b) / (||a|| * ||b||). + * Returns 0.0 for identical vectors, 1.0 for orthogonal, 2.0 for opposite. + * Returns 1.0 if either vector is zero (convention: maximal distance). + */ + def cosineDistance(a: Array[Double], b: Array[Double]): Double = { + require(a.length == b.length, s"Vector dimension mismatch: ${a.length} vs ${b.length}") + var dot = 0.0; var normA = 0.0; var normB = 0.0; var i = 0 + while (i < a.length) { + dot += a(i) * b(i); normA += a(i) * a(i); normB += b(i) * b(i); i += 1 + } + val denom = math.sqrt(normA) * math.sqrt(normB) + if (denom == 0.0) 1.0 else 1.0 - (dot / denom) + } + + /** + * L2 (Euclidean) distance: sqrt(sum((a[i] - b[i])^2)). + * Returns 0.0 for identical vectors. + */ + def l2Distance(a: Array[Double], b: Array[Double]): Double = { + require(a.length == b.length, s"Vector dimension mismatch: ${a.length} vs ${b.length}") + var sum = 0.0; var i = 0 + while (i < a.length) { val d = a(i) - b(i); sum += d * d; i += 1 } + math.sqrt(sum) + } + + /** + * Negated dot product distance: -(a . b). + * Lower values indicate higher similarity (for ascending sort compatibility). + */ + def dotProductDistance(a: Array[Double], b: Array[Double]): Double = { + require(a.length == b.length, s"Vector dimension mismatch: ${a.length} vs ${b.length}") + var dot = 0.0; var i = 0 + while (i < a.length) { dot += a(i) * b(i); i += 1 } + -dot + } + + // ======================== Spark UDF Factories ======================== + + /** + * Creates a Spark UDF for the given distance metric and corpus element type. + * Supports Float, Double, and Byte element types. + */ + def createDistanceUdf(metric: DistanceMetric.Value, elementType: DataType): UserDefinedFunction = + elementType match { + case FloatType => createFloatDistanceUdf(metric) + case DoubleType => createDoubleDistanceUdf(metric) + case ByteType => createByteDistanceUdf(metric) + case _ => throw new HoodieAnalysisException( + s"Unsupported vector element type for distance computation: $elementType") + } + + private def createFloatDistanceUdf(metric: DistanceMetric.Value): UserDefinedFunction = metric match { + case DistanceMetric.COSINE => udf((a: Seq[Float], b: Seq[Float]) => { + requireSameLength(a.length, b.length) + var dot = 0.0f; var normA = 0.0f; var normB = 0.0f; var i = 0 Review Comment: For high-dimensional vectors (hundreds/thousands of dims), Float accumulation loses significant precision. The cosine distance `denominator sqrt(normA) * sqrt(normB)` magnifies the error. Possible to do acucmulation or intermediate calcs in double? ########## hudi-spark-datasource/hudi-spark-common/src/main/scala/org/apache/spark/sql/hudi/analysis/HoodieSparkBaseAnalysis.scala: ########## @@ -310,6 +332,50 @@ case class ResolveReferences(spark: SparkSession) extends Rule[LogicalPlan] sparkAdapter.getCatalystPlanUtils.unapplyMergeIntoTable(plan) } + private def resolveTableToDf(tableName: String): DataFrame = { + if (tableName.contains(StoragePath.SEPARATOR)) { + spark.read.format("hudi").load(tableName) + } else { + spark.table(tableName) + } + } + + private def evaluateQueryVector(expr: Expression): Array[Double] = { + if (!expr.foldable) { + throw new HoodieAnalysisException( + s"Function '${HoodieVectorSearchTableValuedFunction.FUNC_NAME}': " + + "query vector must be a constant expression (e.g., ARRAY(1.0, 2.0, 3.0))") + } + val value = expr.eval(null) + if (value == null) { + throw new HoodieAnalysisException( + s"Function '${HoodieVectorSearchTableValuedFunction.FUNC_NAME}': query vector cannot be null") + } + + val arrayData = value.asInstanceOf[ArrayData] + val numElements = arrayData.numElements() + val elementType = expr.dataType.asInstanceOf[ArrayType].elementType + val result = new Array[Double](numElements) + var i = 0 + elementType match { + case DoubleType => + while (i < numElements) { result(i) = arrayData.getDouble(i); i += 1 } + case FloatType => + while (i < numElements) { result(i) = arrayData.getFloat(i).toDouble; i += 1 } + case IntegerType => + while (i < numElements) { result(i) = arrayData.getInt(i).toDouble; i += 1 } + // Spark SQL infers untyped decimal literals (e.g. ARRAY(1.0, 0.5)) as DecimalType, + // not DoubleType. Accept any DecimalType and convert to Double. + case d: DecimalType => Review Comment: Long Type not supported? Is this intentional? -- 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]
