yihua commented on code in PR #18432: URL: https://github.com/apache/hudi/pull/18432#discussion_r3041968578
########## hudi-spark-datasource/hudi-spark-common/src/main/scala/org/apache/spark/sql/hudi/analysis/VectorDistanceUtils.scala: ########## @@ -0,0 +1,161 @@ +/* + * 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.ml.linalg.{DenseVector, Vectors} +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. + * + * 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}") + val va = new DenseVector(a) + val vb = new DenseVector(b) + val denom = Vectors.norm(va, 2.0) * Vectors.norm(vb, 2.0) + if (denom == 0.0) 1.0 else math.min(2.0, math.max(0.0, 1.0 - (va.dot(vb) / 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}") + math.sqrt(Vectors.sqdist(new DenseVector(a), new DenseVector(b))) + } + + /** + * 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}") + -(new DenseVector(a)).dot(new DenseVector(b)) + } + + /** + * Creates a Spark UDF for the given distance metric and corpus element type. + * Both arguments (corpus vector and query vector) are evaluated per row. + * Prefer [[createSingleQueryDistanceUdf]] when the query vector is constant. + * 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") + } + + /** + * Creates a Spark UDF optimized for single-query mode: the query vector's + * [[DenseVector]] is pre-computed once and closed over, so only the corpus + * vector is converted per row. + */ + def createSingleQueryDistanceUdf( + metric: DistanceMetric.Value, + elementType: DataType, + queryVector: Array[Double]): UserDefinedFunction = { + val queryDv = new DenseVector(queryVector) + val queryNorm = Vectors.norm(queryDv, 2.0) + val distFn = resolveDistanceFn(metric) + + elementType match { + case FloatType => udf((corpus: Seq[Float]) => { + requireSameLength(corpus.length, queryVector.length) + distFn(new DenseVector(corpus.iterator.map(_.toDouble).toArray), queryDv, queryNorm) + }) + case DoubleType => udf((corpus: Seq[Double]) => { Review Comment: Got it. Overall, my concern is that per-vector processing through `DenseVector` could introduce latency overhead. Given this is the initial implementation, we can check in the code and should follow up with micro-benchmarks. -- 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]
