chrevanthreddy commented on code in PR #20107: URL: https://github.com/apache/hudi/pull/20107#discussion_r4384293478
########## hudi-spark-datasource/hudi-spark-common/src/main/scala/org/apache/spark/sql/hudi/analysis/VectorExactSearchPlanner.scala: ########## @@ -0,0 +1,390 @@ +/* + * 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.hudi.SparkAdapterSupport +import org.apache.hudi.client.common.HoodieSparkEngineContext +import org.apache.hudi.common.config.HoodieMetadataConfig +import org.apache.hudi.common.config.TypedProperties +import org.apache.hudi.common.index.vector.{ScoredVectorPostingMatch, VectorDistanceMetric, VectorIndexArbiter, VectorIndexMdtSearchUtils, VectorIndexMetadataCache, VectorIndexOptions, VectorPostingMatch, VectorStalePolicy} +import org.apache.hudi.common.model.{FileSlice, HoodieIndexDefinition, HoodieRecord} +import org.apache.hudi.common.schema.HoodieSchema +import org.apache.hudi.common.table.{HoodieTableMetaClient, TableSchemaResolver} +import org.apache.hudi.common.table.timeline.{HoodieTimeline, InstantComparison} +import org.apache.hudi.common.table.timeline.InstantComparison.compareTimestamps +import org.apache.hudi.common.table.view.HoodieTableFileSystemView +import org.apache.hudi.data.HoodieJavaRDD +import org.apache.hudi.hadoop.fs.HadoopFSUtils +import org.apache.hudi.metadata.{HoodieTableMetadata, HoodieTableMetadataUtil} + +import org.apache.spark.api.java.JavaSparkContext +import org.apache.spark.rdd.RDD +import org.apache.spark.sql.{AnalysisException, DataFrame, Row, SparkSession} +import org.apache.spark.sql.catalyst.InternalRow +import org.apache.spark.sql.catalyst.parser.ParseException +import org.apache.spark.sql.catalyst.plans.logical.HoodieVectorSearchTableValuedFunction +import org.apache.spark.sql.catalyst.plans.logical.HoodieVectorSearchTableValuedFunction.{DistanceMetric, SearchAlgorithm} +import org.apache.spark.sql.catalyst.plans.logical.LogicalPlan +import org.apache.spark.sql.expressions.Window +import org.apache.spark.sql.functions.{broadcast, col, monotonically_increasing_id, row_number} +import org.apache.spark.sql.hudi.command.exception.HoodieAnalysisException +import org.apache.spark.sql.types.{ArrayType, ByteType, DataType, DoubleType, FloatType, Metadata, StructField, StructType} +import org.apache.spark.util.LongAccumulator +import org.slf4j.LoggerFactory + +import scala.collection.JavaConverters._ +import scala.collection.mutable.ArrayBuffer +import scala.util.{Failure, Success, Try} + +/** Exact reranking and file-slice resolution for the IVF/RaBitQ search plan. */ +private[analysis] object VectorExactSearchPlanner extends SparkAdapterSupport { + import HoodieVectorSearchPlanBuilder._ + import IvfRaBitQMdtSearchAlgorithm.{isLogResidentCandidate, isPartitionedRecordIndex, toVectorMetric, VectorSearchPlanContext} + + private val LOG = LoggerFactory.getLogger(getClass) + + def buildRaBitQExactCandidateDf( + spark: SparkSession, + ctx: VectorSearchPlanContext, + outputSchema: StructType, + embeddingCol: String, + vectorSchema: HoodieSchema.Vector, + queryVector: Array[Float], + exactQueryVector: Array[Double], + k: Int): DataFrame = { + val planStartNs = System.nanoTime() + val storageConf = HadoopFSUtils.getStorageConf(spark.sessionState.newHadoopConf()) + val cacheLoadStartNs = System.nanoTime() + val cache = ctx.cache + val mergedOptions = ctx.tuningOptions + val metric = toVectorMetric(ctx.metric) + val residualEncoding = cache.isResidualEncoding + val numProbes = math.min(math.max(1, VectorIndexOptions.getNumProbes(mergedOptions.asJava)), cache.numClusters()) + val topClusters = cache.findTopClusters(queryVector, numProbes, metric) + val topClusterSeq = topClusters.toSeq + val cacheLoadMs = elapsedMs(cacheLoadStartNs) + + // Shard counts from cache — no extra MDT IO. + val shardCounts = cache.getShardCounts(topClusters) + val shardCountsScala = shardCounts.asScala.map { case (clusterId, count) => + clusterId.intValue() -> count.intValue() + }.toMap + val postingPrefixCount = VectorIndexMdtSearchUtils.buildPostingPrefixes(cache.getGenerationId, shardCounts).size() + LOG.info( + s"[vector_search][plan] basePath=${ctx.basePath} indexPartition=${ctx.indexPartition} cacheLoadMs=$cacheLoadMs " + + s"generationId=${cache.getGenerationId} dimension=${cache.getDimension} numClusters=${cache.numClusters()} " + + s"numProbes=$numProbes refineFactor=${VectorIndexOptions.getRefineFactor(mergedOptions.asJava)} residualEncoding=$residualEncoding " + + s"topClusters=${summarizeInts(topClusterSeq)} shardCounts=${summarizeShardCounts(shardCountsScala)} " + + s"postingPrefixes=$postingPrefixCount") + if (shardCounts.isEmpty) { + sparkAdapter.getUnsafeUtils.createDataFrameFromRDD( + spark, spark.sparkContext.emptyRDD[InternalRow], outputSchema) + } else { + val refineK = math.max(k, k * VectorIndexOptions.getRefineFactor(mergedOptions.asJava)) + val emptyDf = sparkAdapter.getUnsafeUtils.createDataFrameFromRDD( + spark, spark.sparkContext.emptyRDD[InternalRow], outputSchema) + + val topCandidates = VectorIndexMdtSearchUtils.scanPostingCandidates( + ctx.metadataTable, ctx.indexPartition, cache.getGenerationId, shardCounts, + queryVector, cache.getDimension, cache.getQuantizerSeed, cache.getRaBitQBits, cache.isAssumeNormalized, + metric, VectorIndexOptions.isRaBitQAsymmetricScoring(mergedOptions.asJava), + residualEncoding, if (residualEncoding) cache.getCentroids else null, refineK) + val topCandidateMaterializeStartNs = System.nanoTime() + val materializedCandidates = try { + topCandidates.collectAsList().asScala.toSeq + } finally { + topCandidates.unpersistWithDependencies() + } + val topCandidateList = materializedCandidates + .filter(candidate => candidate.getFileGroupId != null && candidate.getFileGroupId.nonEmpty) + val candidatesWithoutLocator = materializedCandidates.size - topCandidateList.size + LOG.info( + s"[vector_search][stage][materialize_top_candidates] refineK=$refineK rerankCandidates=${materializedCandidates.size} " + + s"candidatesWithLocators=${topCandidateList.size} candidatesWithoutLocators=$candidatesWithoutLocator " + + s"distinctFileGroups=${topCandidateList.map(_.getFileGroupId).distinct.size} elapsedMs=${elapsedMs(topCandidateMaterializeStartNs)}") + if (candidatesWithoutLocator > 0) { + throw new HoodieAnalysisException( + s"Vector exact search found $candidatesWithoutLocator rerank candidates without file-group locators. " + + s"Exact rerank requires per-candidate posting locators.") + } + + // RFC-109 RLI finalist arbiter. Exact mode uses refreshed RLI locations for every live + // candidate: SERVE retains positional trust; STALE is resolved by key in the live base file. + val arbiterEnabled = VectorIndexOptions.isFinalistArbiterEnabled(mergedOptions.asJava) + val serveCandidates: Seq[ScoredVectorPostingMatch] = + if (arbiterEnabled) { + val arb = VectorIndexMdtSearchUtils.arbitrateMaterializedFinalists( + ctx.metadataTable, topCandidateList.asJava, isPartitionedRecordIndex(ctx)) + val staleFailPolicy = VectorIndexOptions.isStaleLocatorPolicyFail(mergedOptions.asJava) + LOG.info( + s"[vector_search][exact][arbiter] arbiterExclusions{stale=${arb.staleCount}, deleted=${arb.deletedCount}} " + + s"serve=${arb.serve.size} stalePolicy=${if (staleFailPolicy) "fail" else "fallback"}") + if (staleFailPolicy && !arb.stale.isEmpty) { + throw new HoodieAnalysisException( + s"Vector exact search rejected ${arb.staleCount} stale finalist(s) because " + + "vector.stale.locator.policy=FAIL.") + } + arb.serve.asScala.toSeq ++ arb.stale.asScala.toSeq + } else { + topCandidateList + } + + if (serveCandidates.isEmpty) { + emptyDf + } else { + val candidatesByPartition = serveCandidates + .groupBy(candidate => Option(candidate.getLocation).map(_.getPartitionPath).getOrElse("")) + .map { case (partitionPath, candidates) => + partitionPath -> candidates.flatMap(candidate => Option(candidate.getLocation).map(_.getFileId)).toSet + } + val fileSliceMap = preResolveFileSlices( + ctx.metadataTable, ctx.metaClient, ctx.timeline, ctx.latestInstantTime, candidatesByPartition) + if (fileSliceMap.isEmpty) { + throw new HoodieAnalysisException( + s"Vector exact search resolved zero file slices for ${serveCandidates.size} rerank candidates " + + s"across ${candidatesByPartition.values.map(_.size).sum} file groups. Candidate locators are present, so this indicates " + + s"a file-slice resolution bug or stale vector index.") + } + val bFileSliceMap = spark.sparkContext.broadcast(fileSliceMap) + // MOR log-resident candidates carry their raw embedding in a log block, not the base Parquet + // file, so the positional fetcher cannot read them. Split them out and materialize them via a + // merged (base + log) read; base-resident candidates keep the fast positional path. + val (logResidentCandidates, baseCandidates) = serveCandidates.partition { candidate => + bFileSliceMap.value.get(candidate.getLocation.getFileId).exists { slice => + isLogResidentCandidate(candidate, slice.getBaseInstantTime, slice.getLogFiles.findAny().isPresent) + } + } + val exactFetchParallelism = math.max( + 1, + math.min(refineK, math.max(1, spark.sparkContext.defaultParallelism * 4))) + val groupedCandidates = baseCandidates.groupBy(_.getLocation.getFileId).toSeq.sortBy(_._1) + val targetBatchesPerGroup = math.max(1, exactFetchParallelism / math.max(1, groupedCandidates.size)) + val fetchBatches = groupedCandidates.flatMap { case (fgId, candidates) => + val sorted = candidates.sortBy(candidate => candidate.getLocation.getPosition) + val batchSize = math.max(1, math.ceil(sorted.size.toDouble / targetBatchesPerGroup).toInt) + sorted.grouped(batchSize).map(batch => (fgId, batch)).toSeq + } + val byFetchBatch: RDD[(String, Seq[ScoredVectorPostingMatch])] = + if (fetchBatches.isEmpty) { + spark.sparkContext.emptyRDD[(String, Seq[ScoredVectorPostingMatch])] + } else { + spark.sparkContext.parallelize(fetchBatches, math.min(exactFetchParallelism, fetchBatches.size)) + } + LOG.info( + s"[vector_search][stage][exact_read_plan] rerankCandidates=${serveCandidates.size} " + + s"candidateGroups=${groupedCandidates.size} resolvedSlices=${fileSliceMap.size} " + + s"fileGroupPartitions=${byFetchBatch.getNumPartitions} exactFetchParallelism=$exactFetchParallelism") + + val vectorOnlyProjection = mergedOptions.get(VECTOR_ONLY_PROJECTION_OPT) + .orElse(spark.conf.getOption(VECTOR_ONLY_PROJECTION_SPARK_CONF)) + .exists(_.equalsIgnoreCase("true")) + val fetchSkippedLogResidentAcc: LongAccumulator = + spark.sparkContext.longAccumulator("vector_fetch_skipped_log_resident") + val fetchKeyMismatchAcc: LongAccumulator = spark.sparkContext.longAccumulator("vector_fetch_key_mismatch") + val rowsMaterializedAcc: LongAccumulator = spark.sparkContext.longAccumulator("vector_rows_materialized") + val exactRows: RDD[InternalRow] = byFetchBatch.mapPartitions { batches => + val startNs = System.nanoTime() + val rows = ArrayBuffer.empty[InternalRow] + var batchCount = 0 + var resolvedSlices = 0 + var missingSlices = 0 + var candidateKeys = 0L + var baseFiles = 0 + var logFiles = 0L + var totalSliceBytes = 0L + var staleCandidates = 0L + var fetchSkippedLogResident = 0L + var fetchKeyMismatch = 0L + var keyLookupRowsDecoded = 0L + var metrics = PositionalParquetFetcher.FetchMetrics() + val fetcher = new PositionalParquetFetcher( + storageConf.newInstance().unwrap(), + embeddingCol, + vectorSchema, + exactQueryVector, + ctx.metric, Review Comment: **Fixed** `3cde5283ea`. The closure captures only the distance metric. ########## hudi-spark-datasource/hudi-spark-common/src/main/scala/org/apache/spark/sql/hudi/analysis/VectorExactSearchPlanner.scala: ########## @@ -0,0 +1,390 @@ +/* + * 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.hudi.SparkAdapterSupport +import org.apache.hudi.client.common.HoodieSparkEngineContext +import org.apache.hudi.common.config.HoodieMetadataConfig +import org.apache.hudi.common.config.TypedProperties +import org.apache.hudi.common.index.vector.{ScoredVectorPostingMatch, VectorDistanceMetric, VectorIndexArbiter, VectorIndexMdtSearchUtils, VectorIndexMetadataCache, VectorIndexOptions, VectorPostingMatch, VectorStalePolicy} +import org.apache.hudi.common.model.{FileSlice, HoodieIndexDefinition, HoodieRecord} +import org.apache.hudi.common.schema.HoodieSchema +import org.apache.hudi.common.table.{HoodieTableMetaClient, TableSchemaResolver} +import org.apache.hudi.common.table.timeline.{HoodieTimeline, InstantComparison} +import org.apache.hudi.common.table.timeline.InstantComparison.compareTimestamps +import org.apache.hudi.common.table.view.HoodieTableFileSystemView +import org.apache.hudi.data.HoodieJavaRDD +import org.apache.hudi.hadoop.fs.HadoopFSUtils +import org.apache.hudi.metadata.{HoodieTableMetadata, HoodieTableMetadataUtil} + +import org.apache.spark.api.java.JavaSparkContext +import org.apache.spark.rdd.RDD +import org.apache.spark.sql.{AnalysisException, DataFrame, Row, SparkSession} +import org.apache.spark.sql.catalyst.InternalRow +import org.apache.spark.sql.catalyst.parser.ParseException +import org.apache.spark.sql.catalyst.plans.logical.HoodieVectorSearchTableValuedFunction +import org.apache.spark.sql.catalyst.plans.logical.HoodieVectorSearchTableValuedFunction.{DistanceMetric, SearchAlgorithm} +import org.apache.spark.sql.catalyst.plans.logical.LogicalPlan +import org.apache.spark.sql.expressions.Window +import org.apache.spark.sql.functions.{broadcast, col, monotonically_increasing_id, row_number} +import org.apache.spark.sql.hudi.command.exception.HoodieAnalysisException +import org.apache.spark.sql.types.{ArrayType, ByteType, DataType, DoubleType, FloatType, Metadata, StructField, StructType} +import org.apache.spark.util.LongAccumulator +import org.slf4j.LoggerFactory + +import scala.collection.JavaConverters._ +import scala.collection.mutable.ArrayBuffer +import scala.util.{Failure, Success, Try} + +/** Exact reranking and file-slice resolution for the IVF/RaBitQ search plan. */ +private[analysis] object VectorExactSearchPlanner extends SparkAdapterSupport { + import HoodieVectorSearchPlanBuilder._ + import IvfRaBitQMdtSearchAlgorithm.{isLogResidentCandidate, isPartitionedRecordIndex, toVectorMetric, VectorSearchPlanContext} + + private val LOG = LoggerFactory.getLogger(getClass) + + def buildRaBitQExactCandidateDf( + spark: SparkSession, + ctx: VectorSearchPlanContext, + outputSchema: StructType, + embeddingCol: String, + vectorSchema: HoodieSchema.Vector, + queryVector: Array[Float], + exactQueryVector: Array[Double], + k: Int): DataFrame = { + val planStartNs = System.nanoTime() + val storageConf = HadoopFSUtils.getStorageConf(spark.sessionState.newHadoopConf()) + val cacheLoadStartNs = System.nanoTime() + val cache = ctx.cache + val mergedOptions = ctx.tuningOptions + val metric = toVectorMetric(ctx.metric) + val residualEncoding = cache.isResidualEncoding + val numProbes = math.min(math.max(1, VectorIndexOptions.getNumProbes(mergedOptions.asJava)), cache.numClusters()) + val topClusters = cache.findTopClusters(queryVector, numProbes, metric) + val topClusterSeq = topClusters.toSeq + val cacheLoadMs = elapsedMs(cacheLoadStartNs) + + // Shard counts from cache — no extra MDT IO. + val shardCounts = cache.getShardCounts(topClusters) + val shardCountsScala = shardCounts.asScala.map { case (clusterId, count) => + clusterId.intValue() -> count.intValue() + }.toMap + val postingPrefixCount = VectorIndexMdtSearchUtils.buildPostingPrefixes(cache.getGenerationId, shardCounts).size() + LOG.info( + s"[vector_search][plan] basePath=${ctx.basePath} indexPartition=${ctx.indexPartition} cacheLoadMs=$cacheLoadMs " + + s"generationId=${cache.getGenerationId} dimension=${cache.getDimension} numClusters=${cache.numClusters()} " + + s"numProbes=$numProbes refineFactor=${VectorIndexOptions.getRefineFactor(mergedOptions.asJava)} residualEncoding=$residualEncoding " + + s"topClusters=${summarizeInts(topClusterSeq)} shardCounts=${summarizeShardCounts(shardCountsScala)} " + + s"postingPrefixes=$postingPrefixCount") + if (shardCounts.isEmpty) { + sparkAdapter.getUnsafeUtils.createDataFrameFromRDD( + spark, spark.sparkContext.emptyRDD[InternalRow], outputSchema) + } else { + val refineK = math.max(k, k * VectorIndexOptions.getRefineFactor(mergedOptions.asJava)) + val emptyDf = sparkAdapter.getUnsafeUtils.createDataFrameFromRDD( + spark, spark.sparkContext.emptyRDD[InternalRow], outputSchema) + + val topCandidates = VectorIndexMdtSearchUtils.scanPostingCandidates( + ctx.metadataTable, ctx.indexPartition, cache.getGenerationId, shardCounts, + queryVector, cache.getDimension, cache.getQuantizerSeed, cache.getRaBitQBits, cache.isAssumeNormalized, + metric, VectorIndexOptions.isRaBitQAsymmetricScoring(mergedOptions.asJava), + residualEncoding, if (residualEncoding) cache.getCentroids else null, refineK) + val topCandidateMaterializeStartNs = System.nanoTime() + val materializedCandidates = try { + topCandidates.collectAsList().asScala.toSeq + } finally { + topCandidates.unpersistWithDependencies() + } + val topCandidateList = materializedCandidates + .filter(candidate => candidate.getFileGroupId != null && candidate.getFileGroupId.nonEmpty) + val candidatesWithoutLocator = materializedCandidates.size - topCandidateList.size + LOG.info( + s"[vector_search][stage][materialize_top_candidates] refineK=$refineK rerankCandidates=${materializedCandidates.size} " + + s"candidatesWithLocators=${topCandidateList.size} candidatesWithoutLocators=$candidatesWithoutLocator " + + s"distinctFileGroups=${topCandidateList.map(_.getFileGroupId).distinct.size} elapsedMs=${elapsedMs(topCandidateMaterializeStartNs)}") + if (candidatesWithoutLocator > 0) { + throw new HoodieAnalysisException( + s"Vector exact search found $candidatesWithoutLocator rerank candidates without file-group locators. " + + s"Exact rerank requires per-candidate posting locators.") + } + + // RFC-109 RLI finalist arbiter. Exact mode uses refreshed RLI locations for every live + // candidate: SERVE retains positional trust; STALE is resolved by key in the live base file. + val arbiterEnabled = VectorIndexOptions.isFinalistArbiterEnabled(mergedOptions.asJava) + val serveCandidates: Seq[ScoredVectorPostingMatch] = + if (arbiterEnabled) { + val arb = VectorIndexMdtSearchUtils.arbitrateMaterializedFinalists( + ctx.metadataTable, topCandidateList.asJava, isPartitionedRecordIndex(ctx)) + val staleFailPolicy = VectorIndexOptions.isStaleLocatorPolicyFail(mergedOptions.asJava) + LOG.info( + s"[vector_search][exact][arbiter] arbiterExclusions{stale=${arb.staleCount}, deleted=${arb.deletedCount}} " + + s"serve=${arb.serve.size} stalePolicy=${if (staleFailPolicy) "fail" else "fallback"}") + if (staleFailPolicy && !arb.stale.isEmpty) { + throw new HoodieAnalysisException( + s"Vector exact search rejected ${arb.staleCount} stale finalist(s) because " + + "vector.stale.locator.policy=FAIL.") + } + arb.serve.asScala.toSeq ++ arb.stale.asScala.toSeq + } else { + topCandidateList + } + + if (serveCandidates.isEmpty) { + emptyDf + } else { + val candidatesByPartition = serveCandidates + .groupBy(candidate => Option(candidate.getLocation).map(_.getPartitionPath).getOrElse("")) + .map { case (partitionPath, candidates) => + partitionPath -> candidates.flatMap(candidate => Option(candidate.getLocation).map(_.getFileId)).toSet + } + val fileSliceMap = preResolveFileSlices( + ctx.metadataTable, ctx.metaClient, ctx.timeline, ctx.latestInstantTime, candidatesByPartition) + if (fileSliceMap.isEmpty) { + throw new HoodieAnalysisException( + s"Vector exact search resolved zero file slices for ${serveCandidates.size} rerank candidates " + + s"across ${candidatesByPartition.values.map(_.size).sum} file groups. Candidate locators are present, so this indicates " + + s"a file-slice resolution bug or stale vector index.") + } + val bFileSliceMap = spark.sparkContext.broadcast(fileSliceMap) + // MOR log-resident candidates carry their raw embedding in a log block, not the base Parquet + // file, so the positional fetcher cannot read them. Split them out and materialize them via a + // merged (base + log) read; base-resident candidates keep the fast positional path. + val (logResidentCandidates, baseCandidates) = serveCandidates.partition { candidate => + bFileSliceMap.value.get(candidate.getLocation.getFileId).exists { slice => + isLogResidentCandidate(candidate, slice.getBaseInstantTime, slice.getLogFiles.findAny().isPresent) + } + } + val exactFetchParallelism = math.max( + 1, + math.min(refineK, math.max(1, spark.sparkContext.defaultParallelism * 4))) + val groupedCandidates = baseCandidates.groupBy(_.getLocation.getFileId).toSeq.sortBy(_._1) + val targetBatchesPerGroup = math.max(1, exactFetchParallelism / math.max(1, groupedCandidates.size)) + val fetchBatches = groupedCandidates.flatMap { case (fgId, candidates) => + val sorted = candidates.sortBy(candidate => candidate.getLocation.getPosition) + val batchSize = math.max(1, math.ceil(sorted.size.toDouble / targetBatchesPerGroup).toInt) + sorted.grouped(batchSize).map(batch => (fgId, batch)).toSeq + } + val byFetchBatch: RDD[(String, Seq[ScoredVectorPostingMatch])] = + if (fetchBatches.isEmpty) { + spark.sparkContext.emptyRDD[(String, Seq[ScoredVectorPostingMatch])] + } else { + spark.sparkContext.parallelize(fetchBatches, math.min(exactFetchParallelism, fetchBatches.size)) + } + LOG.info( + s"[vector_search][stage][exact_read_plan] rerankCandidates=${serveCandidates.size} " + + s"candidateGroups=${groupedCandidates.size} resolvedSlices=${fileSliceMap.size} " + + s"fileGroupPartitions=${byFetchBatch.getNumPartitions} exactFetchParallelism=$exactFetchParallelism") + + val vectorOnlyProjection = mergedOptions.get(VECTOR_ONLY_PROJECTION_OPT) + .orElse(spark.conf.getOption(VECTOR_ONLY_PROJECTION_SPARK_CONF)) + .exists(_.equalsIgnoreCase("true")) + val fetchSkippedLogResidentAcc: LongAccumulator = + spark.sparkContext.longAccumulator("vector_fetch_skipped_log_resident") + val fetchKeyMismatchAcc: LongAccumulator = spark.sparkContext.longAccumulator("vector_fetch_key_mismatch") + val rowsMaterializedAcc: LongAccumulator = spark.sparkContext.longAccumulator("vector_rows_materialized") + val exactRows: RDD[InternalRow] = byFetchBatch.mapPartitions { batches => + val startNs = System.nanoTime() + val rows = ArrayBuffer.empty[InternalRow] + var batchCount = 0 + var resolvedSlices = 0 + var missingSlices = 0 + var candidateKeys = 0L + var baseFiles = 0 + var logFiles = 0L + var totalSliceBytes = 0L + var staleCandidates = 0L + var fetchSkippedLogResident = 0L + var fetchKeyMismatch = 0L + var keyLookupRowsDecoded = 0L + var metrics = PositionalParquetFetcher.FetchMetrics() + val fetcher = new PositionalParquetFetcher( + storageConf.newInstance().unwrap(), + embeddingCol, + vectorSchema, + exactQueryVector, + ctx.metric, + outputSchema, + mergedOptions.getOrElse("vector.exact_fetch.offset_index_missing_policy", "fallback").equalsIgnoreCase("fail"), + vectorOnlyProjection, + VectorIndexOptions.shouldVerifyFetchKeys(mergedOptions.asJava)) + val keyLocator = new ParquetRecordKeyLocator(storageConf.newInstance().unwrap()) + while (batches.hasNext) { + val (fgId, candidates) = batches.next() + batchCount += 1 + val fileSliceOpt = bFileSliceMap.value.get(fgId) + if (fileSliceOpt.isEmpty) { + missingSlices += 1 + throw new HoodieAnalysisException( + s"Vector exact search could not resolve file group '$fgId' for ${candidates.size} rerank candidates. " + + s"No record-level-index fallback is active in this path.") + } else { + resolvedSlices += 1 + val fileSlice = fileSliceOpt.get + val logFileCount = fileSlice.getLogFiles.count() + val currentBaseInstantTime = fileSlice.getBaseInstantTime + val logResident = candidates.filter(candidate => + isLogResidentCandidate(candidate, currentBaseInstantTime, logFileCount > 0)) + if (logResident.nonEmpty) { + fetchSkippedLogResident += logResident.size + fetchSkippedLogResidentAcc.add(logResident.size) + LOG.warn( + s"Vector exact fetch skipped ${logResident.size} log-resident candidate(s) in file group '$fgId'; " + + "they remain a result shortfall and the retained candidate pool will refill where candidates remain") + } + val liveCandidates = candidates.filterNot(logResident.toSet) + staleCandidates += liveCandidates.count(candidate => + candidate.getArbiterDecision == VectorIndexArbiter.Decision.STALE) + if (liveCandidates.nonEmpty) { + val positionalCandidates = liveCandidates.filter { candidate => + val location = candidate.getLocation + location != null && location.getPosition >= 0 && + candidate.getArbiterDecision != VectorIndexArbiter.Decision.STALE && + (location.getInstantTime == null || location.getInstantTime == currentBaseInstantTime) + } + val keyCandidates = liveCandidates.filterNot(positionalCandidates.toSet) + val recordKeys = liveCandidates.map(_.getRecordKey).toSet + candidateKeys += recordKeys.size + if (fileSlice.getBaseFile.isEmpty) { + throw new HoodieAnalysisException( + s"Vector exact key fallback requires a base Parquet file for file group '$fgId'; " + + "the remaining candidates are not classified as log-resident.") + } + baseFiles += 1 + logFiles += logFileCount + totalSliceBytes += fileSlice.getTotalFileSize + val baseFilePath = fileSlice.getBaseFile.get.getPath + val locatedByKey = keyLocator.locate(baseFilePath, keyCandidates) + keyLookupRowsDecoded += locatedByKey.metrics.rowsDecoded + if (locatedByKey.metrics.located != keyCandidates.size) { + throw new HoodieAnalysisException( + s"Vector exact key fallback located ${locatedByKey.metrics.located} of ${keyCandidates.size} " + + s"candidate key(s) in the live base file for file group '$fgId'.") + } + val firstFetch = fetcher.fetch(baseFilePath, positionalCandidates ++ locatedByKey.candidates) + rows ++= firstFetch.rows + metrics = metrics.add(firstFetch.metrics) + rowsMaterializedAcc.add(firstFetch.metrics.rowsMaterialized) + if (firstFetch.retryByKey.nonEmpty) { + fetchKeyMismatch += firstFetch.retryByKey.size + fetchKeyMismatchAcc.add(firstFetch.retryByKey.size) + val retries = keyLocator.locate(baseFilePath, firstFetch.retryByKey) + keyLookupRowsDecoded += retries.metrics.rowsDecoded + if (retries.metrics.located != firstFetch.retryByKey.size) { + throw new HoodieAnalysisException( + s"Vector exact fetch could not relocate ${firstFetch.retryByKey.size - retries.metrics.located} " + + s"key mismatch(es) in file group '$fgId'.") + } + val retryFetch = fetcher.fetch(baseFilePath, retries.candidates) + if (retryFetch.retryByKey.nonEmpty) { + throw new HoodieAnalysisException( + s"Vector exact fetch could not self-heal ${retryFetch.retryByKey.size} key mismatch(es) " + + s"in file group '$fgId' after key-based relocation.") + } + rows ++= retryFetch.rows + metrics = metrics.add(retryFetch.metrics) + rowsMaterializedAcc.add(retryFetch.metrics.rowsMaterialized) + } + } + } + } + LOG.info( + s"[vector_search][stage][exact_read] rerankCandidates=$candidateKeys candidateFileCount=$resolvedSlices " + + s"fetchBatches=$batchCount resolvedSlices=$resolvedSlices missingSlices=$missingSlices " + + s"rowsDecoded=${metrics.rowsDecoded} rowsMaterialized=${metrics.rowsMaterialized} " + + s"keyLookupRowsDecoded=$keyLookupRowsDecoded fetchKeyMismatch=$fetchKeyMismatch " + + s"fetchSkippedLogResident=$fetchSkippedLogResident " + + s"rowGroupsTotal=${metrics.rowGroupsTotal} rowGroupsSelected=${metrics.rowGroupsSelected} " + + s"pagesInSelectedRowGroups=${metrics.pagesInSelectedRowGroups} pagesSelected=${metrics.pagesSelected} " + + s"offsetIndexHits=${metrics.offsetIndexHits} offsetIndexMissing=${metrics.offsetIndexMissing} " + + s"rangedGets=${metrics.rangedGets} pageBytesFetched=${metrics.pageBytesFetched} " + + s"physicalBytesRead=${metrics.physicalBytesRead} footerReads=${metrics.footerReads} footerCacheHits=${metrics.footerCacheHits} " + + s"fetchKeyMismatchMetric=${metrics.fetchKeyMismatch} staleCandidates=$staleCandidates " + + s"baseFiles=$baseFiles logFiles=$logFiles " + + s"sliceBytes=$totalSliceBytes footerMs=${metrics.footerMs} offsetIndexMs=${metrics.offsetIndexMs} " + + s"fetchWaitMs=${metrics.fetchWaitMs} decodeMs=${metrics.decodeMs} scoreMs=${metrics.scoreMs} elapsedMs=${elapsedMs(startNs)}") + rows.iterator + } + val logResidentRows: Seq[InternalRow] = + if (logResidentCandidates.isEmpty) { + Seq.empty + } else { + val dataSchema = new TableSchemaResolver(ctx.metaClient).getTableSchema(true) + val scorer = new VectorExactScorer(vectorSchema, exactQueryVector, ctx.metric) + val logFetcher = new LogResidentVectorFetcher( + storageConf, ctx.metaClient, dataSchema, ctx.latestInstantTime, + embeddingCol, outputSchema, scorer) + val fetched = logResidentCandidates + .groupBy(candidate => + (Option(candidate.getLocation.getPartitionPath).getOrElse(""), candidate.getLocation.getFileId)) + .flatMap { case ((partitionPath, fgId), candidates) => + logFetcher.fetchSlice(partitionPath, fileSliceMap(fgId), candidates.map(_.getRecordKey).toSet) + }.toSeq + LOG.info( + s"[vector_search][exact][log_resident] candidates=${logResidentCandidates.size} " + + s"materialized=${fetched.size} " + + s"fileGroups=${logResidentCandidates.map(_.getLocation.getFileId).distinct.size}") + fetched + } + val combinedRows: RDD[InternalRow] = + if (logResidentRows.isEmpty) exactRows + else exactRows.union(spark.sparkContext.parallelize(logResidentRows)) + LOG.info(s"[vector_search][plan] exact candidate DF planned in ${elapsedMs(planStartNs)} ms") + sparkAdapter.getUnsafeUtils.createDataFrameFromRDD(spark, combinedRows, outputSchema) + } + } + } + + private def preResolveFileSlices( + metadataTable: HoodieTableMetadata, + metaClient: HoodieTableMetaClient, + timeline: HoodieTimeline, + latestInstantTime: String, + candidatesByPartition: Map[String, Set[String]]): Map[String, FileSlice] = { + val startNs = System.nanoTime() + val fsView = new HoodieTableFileSystemView(metadataTable, metaClient, timeline) + try { + val result = scala.collection.mutable.Map.empty[String, FileSlice] + var baseFiles = 0 + var logFiles = 0L + var totalSliceBytes = 0L + candidatesByPartition.foreach { case (partitionPath, candidateFgIds) => + val sliceIter = fsView.getLatestFileSlicesBeforeOrOn(partitionPath, latestInstantTime, true).iterator() + while (sliceIter.hasNext) { + val slice = sliceIter.next() + val fgId = slice.getFileGroupId.getFileId + if (candidateFgIds.contains(fgId)) { + result(fgId) = slice Review Comment: **Fixed**. Slices are keyed by `HoodieFileGroupId(partitionPath, fileId)` (`fileGroupIdOf`), so reused bucket file ids in different partitions don't collide. ########## hudi-spark-datasource/hudi-spark-common/src/main/scala/org/apache/spark/sql/hudi/analysis/PositionalParquetFetcher.scala: ########## @@ -0,0 +1,587 @@ +/* + * 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.hudi.common.index.vector.{ScoredVectorPostingMatch, VectorIndexMdtSearchUtils, VectorPostingMatch} +import org.apache.hudi.common.model.HoodieRecord +import org.apache.hudi.common.schema.HoodieSchema + +import org.apache.hadoop.conf.Configuration +import org.apache.hadoop.fs.{FileSystem, Path} +import org.apache.parquet.column.page.PageReadStore +import org.apache.parquet.hadoop.ParquetFileReader +import org.apache.parquet.hadoop.metadata.{BlockMetaData, ColumnChunkMetaData, ParquetMetadata} +import org.apache.parquet.internal.column.columnindex.OffsetIndex +import org.apache.parquet.internal.filter2.columnindex.RowRanges +import org.apache.parquet.io.{ColumnIOFactory, RecordReader} +import org.apache.parquet.io.api.{Binary, Converter, GroupConverter, PrimitiveConverter, RecordMaterializer} +import org.apache.parquet.schema.MessageType +import org.apache.spark.sql.catalyst.InternalRow +import org.apache.spark.sql.catalyst.expressions.GenericInternalRow +import org.apache.spark.sql.catalyst.plans.logical.HoodieVectorSearchTableValuedFunction.DistanceMetric +import org.apache.spark.sql.types._ +import org.apache.spark.unsafe.types.UTF8String +import org.slf4j.LoggerFactory + +import java.util.PrimitiveIterator +import java.util.concurrent.ConcurrentHashMap + +import scala.collection.JavaConverters._ +import scala.collection.mutable + +/** + * Fetches exact vector-search rerank candidates from Parquet by physical row position. + * + * The fetcher uses Parquet page indexes to turn candidate row positions into row ranges, so + * the reader can skip non-candidate pages instead of scanning whole Hudi file slices. + */ +private[analysis] final class PositionalParquetFetcher( + conf: Configuration, + embeddingCol: String, + vectorSchema: HoodieSchema.Vector, + queryVector: Array[Double], + metric: DistanceMetric.Value, + outputSchema: StructType, + failOnMissingOffsetIndex: Boolean, + vectorOnlyProjection: Boolean, + verifyKeys: Boolean) extends Serializable { + + import PositionalParquetFetcher._ + + private val distanceOrdinal = outputSchema.fieldIndex(HoodieVectorSearchPlanBuilder.DISTANCE_COL) + private val recordKeyOrdinal = fieldOrdinal(HoodieRecord.RECORD_KEY_METADATA_FIELD) + private val partitionPathOrdinal = fieldOrdinal(HoodieRecord.PARTITION_PATH_METADATA_FIELD) + private val fileGroupIdOrdinal = fieldOrdinal(HoodieVectorSearchPlanBuilder.FILE_GROUP_ID_COL) + private val scorer = new VectorExactScorer(vectorSchema, queryVector, metric) + private var scoreMs: Long = 0L + + def fetch( + filePath: String, + candidates: Seq[ScoredVectorPostingMatch]): FetchResult = { + val startNs = System.nanoTime() + if (candidates.isEmpty) { + FetchResult(Iterator.empty, Seq.empty, FetchMetrics(elapsedMs = elapsedMs(startNs))) + } else { + fetchNonEmpty(filePath, candidates, startNs) + } + } + + private def fetchNonEmpty( + filePath: String, + candidates: Seq[ScoredVectorPostingMatch], + startNs: Long): FetchResult = { + val path = new Path(filePath) + configureRandomRead(conf) + val fs = path.getFileSystem(conf) + val fileLen = fs.getFileStatus(path).getLen + val fileCacheKey = s"$filePath#$fileLen" + val stats = Option(FileSystem.getStatistics(path.toUri.getScheme, fs.getClass)) + val physicalBefore = stats.map(_.getBytesRead).getOrElse(0L) + + val footerStartNs = System.nanoTime() + val footerResult = footer(conf, path, fileCacheKey) + val footerMs = elapsedMs(footerStartNs) + val metadata = footerResult.metadata + val fileSchema = metadata.getFileMetaData.getSchema + val requestedParquetSchema = requestedSchema(fileSchema) + + val rowGroups = metadata.getBlocks.asScala.toIndexedSeq + val rowGroupStarts = cumulativeRowGroupStarts(rowGroups) + val byRowGroup: Map[Int, Seq[ScoredVectorPostingMatch]] = + candidates.groupBy(candidate => rowGroupOrdinal(rowGroupStarts, rowGroups, rowPosition(candidate))) + + val rows = mutable.ArrayBuffer.empty[InternalRow] + val keyMismatches = mutable.ArrayBuffer.empty[ScoredVectorPostingMatch] + var rowGroupsSelected = 0 + var pagesInSelectedRowGroups = 0L + var pagesSelected = 0L + var offsetIndexHits = 0L + var offsetIndexMissing = 0L + var pageBytesFetched = 0L + var rangedGets = 0L + var rowsDecoded = 0L + var rowsMaterialized = 0L + var offsetIndexMs = 0L + var decodeMs = 0L + scoreMs = 0L + + val reader = new ParquetFileReader(conf, path, metadata) + try { + reader.setRequestedSchema(requestedParquetSchema) + byRowGroup.toSeq.sortBy(_._1).foreach { case (rgOrdinal, rgCandidates) => + val block = rowGroups(rgOrdinal) + val rowGroupStart = rowGroupStarts(rgOrdinal) + val embeddingChunk = findColumn(block, embeddingCol) + .getOrElse(throw new IllegalArgumentException(s"Parquet file '$filePath' is missing vector column '$embeddingCol'.")) + + val offsetStartNs = System.nanoTime() + val embeddingOffsetIndex = offsetIndex(reader, fileCacheKey, rgOrdinal, embeddingChunk) + offsetIndexMs += elapsedMs(offsetStartNs) + val relativeRows = rgCandidates.map(candidate => rowPosition(candidate) - rowGroupStart).sorted + val candidateRows = relativeRows.toSet + val candidateByRelativeRow = rgCandidates + .map(candidate => rowPosition(candidate) - rowGroupStart -> candidate) + .toMap + + if (embeddingOffsetIndex == null && failOnMissingOffsetIndex) { + offsetIndexMissing += 1 + throw new IllegalStateException( + s"Parquet file '$filePath' has no OffsetIndex for vector column '$embeddingCol'; positional exact fetch cannot honor pages.") + } + + rowGroupsSelected += 1 + val (pageStore, rowRanges) = if (embeddingOffsetIndex == null) { + offsetIndexMissing += 1 + LOG.warn( + "Parquet file {} rowGroup={} has no OffsetIndex for vector column {}; falling back to row-group fetch.", + filePath, + Int.box(rgOrdinal), + embeddingCol) + pagesInSelectedRowGroups += 0L + pagesSelected += 0L + rangedGets += 1L + pageBytesFetched += requestedParquetSchema.getFields.asScala + .flatMap(field => findColumn(block, field.getName)) + .map(_.getTotalSize) + .sum + (reader.readRowGroup(rgOrdinal), RowRanges.createSingle(block.getRowCount)) + } else { + offsetIndexHits += 1 + pagesInSelectedRowGroups += embeddingOffsetIndex.getPageCount + val selectedVectorPages = selectedPages(embeddingOffsetIndex, block.getRowCount, relativeRows) + pagesSelected += selectedVectorPages.size + val selectedRanges = requestedParquetSchema.getFields.asScala + .flatMap(field => findColumn(block, field.getName)) + .flatMap { chunk => + val oiStartNs = System.nanoTime() + val chunkOffsetIndex = offsetIndex(reader, fileCacheKey, rgOrdinal, chunk) + offsetIndexMs += elapsedMs(oiStartNs) + if (chunkOffsetIndex == null) { + offsetIndexMissing += 1 + Seq.empty + } else { + offsetIndexHits += 1 + selectedPages(chunkOffsetIndex, block.getRowCount, relativeRows).map { pageOrdinal => + val start = chunkOffsetIndex.getOffset(pageOrdinal) + PageRange(start, start + chunkOffsetIndex.getCompressedPageSize(pageOrdinal)) + } + } + } + val mergedRanges = mergeRanges(selectedRanges.sortBy(_.start).toSeq) + rangedGets += mergedRanges.size + pageBytesFetched += mergedRanges.iterator.map(range => range.end - range.start).sum + val ranges = RowRanges.create(block.getRowCount, primitiveIntIterator(selectedVectorPages), embeddingOffsetIndex) + (reader.readFilteredRowGroup(rgOrdinal, ranges), ranges) + } + + val decodeStartNs = System.nanoTime() + val recordReader = parquetRecordReader(fileSchema, requestedParquetSchema, pageStore) + val rowIterator = rowRanges.iterator() + while (rowIterator.hasNext) { + val relativeRow = rowIterator.nextLong() + val row = recordReader.read() + rowsDecoded += 1 + candidateByRelativeRow.get(relativeRow).foreach { candidate => + if (verifyKeys && decodedRecordKey(row) != candidate.getRecordKey) { + keyMismatches += candidate + } else { + fillSyntheticFields(row, candidate) + rows += row + rowsMaterialized += 1 + } + } + } + decodeMs += elapsedMs(decodeStartNs) + } + } finally { + reader.close() + } + + val physicalAfter = stats.map(_.getBytesRead).getOrElse(physicalBefore) + val metrics = FetchMetrics( + rerankCandidates = candidates.size, + rowGroupsTotal = rowGroups.size, + rowGroupsSelected = rowGroupsSelected, + pagesInSelectedRowGroups = pagesInSelectedRowGroups, + pagesSelected = pagesSelected, + offsetIndexHits = offsetIndexHits, + offsetIndexMissing = offsetIndexMissing, + rangedGets = rangedGets, + pageBytesFetched = pageBytesFetched, + rowsDecoded = rowsDecoded, + rowsMaterialized = rowsMaterialized, + fetchKeyMismatch = keyMismatches.size, + physicalBytesRead = math.max(0L, physicalAfter - physicalBefore), + footerReads = if (footerResult.cacheHit) 0 else 1, + footerCacheHits = if (footerResult.cacheHit) 1 else 0, + footerMs = footerMs, + offsetIndexMs = offsetIndexMs, + fetchWaitMs = 0L, + decodeMs = decodeMs, + scoreMs = scoreMs, + elapsedMs = elapsedMs(startNs)) + FetchResult(rows.iterator, keyMismatches.toSeq, metrics) + } + + private def parquetRecordReader( + fileSchema: MessageType, + requestedParquetSchema: MessageType, + pageStore: PageReadStore): RecordReader[InternalRow] = { + val materializer = new InternalRowMaterializer(requestedParquetSchema, outputSchema, embeddingCol, distanceOrdinal) + new ColumnIOFactory().getColumnIO(requestedParquetSchema, fileSchema) + .getRecordReader(pageStore, materializer) + } + + private def requestedSchema(fileSchema: MessageType): MessageType = { + val parquetFields = + if (vectorOnlyProjection) { + mutable.ArrayBuffer.empty[org.apache.parquet.schema.Type] + } else { + outputSchema.fields + .filterNot(field => isSyntheticOutputField(field.name)) + .map(field => fieldByName(fileSchema, field.name)) + .toBuffer[org.apache.parquet.schema.Type] + } + if ((verifyKeys || vectorOnlyProjection) && !parquetFields.exists(_.getName == HoodieRecord.RECORD_KEY_METADATA_FIELD)) { + parquetFields += fieldByName(fileSchema, HoodieRecord.RECORD_KEY_METADATA_FIELD) + } + if (!parquetFields.exists(_.getName == embeddingCol)) { + parquetFields += fieldByName(fileSchema, embeddingCol) + } + new MessageType(fileSchema.getName, parquetFields.asJava) + } + + private def isSyntheticOutputField(name: String): Boolean = + name == HoodieVectorSearchPlanBuilder.DISTANCE_COL || + name == HoodieRecord.RECORD_KEY_METADATA_FIELD || + name == HoodieRecord.PARTITION_PATH_METADATA_FIELD || + name == HoodieVectorSearchPlanBuilder.FILE_GROUP_ID_COL + + private def fieldByName(schema: MessageType, name: String): org.apache.parquet.schema.Type = + schema.getFields.asScala.find(_.getName == name) + .getOrElse(throw new IllegalArgumentException(s"Parquet schema is missing requested field '$name'.")) + + private def rowPosition(candidate: ScoredVectorPostingMatch): Long = { + val location = candidate.getLocation + if (location == null) { + throw new IllegalArgumentException( + s"Vector exact positional fetch requires a row-position locator for record ${candidate.getRecordKey}.") + } + location.getPosition + } + + private def decodedRecordKey(row: InternalRow): String = + recordKeyOrdinal.flatMap(ordinal => Option(row.getUTF8String(ordinal))).map(_.toString).orNull + + private def fillSyntheticFields(row: InternalRow, candidate: ScoredVectorPostingMatch): Unit = { + val liveLocation = Option(candidate.getLocation) + recordKeyOrdinal.foreach(row.update(_, UTF8String.fromString(candidate.getRecordKey))) + partitionPathOrdinal.foreach(row.update(_, UTF8String.fromString( + liveLocation.map(_.getPartitionPath).orElse(Option(candidate.getPartitionPath)).getOrElse("")))) + fileGroupIdOrdinal.foreach(row.update(_, UTF8String.fromString( + liveLocation.map(_.getFileId).orElse(Option(candidate.getFileGroupId)).getOrElse("")))) + } + + private def fieldOrdinal(name: String): Option[Int] = + outputSchema.fields.indexWhere(_.name == name) match { + case -1 => None + case ordinal => Some(ordinal) + } + + private def scoreVector(bytes: Array[Byte]): Double = scorer.scoreBytes(bytes) + + private final class InternalRowMaterializer( + requestedParquetSchema: MessageType, + outputSchema: StructType, + embeddingCol: String, + distanceOrdinal: Int) extends RecordMaterializer[InternalRow] { + + private val values = new Array[Any](outputSchema.length) + private var current: InternalRow = _ + private var vectorBytes: Array[Byte] = _ + + private val root = new GroupConverter { + private val converters = requestedParquetSchema.getFields.asScala.map(converterFor).toIndexedSeq + + override def getConverter(fieldIndex: Int): Converter = + converters(fieldIndex) + + override def start(): Unit = { + var idx = 0 + while (idx < values.length) { + values(idx) = null + idx += 1 + } + vectorBytes = null + } + + override def end(): Unit = { + if (vectorBytes == null) { + throw new IllegalStateException(s"Vector column '$embeddingCol' was not decoded for positional exact fetch.") + } + val scoreStartNs = System.nanoTime() + values(distanceOrdinal) = scoreVector(vectorBytes) + scoreMs += elapsedMs(scoreStartNs) + current = new GenericInternalRow(values.clone()) + } + } + + override def getCurrentRecord: InternalRow = current + + override def getRootConverter: GroupConverter = root + + private def converterFor(field: org.apache.parquet.schema.Type): PrimitiveConverter = { + val name = field.getName + new PrimitiveConverter { + override def addBinary(value: Binary): Unit = { + if (name == embeddingCol) { + vectorBytes = value.getBytes + } else { + setOutputValue(name, value.getBytes) + } + } + + override def addInt(value: Int): Unit = setOutputValue(name, value) + + override def addLong(value: Long): Unit = setOutputValue(name, value) + + override def addFloat(value: Float): Unit = setOutputValue(name, value) + + override def addDouble(value: Double): Unit = setOutputValue(name, value) + + override def addBoolean(value: Boolean): Unit = setOutputValue(name, value) + } + } + + private def setOutputValue(name: String, raw: Any): Unit = { + if (name != embeddingCol && name != HoodieVectorSearchPlanBuilder.DISTANCE_COL) { + val ordinal = outputSchema.fieldIndex(name) + values(ordinal) = outputSchema.fields(ordinal).dataType match { + case StringType => raw match { + case bytes: Array[Byte] => UTF8String.fromBytes(bytes) + case other => UTF8String.fromString(String.valueOf(other)) + } + case BinaryType => raw.asInstanceOf[Array[Byte]] + case IntegerType => raw.asInstanceOf[Int] + case LongType => raw.asInstanceOf[Long] + case FloatType => raw.asInstanceOf[Float] + case DoubleType => raw.asInstanceOf[Double] + case BooleanType => raw.asInstanceOf[Boolean] + case ByteType => raw.asInstanceOf[Int].toByte + case ShortType => raw.asInstanceOf[Int].toShort + case other => throw new UnsupportedOperationException( + s"Unsupported exact positional fetch output type for field '$name': $other") + } + } + } + } +} + +private[analysis] object PositionalParquetFetcher { + private val LOG = LoggerFactory.getLogger(getClass) + private val FOOTER_CACHE = new ConcurrentHashMap[String, ParquetMetadata]() Review Comment: **Fixed** `3cde5283ea`. Footer/offset-index caches are bounded Caffeine caches. ########## hudi-spark-datasource/hudi-spark-common/src/main/scala/org/apache/spark/sql/hudi/analysis/PositionalParquetFetcher.scala: ########## @@ -0,0 +1,587 @@ +/* + * 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.hudi.common.index.vector.{ScoredVectorPostingMatch, VectorIndexMdtSearchUtils, VectorPostingMatch} +import org.apache.hudi.common.model.HoodieRecord +import org.apache.hudi.common.schema.HoodieSchema + +import org.apache.hadoop.conf.Configuration +import org.apache.hadoop.fs.{FileSystem, Path} +import org.apache.parquet.column.page.PageReadStore +import org.apache.parquet.hadoop.ParquetFileReader +import org.apache.parquet.hadoop.metadata.{BlockMetaData, ColumnChunkMetaData, ParquetMetadata} +import org.apache.parquet.internal.column.columnindex.OffsetIndex +import org.apache.parquet.internal.filter2.columnindex.RowRanges +import org.apache.parquet.io.{ColumnIOFactory, RecordReader} +import org.apache.parquet.io.api.{Binary, Converter, GroupConverter, PrimitiveConverter, RecordMaterializer} +import org.apache.parquet.schema.MessageType +import org.apache.spark.sql.catalyst.InternalRow +import org.apache.spark.sql.catalyst.expressions.GenericInternalRow +import org.apache.spark.sql.catalyst.plans.logical.HoodieVectorSearchTableValuedFunction.DistanceMetric +import org.apache.spark.sql.types._ +import org.apache.spark.unsafe.types.UTF8String +import org.slf4j.LoggerFactory + +import java.util.PrimitiveIterator +import java.util.concurrent.ConcurrentHashMap + +import scala.collection.JavaConverters._ +import scala.collection.mutable + +/** + * Fetches exact vector-search rerank candidates from Parquet by physical row position. + * + * The fetcher uses Parquet page indexes to turn candidate row positions into row ranges, so + * the reader can skip non-candidate pages instead of scanning whole Hudi file slices. + */ +private[analysis] final class PositionalParquetFetcher( + conf: Configuration, + embeddingCol: String, + vectorSchema: HoodieSchema.Vector, + queryVector: Array[Double], + metric: DistanceMetric.Value, + outputSchema: StructType, + failOnMissingOffsetIndex: Boolean, + vectorOnlyProjection: Boolean, + verifyKeys: Boolean) extends Serializable { + + import PositionalParquetFetcher._ + + private val distanceOrdinal = outputSchema.fieldIndex(HoodieVectorSearchPlanBuilder.DISTANCE_COL) + private val recordKeyOrdinal = fieldOrdinal(HoodieRecord.RECORD_KEY_METADATA_FIELD) + private val partitionPathOrdinal = fieldOrdinal(HoodieRecord.PARTITION_PATH_METADATA_FIELD) + private val fileGroupIdOrdinal = fieldOrdinal(HoodieVectorSearchPlanBuilder.FILE_GROUP_ID_COL) + private val scorer = new VectorExactScorer(vectorSchema, queryVector, metric) + private var scoreMs: Long = 0L + + def fetch( + filePath: String, + candidates: Seq[ScoredVectorPostingMatch]): FetchResult = { + val startNs = System.nanoTime() + if (candidates.isEmpty) { + FetchResult(Iterator.empty, Seq.empty, FetchMetrics(elapsedMs = elapsedMs(startNs))) + } else { + fetchNonEmpty(filePath, candidates, startNs) + } + } + + private def fetchNonEmpty( + filePath: String, + candidates: Seq[ScoredVectorPostingMatch], + startNs: Long): FetchResult = { + val path = new Path(filePath) + configureRandomRead(conf) + val fs = path.getFileSystem(conf) + val fileLen = fs.getFileStatus(path).getLen + val fileCacheKey = s"$filePath#$fileLen" + val stats = Option(FileSystem.getStatistics(path.toUri.getScheme, fs.getClass)) + val physicalBefore = stats.map(_.getBytesRead).getOrElse(0L) + + val footerStartNs = System.nanoTime() + val footerResult = footer(conf, path, fileCacheKey) + val footerMs = elapsedMs(footerStartNs) + val metadata = footerResult.metadata + val fileSchema = metadata.getFileMetaData.getSchema + val requestedParquetSchema = requestedSchema(fileSchema) + + val rowGroups = metadata.getBlocks.asScala.toIndexedSeq + val rowGroupStarts = cumulativeRowGroupStarts(rowGroups) + val byRowGroup: Map[Int, Seq[ScoredVectorPostingMatch]] = + candidates.groupBy(candidate => rowGroupOrdinal(rowGroupStarts, rowGroups, rowPosition(candidate))) + + val rows = mutable.ArrayBuffer.empty[InternalRow] + val keyMismatches = mutable.ArrayBuffer.empty[ScoredVectorPostingMatch] + var rowGroupsSelected = 0 + var pagesInSelectedRowGroups = 0L + var pagesSelected = 0L + var offsetIndexHits = 0L + var offsetIndexMissing = 0L + var pageBytesFetched = 0L + var rangedGets = 0L + var rowsDecoded = 0L + var rowsMaterialized = 0L + var offsetIndexMs = 0L + var decodeMs = 0L + scoreMs = 0L + + val reader = new ParquetFileReader(conf, path, metadata) + try { + reader.setRequestedSchema(requestedParquetSchema) + byRowGroup.toSeq.sortBy(_._1).foreach { case (rgOrdinal, rgCandidates) => + val block = rowGroups(rgOrdinal) + val rowGroupStart = rowGroupStarts(rgOrdinal) + val embeddingChunk = findColumn(block, embeddingCol) + .getOrElse(throw new IllegalArgumentException(s"Parquet file '$filePath' is missing vector column '$embeddingCol'.")) + + val offsetStartNs = System.nanoTime() + val embeddingOffsetIndex = offsetIndex(reader, fileCacheKey, rgOrdinal, embeddingChunk) + offsetIndexMs += elapsedMs(offsetStartNs) + val relativeRows = rgCandidates.map(candidate => rowPosition(candidate) - rowGroupStart).sorted + val candidateRows = relativeRows.toSet + val candidateByRelativeRow = rgCandidates + .map(candidate => rowPosition(candidate) - rowGroupStart -> candidate) + .toMap + + if (embeddingOffsetIndex == null && failOnMissingOffsetIndex) { + offsetIndexMissing += 1 + throw new IllegalStateException( + s"Parquet file '$filePath' has no OffsetIndex for vector column '$embeddingCol'; positional exact fetch cannot honor pages.") + } + + rowGroupsSelected += 1 + val (pageStore, rowRanges) = if (embeddingOffsetIndex == null) { + offsetIndexMissing += 1 + LOG.warn( + "Parquet file {} rowGroup={} has no OffsetIndex for vector column {}; falling back to row-group fetch.", + filePath, + Int.box(rgOrdinal), + embeddingCol) + pagesInSelectedRowGroups += 0L + pagesSelected += 0L + rangedGets += 1L + pageBytesFetched += requestedParquetSchema.getFields.asScala + .flatMap(field => findColumn(block, field.getName)) + .map(_.getTotalSize) + .sum + (reader.readRowGroup(rgOrdinal), RowRanges.createSingle(block.getRowCount)) + } else { + offsetIndexHits += 1 + pagesInSelectedRowGroups += embeddingOffsetIndex.getPageCount + val selectedVectorPages = selectedPages(embeddingOffsetIndex, block.getRowCount, relativeRows) + pagesSelected += selectedVectorPages.size + val selectedRanges = requestedParquetSchema.getFields.asScala + .flatMap(field => findColumn(block, field.getName)) + .flatMap { chunk => + val oiStartNs = System.nanoTime() + val chunkOffsetIndex = offsetIndex(reader, fileCacheKey, rgOrdinal, chunk) + offsetIndexMs += elapsedMs(oiStartNs) + if (chunkOffsetIndex == null) { + offsetIndexMissing += 1 + Seq.empty + } else { + offsetIndexHits += 1 + selectedPages(chunkOffsetIndex, block.getRowCount, relativeRows).map { pageOrdinal => + val start = chunkOffsetIndex.getOffset(pageOrdinal) + PageRange(start, start + chunkOffsetIndex.getCompressedPageSize(pageOrdinal)) + } + } + } + val mergedRanges = mergeRanges(selectedRanges.sortBy(_.start).toSeq) + rangedGets += mergedRanges.size + pageBytesFetched += mergedRanges.iterator.map(range => range.end - range.start).sum + val ranges = RowRanges.create(block.getRowCount, primitiveIntIterator(selectedVectorPages), embeddingOffsetIndex) + (reader.readFilteredRowGroup(rgOrdinal, ranges), ranges) + } + + val decodeStartNs = System.nanoTime() + val recordReader = parquetRecordReader(fileSchema, requestedParquetSchema, pageStore) + val rowIterator = rowRanges.iterator() + while (rowIterator.hasNext) { + val relativeRow = rowIterator.nextLong() + val row = recordReader.read() + rowsDecoded += 1 + candidateByRelativeRow.get(relativeRow).foreach { candidate => + if (verifyKeys && decodedRecordKey(row) != candidate.getRecordKey) { + keyMismatches += candidate + } else { + fillSyntheticFields(row, candidate) + rows += row + rowsMaterialized += 1 + } + } + } + decodeMs += elapsedMs(decodeStartNs) + } + } finally { + reader.close() + } + + val physicalAfter = stats.map(_.getBytesRead).getOrElse(physicalBefore) + val metrics = FetchMetrics( + rerankCandidates = candidates.size, + rowGroupsTotal = rowGroups.size, + rowGroupsSelected = rowGroupsSelected, + pagesInSelectedRowGroups = pagesInSelectedRowGroups, + pagesSelected = pagesSelected, + offsetIndexHits = offsetIndexHits, + offsetIndexMissing = offsetIndexMissing, + rangedGets = rangedGets, + pageBytesFetched = pageBytesFetched, + rowsDecoded = rowsDecoded, + rowsMaterialized = rowsMaterialized, + fetchKeyMismatch = keyMismatches.size, + physicalBytesRead = math.max(0L, physicalAfter - physicalBefore), + footerReads = if (footerResult.cacheHit) 0 else 1, + footerCacheHits = if (footerResult.cacheHit) 1 else 0, + footerMs = footerMs, + offsetIndexMs = offsetIndexMs, + fetchWaitMs = 0L, + decodeMs = decodeMs, + scoreMs = scoreMs, + elapsedMs = elapsedMs(startNs)) + FetchResult(rows.iterator, keyMismatches.toSeq, metrics) + } + + private def parquetRecordReader( + fileSchema: MessageType, + requestedParquetSchema: MessageType, + pageStore: PageReadStore): RecordReader[InternalRow] = { + val materializer = new InternalRowMaterializer(requestedParquetSchema, outputSchema, embeddingCol, distanceOrdinal) + new ColumnIOFactory().getColumnIO(requestedParquetSchema, fileSchema) + .getRecordReader(pageStore, materializer) + } + + private def requestedSchema(fileSchema: MessageType): MessageType = { + val parquetFields = + if (vectorOnlyProjection) { + mutable.ArrayBuffer.empty[org.apache.parquet.schema.Type] + } else { + outputSchema.fields + .filterNot(field => isSyntheticOutputField(field.name)) + .map(field => fieldByName(fileSchema, field.name)) + .toBuffer[org.apache.parquet.schema.Type] + } + if ((verifyKeys || vectorOnlyProjection) && !parquetFields.exists(_.getName == HoodieRecord.RECORD_KEY_METADATA_FIELD)) { + parquetFields += fieldByName(fileSchema, HoodieRecord.RECORD_KEY_METADATA_FIELD) + } + if (!parquetFields.exists(_.getName == embeddingCol)) { + parquetFields += fieldByName(fileSchema, embeddingCol) + } + new MessageType(fileSchema.getName, parquetFields.asJava) + } + + private def isSyntheticOutputField(name: String): Boolean = + name == HoodieVectorSearchPlanBuilder.DISTANCE_COL || + name == HoodieRecord.RECORD_KEY_METADATA_FIELD || + name == HoodieRecord.PARTITION_PATH_METADATA_FIELD || + name == HoodieVectorSearchPlanBuilder.FILE_GROUP_ID_COL + + private def fieldByName(schema: MessageType, name: String): org.apache.parquet.schema.Type = + schema.getFields.asScala.find(_.getName == name) + .getOrElse(throw new IllegalArgumentException(s"Parquet schema is missing requested field '$name'.")) + + private def rowPosition(candidate: ScoredVectorPostingMatch): Long = { + val location = candidate.getLocation + if (location == null) { + throw new IllegalArgumentException( + s"Vector exact positional fetch requires a row-position locator for record ${candidate.getRecordKey}.") + } + location.getPosition + } + + private def decodedRecordKey(row: InternalRow): String = + recordKeyOrdinal.flatMap(ordinal => Option(row.getUTF8String(ordinal))).map(_.toString).orNull + + private def fillSyntheticFields(row: InternalRow, candidate: ScoredVectorPostingMatch): Unit = { + val liveLocation = Option(candidate.getLocation) + recordKeyOrdinal.foreach(row.update(_, UTF8String.fromString(candidate.getRecordKey))) + partitionPathOrdinal.foreach(row.update(_, UTF8String.fromString( + liveLocation.map(_.getPartitionPath).orElse(Option(candidate.getPartitionPath)).getOrElse("")))) + fileGroupIdOrdinal.foreach(row.update(_, UTF8String.fromString( + liveLocation.map(_.getFileId).orElse(Option(candidate.getFileGroupId)).getOrElse("")))) + } + + private def fieldOrdinal(name: String): Option[Int] = + outputSchema.fields.indexWhere(_.name == name) match { + case -1 => None + case ordinal => Some(ordinal) + } + + private def scoreVector(bytes: Array[Byte]): Double = scorer.scoreBytes(bytes) + + private final class InternalRowMaterializer( + requestedParquetSchema: MessageType, + outputSchema: StructType, + embeddingCol: String, + distanceOrdinal: Int) extends RecordMaterializer[InternalRow] { + + private val values = new Array[Any](outputSchema.length) + private var current: InternalRow = _ + private var vectorBytes: Array[Byte] = _ + + private val root = new GroupConverter { + private val converters = requestedParquetSchema.getFields.asScala.map(converterFor).toIndexedSeq + + override def getConverter(fieldIndex: Int): Converter = + converters(fieldIndex) + + override def start(): Unit = { + var idx = 0 + while (idx < values.length) { + values(idx) = null + idx += 1 + } + vectorBytes = null + } + + override def end(): Unit = { + if (vectorBytes == null) { + throw new IllegalStateException(s"Vector column '$embeddingCol' was not decoded for positional exact fetch.") + } + val scoreStartNs = System.nanoTime() + values(distanceOrdinal) = scoreVector(vectorBytes) + scoreMs += elapsedMs(scoreStartNs) + current = new GenericInternalRow(values.clone()) + } + } + + override def getCurrentRecord: InternalRow = current + + override def getRootConverter: GroupConverter = root + + private def converterFor(field: org.apache.parquet.schema.Type): PrimitiveConverter = { + val name = field.getName + new PrimitiveConverter { + override def addBinary(value: Binary): Unit = { + if (name == embeddingCol) { + vectorBytes = value.getBytes + } else { + setOutputValue(name, value.getBytes) + } + } + + override def addInt(value: Int): Unit = setOutputValue(name, value) + + override def addLong(value: Long): Unit = setOutputValue(name, value) + + override def addFloat(value: Float): Unit = setOutputValue(name, value) + + override def addDouble(value: Double): Unit = setOutputValue(name, value) + + override def addBoolean(value: Boolean): Unit = setOutputValue(name, value) + } + } + + private def setOutputValue(name: String, raw: Any): Unit = { + if (name != embeddingCol && name != HoodieVectorSearchPlanBuilder.DISTANCE_COL) { + val ordinal = outputSchema.fieldIndex(name) + values(ordinal) = outputSchema.fields(ordinal).dataType match { + case StringType => raw match { + case bytes: Array[Byte] => UTF8String.fromBytes(bytes) + case other => UTF8String.fromString(String.valueOf(other)) + } + case BinaryType => raw.asInstanceOf[Array[Byte]] + case IntegerType => raw.asInstanceOf[Int] + case LongType => raw.asInstanceOf[Long] + case FloatType => raw.asInstanceOf[Float] + case DoubleType => raw.asInstanceOf[Double] + case BooleanType => raw.asInstanceOf[Boolean] + case ByteType => raw.asInstanceOf[Int].toByte + case ShortType => raw.asInstanceOf[Int].toShort + case other => throw new UnsupportedOperationException( Review Comment: **Fixed** `d685311d5f`. The positional fetcher now decodes through Spark's Parquet read support (`sparkAdapter.createParquetReadSupport`) with page-level row-range selection; the log path uses `sparkAdapter.createAvroDeserializer`. Timestamp, date, decimal, struct, array and map are covered for COW and MOR by `testExactSearchReturnsNonPrimitiveColumns`. -- 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]
