zikangh commented on code in PR #59943: URL: https://github.com/apache/spark/pull/59943#discussion_r4485661217
########## sql/core/src/main/scala/org/apache/spark/sql/execution/streaming/runtime/StreamingV2ScanPushDown.scala: ########## @@ -0,0 +1,391 @@ +/* + * 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.streaming.runtime + +import scala.collection.mutable +import scala.util.{Failure, Success, Try} +import scala.util.control.NonFatal + +import org.apache.spark.internal.{Logging, LogKeys} +import org.apache.spark.sql.catalyst.analysis.{MultiInstanceRelation, ResolvedInlineTable} +import org.apache.spark.sql.catalyst.catalog.HiveTableRelation +import org.apache.spark.sql.catalyst.expressions.{Attribute, AttributeReference, AttributeSet, Cast, Expression, Unevaluable} +import org.apache.spark.sql.catalyst.optimizer.{ComputeCurrentTime, ReplaceExpressions, SpecialDatetimeValues} +import org.apache.spark.sql.catalyst.planning.PhysicalOperation +import org.apache.spark.sql.catalyst.plans.logical.{LeafNode, LogicalPlan, Statistics} +import org.apache.spark.sql.catalyst.streaming.StreamingRelationV2 +import org.apache.spark.sql.catalyst.trees.TreePattern.{CAST, CURRENT_LIKE, RUNTIME_REPLACEABLE} +import org.apache.spark.sql.catalyst.types.DataTypeUtils +import org.apache.spark.sql.classic.SparkSession +import org.apache.spark.sql.connector.catalog.{SupportsRead, TableCapability} +import org.apache.spark.sql.connector.read.{Scan, ScanBuilder, SupportsPushDownRequiredColumns} +import org.apache.spark.sql.connector.read.streaming.SparkDataStream +import org.apache.spark.sql.execution.datasources.LogicalRelation +import org.apache.spark.sql.execution.datasources.v2.{DataSourceV2Relation, PushDownUtils, StreamingDataSourceV2Relation, StreamingDataSourceV2ScanRelation} +import org.apache.spark.sql.execution.streaming.checkpointing.{OffsetSeqMetadata, OffsetSeqMetadataBase} +import org.apache.spark.sql.internal.SQLConf +import org.apache.spark.sql.types.{ArrayType, DataType, MapType, StringType, StructType, TimestampLTZNanosType, TimestampType} +import org.apache.spark.sql.util.CaseInsensitiveStringMap + +/** + * Pushes the top-level columns that the optimized plan of a streaming query reads from each Data + * Source V2 streaming source down to the scan of the source, as `V2ScanRelationPushDown` does for + * the scans of batch queries, if the table of the source has the capability + * `STREAMING_COLUMN_PRUNING`. + * + * The stream of a source is created from its scan, and has to exist before any micro-batch is + * planned, so the scan is built when the query starts. The query is optimized once at that point, + * by the optimizer of its micro-batches and with their configuration, and the scan of each source + * reads the columns that the planner takes from the optimized plan for the relation of the source + * (see `DataSourceV2Strategy`): those of the projections and the filters over each occurrence of + * the relation. + * + * The plan of each micro-batch is the one without this pushdown: the optimizer keeps deciding + * which columns every operator gets, and in which order, including the stateful ones. Only the + * scan node of the physical plan outputs the columns that the scan reads, and the planner checks + * that the micro-batch reads no other column of the relation. + * + * The other operations that the scans of batch queries can apply, such as aggregates and limits, + * apply to the data of all the micro-batches of a streaming query, which the scan of a micro-batch + * does not read, so they are not pushed down. + */ +private[sql] object StreamingV2ScanPushDown extends Logging { + + /** + * What the optimized plan of a streaming query reads from a streaming relation. + * + * @param requiredColumns the columns of the relation that the plan reads + */ + case class Reads(requiredColumns: AttributeSet) + + /** + * The scan of a streaming source, and its stream. + * + * @param prunedOutput the columns of the relation that the scan reads, in the order of its read + * schema, unless these are all its columns + */ + case class StreamingScan( + scan: Scan, + stream: SparkDataStream, + prunedOutput: Option[Seq[AttributeReference]]) { + + /** Returns the scan relation of the source whose relation is `relation`. */ + def toRelation( + relation: StreamingDataSourceV2Relation): StreamingDataSourceV2ScanRelation = { + StreamingDataSourceV2ScanRelation(relation, scan, relation.output, stream, + prunedOutput = prunedOutput) + } + } + + /** + * Returns what the optimized `plan` reads from each of its streaming relations of Data Source + * V2: the columns that the planner takes from the projections and the filters over each + * occurrence of the relation, or all its columns where nothing is projected or filtered over it. + * A relation that the optimized plan does not have has no entry, so that its scan reads all its + * columns, as no relation has if the optimizer fails, or if the optimized plan has a streaming + * relation that `plan` does not have. + * + * The plan is optimized as the micro-batches are: with the optimizer and the configuration of + * the stream, the values of the configurations that the offset log of a restarted query pins, and + * without the cost-based reordering of joins, which `StreamExecution` disables for its queries. + * The plan does not depend on when it is optimized: + * - The expressions whose values depend on when they are evaluated or on the state of the + * session, such as the current time, the casts of special strings such as 'today' or 'now' + * to dates and timestamps, the casts of strings to timestamps with local time zone, which take + * the current date for a string with only a time, or the current database, which each + * micro-batch gets the values of, become placeholders that the optimizer does not take as + * deterministic, as the current time of a micro-batch, so the columns that the plan reads do + * not depend on them. + * - The relations of batch scans become placeholders, so that their scans are neither built + * nor planned: they do not change what the plan reads from its streaming relations. + * + * @param session the session of the stream + * @param offsetLogMetadata the metadata of the last entry of the offset log of the query, if any + */ + def reads( + session: SparkSession, + plan: LogicalPlan, + offsetLogMetadata: Option[OffsetSeqMetadataBase]): Map[StreamingRelationV2, Reads] = { + val conf = session.sessionState.conf.clone() + offsetLogMetadata.foreach(OffsetSeqMetadata.setSessionConf(_, conf, logChanges = false)) + conf.setConf(SQLConf.CBO_ENABLED, false) + session.withActive { + SQLConf.withExistingConf(conf) { + try { + val optimized = session.sessionState.optimizer.execute(forStartup(plan)) + // The planner of a micro-batch gets the values of the expressions that depend on when + // they are evaluated, which are deterministic. + val planned = optimized.transformUpWithSubqueries { + case node => node.transformExpressions { + case value: StartupValue => value.copy(planned = true) + } + } + val occurrences = mutable.ArrayBuffer.empty[(StreamingRelationV2, Reads)] + // Visits the nodes as the planner plans them: the projections and the filters over a + // relation are planned with its scan, and the other nodes on their own. + def visit(node: LogicalPlan): Unit = node match { + case PhysicalOperation(project, filters, relation: StreamingRelationV2) => + occurrences += relation -> Reads( + AttributeSet((project ++ filters).flatMap(_.references))) + case _ => + node.children.foreach(visit) + } + visit(planned) + val relations = plan.collect { case relation: StreamingRelationV2 => relation }.toSet + val unknown = occurrences.map(_._1).filterNot(relations.contains) + if (unknown.nonEmpty) { + // The optimizer does not copy streaming relations: `DeduplicateRelations` keeps + // streaming plans as they are, so all the references to a common table expression, + // for example, have the same relation. A copy would share the scan of its relation in + // the micro-batches, so the scan would have to read what the copy reads too. + logWarning(log"The optimized streaming query reads the streaming relations " + + log"${MDC(LogKeys.STREAMING_TABLE, unknown.distinct.mkString(", "))} that its " + + log"plan does not have, so the scans of its sources read all their columns.") + Map.empty + } else { + combine(occurrences.toSeq) + } + } catch { + case NonFatal(e) => + logWarning(log"Could not optimize the streaming query to push down the columns " + + log"that it reads to the scans of its sources, which read all their columns.", e) + Map.empty + } + } + } + } + + /** + * Returns what the occurrences of each relation read, combined: the scan of a relation reads the + * rows of all its occurrences, so it reads the columns that any of them reads. + */ + private def combine( + occurrences: Seq[(StreamingRelationV2, Reads)]): Map[StreamingRelationV2, Reads] = { + occurrences.groupBy(_._1).map { case (relation, reads) => + relation -> Reads(AttributeSet.fromAttributeSets(reads.map(_._2.requiredColumns))) + } + } + + /** + * Returns `plan` for the optimization when the query starts, see `reads`, with its subqueries. + * The expressions that are replaced keep their children, so that the plan still reads the same + * columns. Visible for tests. + */ + private[sql] def forStartup(plan: LogicalPlan): LogicalPlan = plan.transformUpWithSubqueries { + case relation: DataSourceV2Relation => BatchRelation(relation.output) Review Comment: reads doesn't apply the cache, so a cached static side is transparent here but an opaque InMemoryRelation in each micro-batch. e.g. joining with `dim.withColumn("on", lit(false)).cache()` and selecting `when($"on", $"data").otherwise($"extra")` prunes data at startup, and every micro-batch then fails the column check. -- 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]
