HyukjinKwon commented on code in PR #59943: URL: https://github.com/apache/spark/pull/59943#discussion_r4479845003
########## sql/core/src/main/scala/org/apache/spark/sql/execution/streaming/runtime/StreamingV2ScanPushDown.scala: ########## @@ -0,0 +1,475 @@ +/* + * 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.{And, Attribute, AttributeReference, AttributeSet, Cast, Expression, SubqueryExpression, Unevaluable} +import org.apache.spark.sql.catalyst.optimizer.{ComputeCurrentTime, InferFiltersFromConstraints, ReplaceExpressions, SpecialDatetimeValues} +import org.apache.spark.sql.catalyst.planning.PhysicalOperation +import org.apache.spark.sql.catalyst.plans.logical.{Filter, 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.catalyst.util.DateTimeUtils +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.{DataSourceStrategy, 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.{DataType, StringType, StructType, TimestampLTZNanosType, TimestampType} +import org.apache.spark.sql.util.CaseInsensitiveStringMap +import org.apache.spark.unsafe.types.UTF8String + +/** + * Pushes the top-level columns and the filters 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` or `STREAMING_FILTER_PUSHDOWN`. + * + * 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 + * gets what the planner takes from the optimized plan for the relation of the source (see + * `DataSourceV2Strategy`): the projections and the filters over each occurrence of the relation. + * + * The plan of each micro-batch is the one without these pushdowns, but for a filter of the rows + * of the relation with the filters that the scan builder got, which the query applies anyway, so + * that the scan cannot drop a row that the plan keeps (see `withPushedFilters`). The optimizer + * keeps deciding which columns every operator gets, and in which order, including the stateful + * ones, and Spark still evaluates all the filters over the relation. 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 + * @param filters the filters that the scan of the relation may get, in the order in which the + * plan evaluates them: those that every occurrence of the relation applies first, + * see `pushableFilters` + */ + case class Reads(requiredColumns: AttributeSet, filters: Seq[Expression]) + + /** + * 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 + * @param pushedFilters the filters that the scan builder got, which the scan may apply in any + * form, so the plan of each micro-batch applies them over the relation + */ + case class StreamingScan( + scan: Scan, + stream: SparkDataStream, + prunedOutput: Option[Seq[AttributeReference]], + pushedFilters: Seq[Expression]) { + + /** Returns the scan relation of the source whose relation is `relation`. */ + def toRelation( + relation: StreamingDataSourceV2Relation): StreamingDataSourceV2ScanRelation = { + StreamingDataSourceV2ScanRelation(relation, scan, relation.output, stream, + prunedOutput = prunedOutput, pushedFilters = pushedFilters) + } + } + + /** + * Returns what the optimized `plan` reads from each of its streaming relations of Data Source + * V2: what the planner takes from the projections and the filters over each occurrence of the + * relation, or all its columns and no filter 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 and gets no filter, 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, and the scan of a relation only gets what the + * query reads from that relation: + * - 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 with only a time to timestamps, 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 neither + * its filters nor its columns depend on them, and no filter on them is pushed down. + * - The filters that the optimizer infers from the constraints of the plan, such as those of the + * filters over another relation joined with the relation, are not inferred, so that the scan + * only gets the filters that the query applies to the relation itself. + * - 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) + conf.setConfString(SQLConf.OPTIMIZER_EXCLUDED_RULES.key, + (conf.optimizerExcludedRules.toSeq :+ InferFiltersFromConstraints.ruleName).mkString(",")) + 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)), + pushableFilters(filters)) + 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 with " + + log"no filter.") + Map.empty + } else { + combine(occurrences.toSeq) + } + } catch { + case NonFatal(e) => + logWarning(log"Could not optimize the streaming query to push down what it reads " + + log"to the scans of its sources, which read all their columns with no filter.", 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, and can only + * apply the filters that all of them apply. Each occurrence evaluates the pushed filters before + * its own, so the pushed filters are those that all the occurrences evaluate first, in the same + * order: the longest common prefix of their filters, see `pushableFilters`. + */ + private def combine( + occurrences: Seq[(StreamingRelationV2, Reads)]): Map[StreamingRelationV2, Reads] = { + occurrences.groupBy(_._1).map { case (relation, reads) => + val filters = reads.map(_._2.filters).reduce { (left, right) => + left.zip(right).takeWhile { case (l, r) => l.semanticEquals(r) }.map(_._1) + } + relation -> Reads(AttributeSet.fromAttributeSets(reads.map(_._2.requiredColumns)), filters) + } + } + + /** + * Returns the filters that the scan of a relation may get of `filters`, the filters over an + * occurrence of the relation in the order in which the plan evaluates them: the deterministic + * ones without subqueries or placeholders, up to the first filter that is not, which is not + * pushed down, and neither are those after it. The plan of each micro-batch evaluates the pushed + * filters first, over the relation (see `withPushedFilters`), so a pushed filter does not see a + * row that a filter written before it drops: it could fail on such a row, as a filter with a + * cast or a division does under ANSI, or with `raise_error`. + */ + private def pushableFilters(filters: Seq[Expression]): Seq[Expression] = { + filters.takeWhile { filter => + filter.deterministic && !SubqueryExpression.hasSubquery(filter) && + !filter.exists(_.isInstanceOf[StartupValue]) + } + } + + /** + * 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) + case relation: LogicalRelation if !relation.isStreaming => BatchRelation(relation.output) + case relation: HiveTableRelation => BatchRelation(relation.output) + // The rows of an inline table may have values of the current time, which the optimizer + // evaluates. + case table: ResolvedInlineTable => BatchRelation(table.output) + case node => + // The expressions that ReplaceExpressions replaces with others when the optimization starts + // are replaced first, as the optimizer does, so that the casts that they stand for, such as + // the one of to_date('today'), are replaced below. + val replaced = if (node.containsPattern(RUNTIME_REPLACEABLE)) { + node.mapExpressions(ReplaceExpressions.replace) + } else { + node + } + replaced.transformExpressionsUpWithPruning( + bits => bits.containsPattern(CURRENT_LIKE) || bits.containsPattern(CAST)) { + // The children of an expression are replaced before it, as are the subqueries, so an + // expression that still has the pattern CURRENT_LIKE has it itself: its value is the + // current time, or comes from the state of the session, as the current database does. + case e if e.containsPattern(CURRENT_LIKE) => StartupValue(e, e.children) + // So does a cast that ComputeCurrentTime or SpecialDatetimeValues rewrites: a cast of a + // time to a timestamp, which takes the current date, or a cast of a special string, such + // as 'today' or 'now', to a date or a timestamp. These rules only rewrite casts besides + // the expressions with the pattern CURRENT_LIKE. A cast of a string with only a time, + // such as '12:00:00', to a timestamp takes the current date too, which the optimizer + // folds in. + case cast: Cast if ComputeCurrentTime.applyForExpression(cast) != cast || + SpecialDatetimeValues.applyForExpression(cast) != cast || castsTimeOnly(cast) => + StartupValue(cast, cast.children) + } + } + + /** + * Returns whether `cast` casts a foldable string with only a time to a timestamp with local time + * zone, of any precision, which takes the date of when it is evaluated. A cast to a timestamp + * without time zone or to a date does not take the current date: it rejects such a string. + */ + private def castsTimeOnly(cast: Cast): Boolean = { + val toTimestampLTZ = cast.dataType match { + case TimestampType | _: TimestampLTZNanosType => true + case _ => false + } + toTimestampLTZ && cast.child.dataType.isInstanceOf[StringType] && cast.child.foldable && Review Comment: Done in 89114e3: `forStartup` now replaces every cast of a string to `TIMESTAMP` or `TIMESTAMP_LTZ(p)`, whatever its child is, so the cast of `cutoff` becomes a placeholder before the optimizer propagates the string. With the filter pushdown moved to SPARK-60926 (#60129), what these placeholders decide here is the columns of a scan. The new test "read the columns of a condition on a cast whose string is a literal only once optimized" checks your `withColumn("cutoff", lit("12:00:00"))` case in a `CASE WHEN`: the scan reads the columns of both branches. --- <!-- GITHUB_MCP_FOOTER: This attribution is automatically appended by GitHub MCP. --> _This comment was generated with [GitHub MCP](http://go/mcps)._ -- 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]
