dwsmith1983 commented on code in PR #5365: URL: https://github.com/apache/datafusion-comet/pull/5365#discussion_r3861279048
########## contrib/delta-spark/src/main/scala/org/apache/spark/sql/comet/CometDeltaNativeScanExec.scala: ########## @@ -0,0 +1,301 @@ +/* + * 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.comet + +import org.apache.spark.rdd.RDD +import org.apache.spark.sql.catalyst.expressions._ +import org.apache.spark.sql.catalyst.plans.QueryPlan +import org.apache.spark.sql.catalyst.plans.physical.{Partitioning, UnknownPartitioning} +import org.apache.spark.sql.execution.{FileSourceScanExec, InSubqueryExec, ReusedSubqueryExec, ScalarSubquery, SparkPlan, SubqueryAdaptiveBroadcastExec} +import org.apache.spark.sql.execution.datasources.HadoopFsRelation +import org.apache.spark.sql.execution.metric.SQLMetric +import org.apache.spark.sql.types.StructType +import org.apache.spark.sql.vectorized.ColumnarBatch + +import org.apache.comet.contrib.delta.DeltaSparkScanEnvelope +import org.apache.comet.serde.OperatorOuterClass +import org.apache.comet.serde.OperatorOuterClass.Operator + +/** + * Native scan node for Delta Lake tables (contrib). Delta's own planning (log replay, snapshot + * resolution, partition pruning) has already run inside delta-spark by the time this node is + * created from the DSv1 [[FileSourceScanExec]]; file listing and split planning are delegated to + * a [[CometScanExec]] helper, and data reads execute through Comet's native DataFusion parquet + * machinery, inheriting row-group and page-index pruning. + * + * DPP: `runtimeFilters` is a constructor field included in equality, so + * `CometPlanAdaptiveDynamicPruningFilters`'s rewrite (via [[CometScanWithPlanData]]) survives + * plan copies, the lesson from CometIcebergNativeScanExec (a transient field is dropped by + * `TreeNode.makeCopy` on MERGE re-planning). + */ +case class CometDeltaNativeScanExec( + override val nativeOp: Operator, + override val output: Seq[Attribute], + requiredSchema: StructType, + runtimeFilters: Seq[Expression], + dataFilters: Seq[Expression], + @transient relation: HadoopFsRelation, + originalPlan: FileSourceScanExec, + override val serializedPlanOpt: SerializedPlan, + sourceKey: String) + extends CometLeafExec + with CometScanWithPlanData { + + override val nodeName: String = s"CometDeltaNativeScan $relation" + + // Derived from (originalPlan, runtimeFilters), never stored: any copy of this node, our + // own withDynamicPruningFilters, or a generic Catalyst expression rewrite going through + // TreeNode.makeCopy, automatically gets a helper consistent with ITS runtimeFilters. A + // stored helper field would desync from rewritten filters (the #3510 class of bug). The + // cost is that file listing runs once per executed instance (planning listed separately in + // the rule extension); correctness over the duplicate driver-side listing. + // + // Forcing invariant: this lazy val is forced by the `metrics` override below (via + // `scanHelper.metrics`), and AQE's UI plan-walk calls `.metrics` on every node MID-PLANNING -- + // including while a DPP subquery is still an adaptive placeholder or a partition filter holds + // an unresolved ScalarSubquery (see `hasUnevaluableSubqueryFilter` just below). Constructing + // `scanHelper` here, and reading `CometScanExec.metrics` off it, is safe ONLY because that + // construction is a cheap case-class build with no file listing, and core's + // `CometScanExec.metrics` (spark/.../CometScanExec.scala) touches only `wrapped.driverMetrics` + // -- populated by Spark's own planning, not by this scan -- plus a static metric-node + // constructor. Neither does file listing (`selectedPartitions`/`getFilePartitions`) or + // subquery resolution. If core's `metrics` is ever changed to touch either, forcing + // `scanHelper` from this `metrics` override would resurrect the two AQE mid-planning crashes + // this invariant was written to prevent. + @transient private lazy val scanHelper: CometScanExec = + CometDeltaNativeScanExec.planningHelper(originalPlan, runtimeFilters) + + // NOT lazy val: while a DPP subquery is still an adaptive placeholder, or a partition filter + // holds an unresolved scalar subquery, this returns a temporary value that must not be + // memoized -- after CometPlanAdaptiveDynamicPruningFilters rewrites the filters (DPP case) or + // AQE resolves the subquery (scalar case), later reads must see the real post-pruning + // partition count. + override def outputPartitioning: Partitioning = + if (hasUnevaluableSubqueryFilter) UnknownPartitioning(0) + else UnknownPartitioning(perPartitionData.length) + + // runtimeFilters IS scanHelper.partitionFilters element-for-element (planningHelper passes + // partitionFilters = runtimeFilters into CometScanExec's plain constructor field below), so + // checking runtimeFilters here avoids constructing/forcing the derived scanHelper just to + // read partitioning. The InSubqueryExec placeholder shapes mirror + // CometPlanAdaptiveDynamicPruningFilters.extractSABData + hasWrappedSAB -- keep these two + // sets in sync; if that rule learns to unwrap a new wrapper form, mirror it here too. The + // ScalarSubquery case (e.g. `p = (SELECT max(p) FROM dim ...)`, a partition-column filter, + // so it lands in runtimeFilters rather than dataFilters) is presence-based, unlike the + // InSubqueryExec check above: Spark exposes no public finished/updated probe on + // ExecSubqueryExpression, so any partition filter containing one forces UnknownPartitioning(0) + // even once it has actually resolved. That is safe one-directionally only -- worst case an + // extra shuffle from an overly conservative partitioning, never a wrong answer, since + // execution reads perPartitionData directly and never goes through this getter. + private def hasUnevaluableSubqueryFilter: Boolean = + runtimeFilters.exists(_.exists { + // Match `e: InSubqueryExec` and dispatch on e.plan rather than unapplying InSubqueryExec + // directly: its unapply arity differs across Spark versions and this module ships no + // version shim. + case e: InSubqueryExec => isAdaptivePlaceholder(e.plan) + case _: ScalarSubquery => true Review Comment: Fixed. The guard now probes resolution instead of presence: eval on a resolved scalar subquery is a pure cached read (verified across the 3.5 through 4.2 jars), so once prepare resolves it the getter returns the real partition count, and only the unresolved case reports UnknownPartitioning(0). Your fused-parent shape reproduced the per-partition length failure locally and is now a regression that requires the native scan beneath a native aggregate. -- 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]
