Akash3121 commented on code in PR #11174: URL: https://github.com/apache/paimon/pull/11174#discussion_r4581526537
########## paimon-spark/paimon-spark-common/src/main/scala/org/apache/paimon/spark/commands/ReplaceByKeysCommand.scala: ########## @@ -0,0 +1,329 @@ +/* + * 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.paimon.spark.commands + +import org.apache.paimon.{CoreOptions, Snapshot} +import org.apache.paimon.spark.{PaimonScanBuilder, SparkTable} +import org.apache.paimon.spark.catalyst.analysis.PaimonOutputResolver +import org.apache.paimon.table.{FileStoreTable, KnownSplitsTable} +import org.apache.paimon.table.sink.{CommitMessage, CommitMessageImpl} +import org.apache.paimon.table.source.{DataSplit, Split} +import org.apache.paimon.table.source.snapshot.TimeTravelUtil + +import org.apache.spark.sql.{DataFrame, Row, SparkSession} +import org.apache.spark.sql.PaimonUtils.createDataset +import org.apache.spark.sql.catalyst.expressions.{Attribute, SubqueryExpression} +import org.apache.spark.sql.catalyst.expressions.Literal.TrueLiteral +import org.apache.spark.sql.catalyst.optimizer.ComputeCurrentTime +import org.apache.spark.sql.catalyst.plans.logical.{Aggregate, Command, CTERelationRef, Deduplicate, GlobalLimit, LocalLimit, LocalRelation, LogicalPlan, OneRowRelation, ParsedStatement, Range, Sample, Tail, Window} +import org.apache.spark.sql.connector.read.ScanBuilder +import org.apache.spark.sql.execution.datasources.v2.DataSourceV2Relation +import org.apache.spark.sql.functions.{col, expr} +import org.apache.spark.sql.paimon.shims.SparkShimLoader +import org.apache.spark.sql.util.CaseInsensitiveStringMap + +import java.util.{Collections, TreeMap} + +import scala.collection.JavaConverters._ +import scala.collection.mutable + +/** Prepare a key-based delete and append, publishing both in one data commit. */ +case class ReplaceByKeysCommand( + override val table: FileStoreTable, + relation: DataSourceV2Relation, + sourceQuery: String, + keyColumns: Seq[String], + targetFilter: Option[String]) + extends PaimonRowLevelCommand { + + override def run(spark: SparkSession): Seq[Row] = { + require(coreOptions.dataEvolutionEnabled(), "replace_by_keys requires a Data Evolution table") + require( + coreOptions.deletionVectorsEnabled(), + "replace_by_keys requires deletion-vectors.enabled=true") + val (messages, readSnapshotId) = prepare(spark) + val checked = + if (readSnapshotId < 0) messages + else messages.map(_.asInstanceOf[CommitMessageImpl].withCheckFromSnapshot(readSnapshotId)) Review Comment: `withCheckFromSnapshot(readSnapshotId)` is currently a no-op for the `CommitMessage`s produced by `ReplaceByKeysCommand`: - `FileStoreCommitImpl.configureRowIdCheckFromMessages` enables `DataEvolutionDmlRowIdConflictCheck`, which builds `RowIdColumnConflictChecker.fromDataFiles(schemaManager, deltaFiles, nestedFieldEnabled)` by filtering `deltaFiles.stream().filter(file -> file.firstRowId() != null)`. - Here, the deletion messages only contain `indexIncrement` (DV index files, no `deltaFiles`), and the appended messages only contain newly written files whose `firstRowId` is `null` until `assignRowTracking` runs at commit time. As a result, `rowIdConflictChecker.isEmpty()` is always `true` and `checkForRowIdFromSnapshot` immediately returns `Optional.empty()` without inspecting snapshots `(readSnapshotId, latestSnapshot]`. While concurrent DV changes on the same anchor file are still caught by `checkDeleteInEntries`, a concurrent Data Evolution partial-column update committing a new column-slice file (`firstRowId != null`) on the matched row range after `readSnapshotId` will not be detected. ########## paimon-spark/paimon-spark-common/src/main/scala/org/apache/paimon/spark/commands/ReplaceByKeysCommand.scala: ########## @@ -0,0 +1,329 @@ +/* + * 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.paimon.spark.commands + +import org.apache.paimon.{CoreOptions, Snapshot} +import org.apache.paimon.spark.{PaimonScanBuilder, SparkTable} +import org.apache.paimon.spark.catalyst.analysis.PaimonOutputResolver +import org.apache.paimon.table.{FileStoreTable, KnownSplitsTable} +import org.apache.paimon.table.sink.{CommitMessage, CommitMessageImpl} +import org.apache.paimon.table.source.{DataSplit, Split} +import org.apache.paimon.table.source.snapshot.TimeTravelUtil + +import org.apache.spark.sql.{DataFrame, Row, SparkSession} +import org.apache.spark.sql.PaimonUtils.createDataset +import org.apache.spark.sql.catalyst.expressions.{Attribute, SubqueryExpression} +import org.apache.spark.sql.catalyst.expressions.Literal.TrueLiteral +import org.apache.spark.sql.catalyst.optimizer.ComputeCurrentTime +import org.apache.spark.sql.catalyst.plans.logical.{Aggregate, Command, CTERelationRef, Deduplicate, GlobalLimit, LocalLimit, LocalRelation, LogicalPlan, OneRowRelation, ParsedStatement, Range, Sample, Tail, Window} +import org.apache.spark.sql.connector.read.ScanBuilder +import org.apache.spark.sql.execution.datasources.v2.DataSourceV2Relation +import org.apache.spark.sql.functions.{col, expr} +import org.apache.spark.sql.paimon.shims.SparkShimLoader +import org.apache.spark.sql.util.CaseInsensitiveStringMap + +import java.util.{Collections, TreeMap} + +import scala.collection.JavaConverters._ +import scala.collection.mutable + +/** Prepare a key-based delete and append, publishing both in one data commit. */ +case class ReplaceByKeysCommand( + override val table: FileStoreTable, + relation: DataSourceV2Relation, + sourceQuery: String, + keyColumns: Seq[String], + targetFilter: Option[String]) + extends PaimonRowLevelCommand { + + override def run(spark: SparkSession): Seq[Row] = { + require(coreOptions.dataEvolutionEnabled(), "replace_by_keys requires a Data Evolution table") + require( + coreOptions.deletionVectorsEnabled(), + "replace_by_keys requires deletion-vectors.enabled=true") + val (messages, readSnapshotId) = prepare(spark) + val checked = + if (readSnapshotId < 0) messages + else messages.map(_.asInstanceOf[CommitMessageImpl].withCheckFromSnapshot(readSnapshotId)) + // Once publication is attempted an exception can have an unknown outcome. Do not abort it. + writer.commit(checked, Snapshot.Operation.MERGE) + Seq.empty + } + + private[spark] def prepare(spark: SparkSession): (Seq[CommitMessage], Long) = { + val snapshot = table.snapshotManager().latestSnapshot() + def byName(attributes: Seq[Attribute]): TreeMap[String, Attribute] = { + val result = + if (spark.sessionState.conf.caseSensitiveAnalysis) new TreeMap[String, Attribute]() + else new TreeMap[String, Attribute](String.CASE_INSENSITIVE_ORDER) + attributes.foreach { + attribute => + require(!result.containsKey(attribute.name), s"Ambiguous column name: ${attribute.name}") + result.put(attribute.name, attribute) + } + result + } + val fields = byName(relation.output) + require(keyColumns.nonEmpty, "key_columns must not be empty") + val keys = keyColumns.map { + key => + require(fields.containsKey(key), s"Unknown target key column: $key") + fields.get(key).name + } + require(keys.distinct.size == keyColumns.size, "Duplicate key columns") + val source = stableSource(spark, snapshot) + val sourceFields = byName(source.output) + require( + sourceFields.keySet() == fields.keySet(), + "Source query must supply all target columns by name") + // Reuse INSERT's assignment policy, nested/nullability checks and CHAR/VARCHAR handling. + val aligned = PaimonOutputResolver.resolveOutputColumns( + table.name(), + relation.output, + source, + byName = true, + mergeSchemaEnabled = false) + val input = createDataset(spark, aligned) + // Match original query keys: an INSERT cast must not change which old rows are deleted. + val sourceKeys = createDataset(spark, source) + .select(keys.map(k => col(quoted(sourceFields.get(k).name)).as(k)): _*) + .alias("source_keys") + + var staged = Seq.empty[CommitMessage] + try { + staged = prepareDeletion(spark, snapshot, sourceKeys, keys) + staged = staged ++ writer.write(input) + (staged, if (snapshot == null) -1L else snapshot.id()) + } catch { + case failure: Throwable => + if (staged.nonEmpty) { + try { + val cleanup = writer.writeBuilder.newCommit() + try cleanup.abort(staged.asJava) + finally cleanup.close() + } catch { + case cleanupFailure: Throwable => + if (cleanupFailure ne failure) failure.addSuppressed(cleanupFailure) + } + } + throw failure + } + } + + // Release the target scan and full file inventory before starting the append job. + private def prepareDeletion( + spark: SparkSession, + snapshot: Snapshot, + sourceKeys: DataFrame, + keys: Seq[String]): Seq[CommitMessage] = { + val reader = table.newSnapshotReader() + if (snapshot != null) reader.withSnapshot(snapshot) + val predicate = targetFilter + .map { + text => + val parsed = spark.sessionState.sqlParser.parseExpression(text) Review Comment: When the target table is initially empty at the start of `prepare`, `snapshot` is `null` (line 70). Because `if (snapshot != null) reader.withSnapshot(snapshot)` is skipped here, `reader` remains unpinned and `reader.read()` on line 161 will read `latestSnapshot()` from disk. If a concurrent writer commits Snapshot 1 between line 70 and line 161, `prepareDeletion` will scan Snapshot 1's files and call `writer.persistDeletionVectors(vectors, null)` while `prepare` still returns `readSnapshotId = -1L` (skipping conflict checks from a baseline snapshot). After validating `targetFilter` (lines 137–153), if `snapshot == null`, `prepareDeletion` should return `Seq.empty` immediately without calling `reader.read()`. ########## paimon-spark/paimon-spark-common/src/main/scala/org/apache/paimon/spark/commands/ReplaceByKeysCommand.scala: ########## @@ -0,0 +1,329 @@ +/* + * 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.paimon.spark.commands + +import org.apache.paimon.{CoreOptions, Snapshot} +import org.apache.paimon.spark.{PaimonScanBuilder, SparkTable} +import org.apache.paimon.spark.catalyst.analysis.PaimonOutputResolver +import org.apache.paimon.table.{FileStoreTable, KnownSplitsTable} +import org.apache.paimon.table.sink.{CommitMessage, CommitMessageImpl} +import org.apache.paimon.table.source.{DataSplit, Split} +import org.apache.paimon.table.source.snapshot.TimeTravelUtil + +import org.apache.spark.sql.{DataFrame, Row, SparkSession} +import org.apache.spark.sql.PaimonUtils.createDataset +import org.apache.spark.sql.catalyst.expressions.{Attribute, SubqueryExpression} +import org.apache.spark.sql.catalyst.expressions.Literal.TrueLiteral +import org.apache.spark.sql.catalyst.optimizer.ComputeCurrentTime +import org.apache.spark.sql.catalyst.plans.logical.{Aggregate, Command, CTERelationRef, Deduplicate, GlobalLimit, LocalLimit, LocalRelation, LogicalPlan, OneRowRelation, ParsedStatement, Range, Sample, Tail, Window} +import org.apache.spark.sql.connector.read.ScanBuilder +import org.apache.spark.sql.execution.datasources.v2.DataSourceV2Relation +import org.apache.spark.sql.functions.{col, expr} +import org.apache.spark.sql.paimon.shims.SparkShimLoader +import org.apache.spark.sql.util.CaseInsensitiveStringMap + +import java.util.{Collections, TreeMap} + +import scala.collection.JavaConverters._ +import scala.collection.mutable + +/** Prepare a key-based delete and append, publishing both in one data commit. */ +case class ReplaceByKeysCommand( + override val table: FileStoreTable, + relation: DataSourceV2Relation, + sourceQuery: String, + keyColumns: Seq[String], + targetFilter: Option[String]) + extends PaimonRowLevelCommand { + + override def run(spark: SparkSession): Seq[Row] = { + require(coreOptions.dataEvolutionEnabled(), "replace_by_keys requires a Data Evolution table") + require( + coreOptions.deletionVectorsEnabled(), + "replace_by_keys requires deletion-vectors.enabled=true") + val (messages, readSnapshotId) = prepare(spark) + val checked = + if (readSnapshotId < 0) messages + else messages.map(_.asInstanceOf[CommitMessageImpl].withCheckFromSnapshot(readSnapshotId)) + // Once publication is attempted an exception can have an unknown outcome. Do not abort it. + writer.commit(checked, Snapshot.Operation.MERGE) + Seq.empty + } + + private[spark] def prepare(spark: SparkSession): (Seq[CommitMessage], Long) = { + val snapshot = table.snapshotManager().latestSnapshot() + def byName(attributes: Seq[Attribute]): TreeMap[String, Attribute] = { + val result = + if (spark.sessionState.conf.caseSensitiveAnalysis) new TreeMap[String, Attribute]() + else new TreeMap[String, Attribute](String.CASE_INSENSITIVE_ORDER) + attributes.foreach { + attribute => + require(!result.containsKey(attribute.name), s"Ambiguous column name: ${attribute.name}") + result.put(attribute.name, attribute) + } + result + } + val fields = byName(relation.output) + require(keyColumns.nonEmpty, "key_columns must not be empty") + val keys = keyColumns.map { + key => + require(fields.containsKey(key), s"Unknown target key column: $key") + fields.get(key).name + } + require(keys.distinct.size == keyColumns.size, "Duplicate key columns") + val source = stableSource(spark, snapshot) + val sourceFields = byName(source.output) + require( + sourceFields.keySet() == fields.keySet(), + "Source query must supply all target columns by name") + // Reuse INSERT's assignment policy, nested/nullability checks and CHAR/VARCHAR handling. + val aligned = PaimonOutputResolver.resolveOutputColumns( + table.name(), + relation.output, + source, + byName = true, + mergeSchemaEnabled = false) + val input = createDataset(spark, aligned) + // Match original query keys: an INSERT cast must not change which old rows are deleted. + val sourceKeys = createDataset(spark, source) + .select(keys.map(k => col(quoted(sourceFields.get(k).name)).as(k)): _*) + .alias("source_keys") + + var staged = Seq.empty[CommitMessage] + try { + staged = prepareDeletion(spark, snapshot, sourceKeys, keys) + staged = staged ++ writer.write(input) + (staged, if (snapshot == null) -1L else snapshot.id()) + } catch { + case failure: Throwable => + if (staged.nonEmpty) { + try { + val cleanup = writer.writeBuilder.newCommit() + try cleanup.abort(staged.asJava) + finally cleanup.close() + } catch { + case cleanupFailure: Throwable => + if (cleanupFailure ne failure) failure.addSuppressed(cleanupFailure) + } + } + throw failure + } + } + + // Release the target scan and full file inventory before starting the append job. + private def prepareDeletion( + spark: SparkSession, + snapshot: Snapshot, + sourceKeys: DataFrame, + keys: Seq[String]): Seq[CommitMessage] = { + val reader = table.newSnapshotReader() + if (snapshot != null) reader.withSnapshot(snapshot) + val predicate = targetFilter + .map { + text => + val parsed = spark.sessionState.sqlParser.parseExpression(text) + require( + !SubqueryExpression.hasSubquery(parsed), + "target_filter must be a deterministic predicate over target columns") + val resolved = createDataset(spark, relation) + .filter(expr(text)) + .queryExecution + .analyzed + .collect { case f: org.apache.spark.sql.catalyst.plans.logical.Filter => f.condition } + .head + require(resolved.deterministic, "target_filter must be deterministic") + resolved + } + .getOrElse(TrueLiteral) + convertConditionToPaimonPredicate( + predicate, + relation.output, + rowType, + ignorePartialFailure = true) + .foreach(reader.withFilter) + if (coreOptions.manifestDeleteFileDropStats()) reader.dropStats() + val splits = reader.read().splits().asScala.collect { case d: DataSplit => d }.toSeq + if (splits.nonEmpty) { + val rowId = (relation.output ++ relation.metadataOutput).find(_.name == "_ROW_ID").get + val scan = createNewScanPlan(splits, relation).copy(output = relation.output :+ rowId) + val target = createDataset(spark, scan).filter(toColumn(predicate)).alias("target") + val matches = keys + .map(key => col("target." + quoted(key)) === col("source_keys." + quoted(key))) + .reduce(_ && _) + // Semi join has the same TRUE-only membership semantics as DELETE WHERE key IN source. + val rowIds = target.join(sourceKeys, matches, "left_semi").select(col("target._ROW_ID")) + val vectors = + collectDataEvolutionDeletionVectors(splits, candidateFileMap(splits), rowIds, spark) + writer.persistDeletionVectors(vectors, snapshot) + } else Seq.empty + } + + private def stableSource(spark: SparkSession, targetSnapshot: Snapshot): LogicalPlan = { + require(sourceQuery != null && sourceQuery.trim.nonEmpty, "source_query must not be empty") + val parsed = spark.sessionState.sqlParser.parsePlan(sourceQuery) + parsed.foreach { + case _: Command | _: ParsedStatement => + throw new IllegalArgumentException("source_query must be a query, not a command") + case _ => + } + // Dataset construction eagerly executes commands. Analyze without creating one first. + val analyzed = spark.sessionState.executePlan(parsed).analyzed + analyzed.foreach { + case _: Command => + throw new IllegalArgumentException("source_query must be a query, not a command") + case _ => + } + require(!analyzed.isStreaming, "Streaming source queries are not supported") + val selectors = Seq( + CoreOptions.SCAN_SNAPSHOT_ID.key(), + CoreOptions.SCAN_TAG_NAME.key(), + CoreOptions.SCAN_WATERMARK.key(), + CoreOptions.SCAN_TIMESTAMP.key(), + CoreOptions.SCAN_TIMESTAMP_MILLIS.key(), + CoreOptions.SCAN_VERSION.key() + ) + val snapshots = mutable.Map + .empty[(String, String, Seq[Option[String]], CoreOptions.StartupMode), Option[Snapshot]] + snapshots.put( + ( + table.location().toString, + table.coreOptions().branch(), + selectors.map(k => Option(table.options().get(k))), + CoreOptions.StartupMode.LATEST_FULL), + Option(targetSnapshot) + ) + // Query-time values must not be evaluated afresh by the two preparation jobs. + ComputeCurrentTime(analyzed).transformUpWithSubqueries { Review Comment: `ComputeCurrentTime(analyzed)` only runs `transformAllExpressionsWithPruning` on the top-level plan nodes and does not recurse into `SubqueryExpression.plan` (in Spark's normal optimizer, `OptimizeSubqueries` is what runs `ComputeCurrentTime` inside subqueries). Because `CurrentDate` / `CurrentTimestamp` have `deterministic = true` in Catalyst, `require(p.expressions.forall(_.deterministic))` will pass even when `current_timestamp()` / `current_date()` appears inside a subquery (e.g. `WHERE k IN (SELECT ... WHERE ts <= current_timestamp())`), allowing the two Spark jobs (`prepareDeletion` and `writer.write(input)`) to evaluate different timestamps for the subquery. Consider applying `ComputeCurrentTime` to subquery plans inside `transformUpWithSubqueries` as well. -- 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]
