danny0405 commented on code in PR #20516:
URL: https://github.com/apache/hudi/pull/20516#discussion_r4533510961


##########
hudi-spark-datasource/hudi-spark-common/src/main/scala/org/apache/spark/sql/hudi/streaming/HoodieStreamSourceV2.scala:
##########
@@ -113,62 +117,68 @@ class HoodieStreamSourceV2(sqlContext: SQLContext,
       .getOrElse(initialOffsets)
     var endOffset = HoodieSourceOffset(end)
 
-    // We update the offsets here since until this point the latest offsets 
have been
-    // calculated no matter if it is in the expected version.
-    // We translate them here, then the rest logic should be intact.
-    startOffset = 
HoodieSourceOffset(translateCheckpoint(startOffset.offsetCommitTime))
-    endOffset = 
HoodieSourceOffset(translateCheckpoint(endOffset.offsetCommitTime))
-
-    if (startOffset == endOffset) {
-      sparkAdapter.internalCreateDataFrame(sqlContext.sparkSession,
-        sqlContext.sparkContext.emptyRDD[InternalRow].setName("empty"), 
schema, isStreaming = true)
-    } else {
-      val (startCompletionTime, rangeType) = 
getStartCompletionTimeAndRangeType(startOffset)
-      if (isCDCQuery) {
-        val cdcOptions = Map(
-          DataSourceReadOptions.START_COMMIT.key() -> startCompletionTime,
-          DataSourceReadOptions.END_COMMIT.key() -> endOffset.offsetCommitTime
-        )
-        val relation = if (tableType == HoodieTableType.COPY_ON_WRITE) {
-          new HoodieCopyOnWriteCDCHadoopFsRelationFactory(
-            sqlContext, metaClient, parameters ++ cdcOptions, schemaOption, 
isBootstrappedTable, rangeType).build()
-        } else {
-          new HoodieMergeOnReadCDCHadoopFsRelationFactory(
-            sqlContext, metaClient, parameters ++ cdcOptions, schemaOption, 
isBootstrappedTable, rangeType).build()
-        }
-        sparkAdapter.createStreamingDataFrame(sqlContext, relation, 
HoodieCDCFileIndex.FULL_CDC_SPARK_SCHEMA)
+    val recovery = if (isCDCQuery) None else 
HoodieStreamingSourceSafety.recover(
+      sqlContext, metaClient, schema, parameters, recoveryStrategy, 
startOffset.offsetCommitTime,
+      endOffset.offsetCommitTime, completionTime = true, resumed = 
start.isDefined)
+    recovery.getOrElse {
+      // We update the offsets here since until this point the latest offsets 
have been
+      // calculated no matter if it is in the expected version.
+      // We translate them here, then the rest logic should be intact.
+      startOffset = 
HoodieSourceOffset(translateCheckpoint(startOffset.offsetCommitTime))
+      endOffset = 
HoodieSourceOffset(translateCheckpoint(endOffset.offsetCommitTime))
+
+      if (startOffset == endOffset) {
+        sparkAdapter.internalCreateDataFrame(sqlContext.sparkSession,
+          sqlContext.sparkContext.emptyRDD[InternalRow].setName("empty"), 
schema, isStreaming = true)
       } else {
-        // Consume the data between (startCommitTime, endCommitTime]
-        val incParams = parameters ++ Map(
-          DataSourceReadOptions.QUERY_TYPE.key -> 
DataSourceReadOptions.QUERY_TYPE_INCREMENTAL_OPT_VAL,
-          DataSourceReadOptions.START_COMMIT.key -> startCompletionTime,
-          DataSourceReadOptions.END_COMMIT.key -> endOffset.offsetCommitTime
-        )
-
-        if (enableFileGroupReader) {
+        val (startCompletionTime, rangeType) = 
getStartCompletionTimeAndRangeType(startOffset)
+        if (isCDCQuery) {
+          val cdcOptions = Map(
+            DataSourceReadOptions.START_COMMIT.key() -> startCompletionTime,
+            DataSourceReadOptions.END_COMMIT.key() -> 
endOffset.offsetCommitTime
+          )
           val relation = if (tableType == HoodieTableType.COPY_ON_WRITE) {
-            new 
HoodieCopyOnWriteIncrementalHadoopFsRelationFactoryV2(sqlContext, metaClient, 
incParams, Option(schema), isBootstrappedTable, rangeType)
-              .build()
+            new HoodieCopyOnWriteCDCHadoopFsRelationFactory(
+              sqlContext, metaClient, parameters ++ cdcOptions, schemaOption, 
isBootstrappedTable, rangeType).build()
           } else {
-            new 
HoodieMergeOnReadIncrementalHadoopFsRelationFactoryV2(sqlContext, metaClient, 
incParams, Option(schema), isBootstrappedTable, rangeType)
-              .build()
+            new HoodieMergeOnReadCDCHadoopFsRelationFactory(
+              sqlContext, metaClient, parameters ++ cdcOptions, schemaOption, 
isBootstrappedTable, rangeType).build()
           }
-          sparkAdapter.createStreamingDataFrame(sqlContext, relation, schema)
+          sparkAdapter.createStreamingDataFrame(sqlContext, relation, 
HoodieCDCFileIndex.FULL_CDC_SPARK_SCHEMA)
         } else {
-          val requiredColumns = schema.fields.map(_.name)
-          val rdd = tableType match {
-            case HoodieTableType.COPY_ON_WRITE =>
-              val serDe = HoodieSparkUtils.getCatalystRowSerDe(schema)
-              new IncrementalRelationV2(sqlContext, incParams, Some(schema), 
metaClient, rangeType)
-                .buildScan(requiredColumns)
-                .map(serDe.serializeRow)
-            case HoodieTableType.MERGE_ON_READ =>
-              MergeOnReadIncrementalRelationV2(sqlContext, incParams, 
metaClient, Some(schema), rangeType = rangeType)
-                .buildScan(requiredColumns, Array.empty[Filter])
-                .asInstanceOf[RDD[InternalRow]]
-            case _ => throw new IllegalArgumentException(s"UnSupport 
tableType: $tableType")
-          }
-          sparkAdapter.internalCreateDataFrame(sqlContext.sparkSession, rdd, 
schema, isStreaming = true)}
+          // Consume the data between (startCommitTime, endCommitTime]
+          val incParams = parameters ++ Map(
+            
DataSourceReadOptions.INCREMENTAL_FALLBACK_TO_FULL_TABLE_SCAN.key() -> "false",

Review Comment:
   Fixed in 57883c3ccfe. Both sources override the legacy fallback only after 
INIT_OFFSET; fresh earliest starts retain the configured fallback. The 
operational guide, migration note, and config description now call out this 
exception. The real-cleaner tests also execute a fresh earliest read and check 
its rows.



##########
hudi-spark-datasource/hudi-spark-common/src/main/scala/org/apache/spark/sql/hudi/streaming/HoodieStreamingSourceSafety.scala:
##########
@@ -0,0 +1,163 @@
+/*
+ * 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.streaming
+
+import org.apache.hudi.{DataSourceReadOptions, DefaultSource, 
SparkAdapterSupport}
+import org.apache.hudi.common.model.WriteOperationType
+import org.apache.hudi.common.table.HoodieTableMetaClient
+import org.apache.hudi.common.table.timeline.{HoodieInstant, HoodieTimeline}
+import org.apache.hudi.common.util.CleanerUtils
+import org.apache.hudi.exception.HoodieException
+
+import org.apache.spark.sql.{DataFrame, SQLContext}
+import org.apache.spark.sql.catalyst.InternalRow
+import org.apache.spark.sql.execution.datasources.HadoopFsRelation
+import org.apache.spark.sql.types.StructType
+
+import scala.collection.JavaConverters._
+
+/** Conservative reset and history validation for checkpointed latest-state 
streaming consumers. */
+private[streaming] object HoodieStreamingSourceSafety extends 
SparkAdapterSupport {
+  private class RebuildRequired(message: String) extends 
HoodieException(message)
+
+  def strategy(parameters: Map[String, String]): String = {
+    val config = DataSourceReadOptions.STREAMING_READ_RECOVERY_STRATEGY
+    val value = parameters.getOrElse(config.key(), config.defaultValue())
+    require(Set("FAIL", "READ_UPTO_LATEST_COMMIT", 
"READ_LATEST").contains(value),
+      s"Invalid ${config.key()}=$value; expected FAIL, READ_UPTO_LATEST_COMMIT 
or READ_LATEST")
+    value
+  }
+
+  def recover(sqlContext: SQLContext,
+              metaClient: HoodieTableMetaClient,
+              schema: => StructType,
+              parameters: Map[String, String],
+              strategy: String,
+              start: String,
+              end: String,
+              completionTime: Boolean,
+              resumed: Boolean): Option[DataFrame] = {
+    lazy val recoverySchema = schema
+    try {
+      validate(metaClient, start, end, completionTime, resumed)

Review Comment:
   Fixed in 57883c3ccfe. Validation and relation construction are deferred 
until execution through a lazy batch RDD, so getBatch can return the discarded 
restart plan without validating its delivered range. Initial-offset capture 
remains eager, and the file-group reader's filters/projection are preserved. 
Tests cover discarded plans after a reset/recovery and after real cleaning, 
then execute a readable resumed batch under FAIL.



##########
hudi-spark-datasource/hudi-spark-common/src/main/scala/org/apache/spark/sql/hudi/streaming/HoodieStreamingSourceSafety.scala:
##########
@@ -0,0 +1,163 @@
+/*
+ * 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.streaming
+
+import org.apache.hudi.{DataSourceReadOptions, DefaultSource, 
SparkAdapterSupport}
+import org.apache.hudi.common.model.WriteOperationType
+import org.apache.hudi.common.table.HoodieTableMetaClient
+import org.apache.hudi.common.table.timeline.{HoodieInstant, HoodieTimeline}
+import org.apache.hudi.common.util.CleanerUtils
+import org.apache.hudi.exception.HoodieException
+
+import org.apache.spark.sql.{DataFrame, SQLContext}
+import org.apache.spark.sql.catalyst.InternalRow
+import org.apache.spark.sql.execution.datasources.HadoopFsRelation
+import org.apache.spark.sql.types.StructType
+
+import scala.collection.JavaConverters._
+
+/** Conservative reset and history validation for checkpointed latest-state 
streaming consumers. */
+private[streaming] object HoodieStreamingSourceSafety extends 
SparkAdapterSupport {
+  private class RebuildRequired(message: String) extends 
HoodieException(message)
+
+  def strategy(parameters: Map[String, String]): String = {
+    val config = DataSourceReadOptions.STREAMING_READ_RECOVERY_STRATEGY
+    val value = parameters.getOrElse(config.key(), config.defaultValue())
+    require(Set("FAIL", "READ_UPTO_LATEST_COMMIT", 
"READ_LATEST").contains(value),
+      s"Invalid ${config.key()}=$value; expected FAIL, READ_UPTO_LATEST_COMMIT 
or READ_LATEST")
+    value
+  }
+
+  def recover(sqlContext: SQLContext,
+              metaClient: HoodieTableMetaClient,
+              schema: => StructType,
+              parameters: Map[String, String],
+              strategy: String,
+              start: String,
+              end: String,
+              completionTime: Boolean,
+              resumed: Boolean): Option[DataFrame] = {
+    lazy val recoverySchema = schema
+    try {
+      validate(metaClient, start, end, completionTime, resumed)
+      None
+    } catch {
+      case error: RebuildRequired => strategy match {
+        case "FAIL" => throw error
+        case "READ_LATEST" =>
+          Some(sparkAdapter.internalCreateDataFrame(sqlContext.sparkSession,
+            sqlContext.sparkContext.emptyRDD[InternalRow], recoverySchema, 
isStreaming = true))
+        case "READ_UPTO_LATEST_COMMIT" =>
+          // Spark owns the batch end/checkpoint. Never read a newer snapshot 
and checkpoint an older end.
+          val writes = 
metaClient.reloadActiveTimeline().getCommitsTimeline.filterCompletedInstants()
+            .getInstants.asScala.toSeq
+          def offset(i: HoodieInstant): String = if (completionTime) 
i.getCompletionTime else i.requestedTime()
+          val included = writes.filter(i => offset(i) <= end)
+          val bound = included.map(_.requestedTime()).reduceOption((a, b) => 
if (a > b) a else b)
+          if (bound.isEmpty || writes.exists(i => offset(i) > end && 
i.requestedTime() <= bound.get)) {

Review Comment:
   Fixed in 57883c3ccfe. The ordering guard now exempts COMPACT, LOG_COMPACT, 
and CLUSTER operations, while still rejecting late data writes. Completion-time 
COW/MOR regressions pin the snapshot end, add each late service completion, and 
verify the exact snapshot rows; a late UPSERT still fails reconstruction.



##########
hudi-spark-datasource/hudi-spark-common/src/main/scala/org/apache/spark/sql/hudi/streaming/HoodieStreamingSourceSafety.scala:
##########
@@ -0,0 +1,163 @@
+/*
+ * 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.streaming
+
+import org.apache.hudi.{DataSourceReadOptions, DefaultSource, 
SparkAdapterSupport}
+import org.apache.hudi.common.model.WriteOperationType
+import org.apache.hudi.common.table.HoodieTableMetaClient
+import org.apache.hudi.common.table.timeline.{HoodieInstant, HoodieTimeline}
+import org.apache.hudi.common.util.CleanerUtils
+import org.apache.hudi.exception.HoodieException
+
+import org.apache.spark.sql.{DataFrame, SQLContext}
+import org.apache.spark.sql.catalyst.InternalRow
+import org.apache.spark.sql.execution.datasources.HadoopFsRelation
+import org.apache.spark.sql.types.StructType
+
+import scala.collection.JavaConverters._
+
+/** Conservative reset and history validation for checkpointed latest-state 
streaming consumers. */
+private[streaming] object HoodieStreamingSourceSafety extends 
SparkAdapterSupport {
+  private class RebuildRequired(message: String) extends 
HoodieException(message)
+
+  def strategy(parameters: Map[String, String]): String = {
+    val config = DataSourceReadOptions.STREAMING_READ_RECOVERY_STRATEGY
+    val value = parameters.getOrElse(config.key(), config.defaultValue())
+    require(Set("FAIL", "READ_UPTO_LATEST_COMMIT", 
"READ_LATEST").contains(value),
+      s"Invalid ${config.key()}=$value; expected FAIL, READ_UPTO_LATEST_COMMIT 
or READ_LATEST")
+    value
+  }
+
+  def recover(sqlContext: SQLContext,
+              metaClient: HoodieTableMetaClient,
+              schema: => StructType,
+              parameters: Map[String, String],
+              strategy: String,
+              start: String,
+              end: String,
+              completionTime: Boolean,
+              resumed: Boolean): Option[DataFrame] = {
+    lazy val recoverySchema = schema
+    try {
+      validate(metaClient, start, end, completionTime, resumed)
+      None
+    } catch {
+      case error: RebuildRequired => strategy match {
+        case "FAIL" => throw error
+        case "READ_LATEST" =>
+          Some(sparkAdapter.internalCreateDataFrame(sqlContext.sparkSession,
+            sqlContext.sparkContext.emptyRDD[InternalRow], recoverySchema, 
isStreaming = true))
+        case "READ_UPTO_LATEST_COMMIT" =>
+          // Spark owns the batch end/checkpoint. Never read a newer snapshot 
and checkpoint an older end.
+          val writes = 
metaClient.reloadActiveTimeline().getCommitsTimeline.filterCompletedInstants()
+            .getInstants.asScala.toSeq
+          def offset(i: HoodieInstant): String = if (completionTime) 
i.getCompletionTime else i.requestedTime()
+          val included = writes.filter(i => offset(i) <= end)
+          val bound = included.map(_.requestedTime()).reduceOption((a, b) => 
if (a > b) a else b)
+          if (bound.isEmpty || writes.exists(i => offset(i) > end && 
i.requestedTime() <= bound.get)) {
+            throw new HoodieException("Cannot reconstruct the recovery 
snapshot at the checkpointed batch end. "
+              + "Rebuild downstream tables and checkpoints together.", error)
+          }
+          val options = parameters ++ Map(
+            DataSourceReadOptions.QUERY_TYPE.key() -> 
DataSourceReadOptions.QUERY_TYPE_SNAPSHOT_OPT_VAL,
+            DataSourceReadOptions.TIME_TRAVEL_AS_OF_INSTANT.key() -> bound.get)
+          val relation = DefaultSource.createRelation(sqlContext, metaClient, 
recoverySchema, options)
+          Some(relation match {
+            case fsRelation: HadoopFsRelation => 
sparkAdapter.createStreamingDataFrame(sqlContext, fsRelation, recoverySchema)
+            case _ => 
sparkAdapter.internalCreateDataFrame(sqlContext.sparkSession,
+              
sqlContext.baseRelationToDataFrame(relation).queryExecution.toRdd, 
recoverySchema, isStreaming = true)
+          })
+      }
+    }
+  }
+
+  def validate(metaClient: HoodieTableMetaClient,
+               start: String,
+               end: String,
+               completionTime: Boolean,
+               resumed: Boolean): Unit = {
+    if (start != HoodieTimeline.INIT_INSTANT_TS) {
+      validateRetainedRange(metaClient, start, end, completionTime, resumed)
+    }
+  }
+
+  private def validateRetainedRange(metaClient: HoodieTableMetaClient,
+                                    start: String,
+                                    end: String,
+                                    completionTime: Boolean,
+                                    resumed: Boolean): Unit = {
+    val timeline = metaClient.reloadActiveTimeline()
+    val completed = 
timeline.filterCompletedInstants().getInstants.asScala.toSeq
+    def offset(instant: HoodieInstant): String = {
+      if (completionTime) instant.getCompletionTime else 
instant.requestedTime()
+    }
+    def fail(reason: String): Nothing = {
+      throw new RebuildRequired(s"Streaming source rebuild required for 
${metaClient.getBasePath}: $reason. "
+        + "Stop downstream consumers and rebuild affected downstream tables 
and checkpoints in dependency order. "
+        + s"To explicitly accept recovery risks, set 
${DataSourceReadOptions.STREAMING_READ_RECOVERY_STRATEGY.key()} "
+        + "to READ_UPTO_LATEST_COMMIT (snapshot replay) or READ_LATEST (skip 
the gap).")
+    }
+
+    val checkpointInstant = completed.find(i => offset(i) == start)
+    if (resumed && checkpointInstant.isEmpty) {
+      fail(s"saved offset $start is absent from the active timeline; history 
expired or the source table was recreated")
+    }
+    if (end < start) {
+      fail(s"source offset moved backwards from $start to $end")
+    }
+
+    val firstWrite = 
timeline.getWriteTimeline.filterCompletedInstants().firstInstant()
+    val requestedStart = checkpointInstant.map(_.requestedTime())
+    val unreadWrites = 
timeline.getCommitsTimeline.filterCompletedInstants().getInstants.asScala
+      .filter(i => offset(i) > start && offset(i) <= end).toSeq
+    val earliestUnread = unreadWrites.map(_.requestedTime()).reduceOption((a, 
b) => if (a < b) a else b)
+    val completedByRequestedTime = completed.map(i => i.requestedTime() -> 
i).toMap
+    // Usually only new cleans matter. Retain earlier cleans for 
late-completing writes,
+    // or a user-specified initial offset whose requested-time boundary is 
unknown.
+    timeline.getCleanerTimeline.filterCompletedInstants().getInstants.asScala
+      .filter(clean => offset(clean) > start || requestedStart.isEmpty ||
+        requestedStart.exists(_ < clean.requestedTime()) || 
earliestUnread.exists(_ < clean.requestedTime()))
+      .foreach { clean =>
+        val metadata = CleanerUtils.getCleanerMetadata(metaClient, clean)
+        val retained = metadata.getEarliestCommitToRetain
+        // File-version cleaning does not expose a retention boundary. Refuse 
to infer safety.
+        if ((retained == null || retained.isEmpty) && 
metadata.getTotalFilesDeleted > 0 && offset(clean) > start) {
+          fail("the cleaner did not record an earliest retained commit; 
KEEP_LATEST_FILE_VERSIONS "
+            + "cannot prove incremental safety after deleting files")
+        }
+        val boundary = Option(retained).filter(_.nonEmpty)
+        val retainedInstant = completedByRequestedTime.get(retained)
+        val expired = boundary.exists { floor =>
+          requestedStart.exists(_ < floor) || earliestUnread.exists(_ < floor) 
||

Review Comment:
   Agreed; removed the requestedStart < floor disjunct in 57883c3ccfe. 
Retention is checked against unread writes. Added completion-time COW/MOR cases 
with a delivered compaction whose requested time is below the cleaner floor; 
the next retained rows remain readable.



##########
hudi-spark-datasource/hudi-spark-common/src/main/scala/org/apache/spark/sql/hudi/streaming/HoodieStreamingSourceSafety.scala:
##########
@@ -0,0 +1,163 @@
+/*
+ * 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.streaming
+
+import org.apache.hudi.{DataSourceReadOptions, DefaultSource, 
SparkAdapterSupport}
+import org.apache.hudi.common.model.WriteOperationType
+import org.apache.hudi.common.table.HoodieTableMetaClient
+import org.apache.hudi.common.table.timeline.{HoodieInstant, HoodieTimeline}
+import org.apache.hudi.common.util.CleanerUtils
+import org.apache.hudi.exception.HoodieException
+
+import org.apache.spark.sql.{DataFrame, SQLContext}
+import org.apache.spark.sql.catalyst.InternalRow
+import org.apache.spark.sql.execution.datasources.HadoopFsRelation
+import org.apache.spark.sql.types.StructType
+
+import scala.collection.JavaConverters._
+
+/** Conservative reset and history validation for checkpointed latest-state 
streaming consumers. */
+private[streaming] object HoodieStreamingSourceSafety extends 
SparkAdapterSupport {
+  private class RebuildRequired(message: String) extends 
HoodieException(message)
+
+  def strategy(parameters: Map[String, String]): String = {
+    val config = DataSourceReadOptions.STREAMING_READ_RECOVERY_STRATEGY
+    val value = parameters.getOrElse(config.key(), config.defaultValue())
+    require(Set("FAIL", "READ_UPTO_LATEST_COMMIT", 
"READ_LATEST").contains(value),
+      s"Invalid ${config.key()}=$value; expected FAIL, READ_UPTO_LATEST_COMMIT 
or READ_LATEST")
+    value
+  }
+
+  def recover(sqlContext: SQLContext,
+              metaClient: HoodieTableMetaClient,
+              schema: => StructType,
+              parameters: Map[String, String],
+              strategy: String,
+              start: String,
+              end: String,
+              completionTime: Boolean,
+              resumed: Boolean): Option[DataFrame] = {
+    lazy val recoverySchema = schema
+    try {
+      validate(metaClient, start, end, completionTime, resumed)
+      None
+    } catch {
+      case error: RebuildRequired => strategy match {
+        case "FAIL" => throw error
+        case "READ_LATEST" =>
+          Some(sparkAdapter.internalCreateDataFrame(sqlContext.sparkSession,
+            sqlContext.sparkContext.emptyRDD[InternalRow], recoverySchema, 
isStreaming = true))
+        case "READ_UPTO_LATEST_COMMIT" =>
+          // Spark owns the batch end/checkpoint. Never read a newer snapshot 
and checkpoint an older end.
+          val writes = 
metaClient.reloadActiveTimeline().getCommitsTimeline.filterCompletedInstants()
+            .getInstants.asScala.toSeq
+          def offset(i: HoodieInstant): String = if (completionTime) 
i.getCompletionTime else i.requestedTime()
+          val included = writes.filter(i => offset(i) <= end)
+          val bound = included.map(_.requestedTime()).reduceOption((a, b) => 
if (a > b) a else b)
+          if (bound.isEmpty || writes.exists(i => offset(i) > end && 
i.requestedTime() <= bound.get)) {
+            throw new HoodieException("Cannot reconstruct the recovery 
snapshot at the checkpointed batch end. "
+              + "Rebuild downstream tables and checkpoints together.", error)
+          }
+          val options = parameters ++ Map(
+            DataSourceReadOptions.QUERY_TYPE.key() -> 
DataSourceReadOptions.QUERY_TYPE_SNAPSHOT_OPT_VAL,
+            DataSourceReadOptions.TIME_TRAVEL_AS_OF_INSTANT.key() -> bound.get)
+          val relation = DefaultSource.createRelation(sqlContext, metaClient, 
recoverySchema, options)
+          Some(relation match {
+            case fsRelation: HadoopFsRelation => 
sparkAdapter.createStreamingDataFrame(sqlContext, fsRelation, recoverySchema)
+            case _ => 
sparkAdapter.internalCreateDataFrame(sqlContext.sparkSession,
+              
sqlContext.baseRelationToDataFrame(relation).queryExecution.toRdd, 
recoverySchema, isStreaming = true)
+          })
+      }
+    }
+  }
+
+  def validate(metaClient: HoodieTableMetaClient,
+               start: String,
+               end: String,
+               completionTime: Boolean,
+               resumed: Boolean): Unit = {
+    if (start != HoodieTimeline.INIT_INSTANT_TS) {
+      validateRetainedRange(metaClient, start, end, completionTime, resumed)
+    }
+  }
+
+  private def validateRetainedRange(metaClient: HoodieTableMetaClient,
+                                    start: String,
+                                    end: String,
+                                    completionTime: Boolean,
+                                    resumed: Boolean): Unit = {
+    val timeline = metaClient.reloadActiveTimeline()
+    val completed = 
timeline.filterCompletedInstants().getInstants.asScala.toSeq
+    def offset(instant: HoodieInstant): String = {
+      if (completionTime) instant.getCompletionTime else 
instant.requestedTime()
+    }
+    def fail(reason: String): Nothing = {
+      throw new RebuildRequired(s"Streaming source rebuild required for 
${metaClient.getBasePath}: $reason. "
+        + "Stop downstream consumers and rebuild affected downstream tables 
and checkpoints in dependency order. "
+        + s"To explicitly accept recovery risks, set 
${DataSourceReadOptions.STREAMING_READ_RECOVERY_STRATEGY.key()} "
+        + "to READ_UPTO_LATEST_COMMIT (snapshot replay) or READ_LATEST (skip 
the gap).")
+    }
+
+    val checkpointInstant = completed.find(i => offset(i) == start)
+    if (resumed && checkpointInstant.isEmpty) {
+      fail(s"saved offset $start is absent from the active timeline; history 
expired or the source table was recreated")
+    }
+    if (end < start) {
+      fail(s"source offset moved backwards from $start to $end")
+    }
+
+    val firstWrite = 
timeline.getWriteTimeline.filterCompletedInstants().firstInstant()
+    val requestedStart = checkpointInstant.map(_.requestedTime())
+    val unreadWrites = 
timeline.getCommitsTimeline.filterCompletedInstants().getInstants.asScala
+      .filter(i => offset(i) > start && offset(i) <= end).toSeq
+    val earliestUnread = unreadWrites.map(_.requestedTime()).reduceOption((a, 
b) => if (a < b) a else b)
+    val completedByRequestedTime = completed.map(i => i.requestedTime() -> 
i).toMap
+    // Usually only new cleans matter. Retain earlier cleans for 
late-completing writes,
+    // or a user-specified initial offset whose requested-time boundary is 
unknown.
+    timeline.getCleanerTimeline.filterCompletedInstants().getInstants.asScala
+      .filter(clean => offset(clean) > start || requestedStart.isEmpty ||
+        requestedStart.exists(_ < clean.requestedTime()) || 
earliestUnread.exists(_ < clean.requestedTime()))
+      .foreach { clean =>
+        val metadata = CleanerUtils.getCleanerMetadata(metaClient, clean)
+        val retained = metadata.getEarliestCommitToRetain
+        // File-version cleaning does not expose a retention boundary. Refuse 
to infer safety.
+        if ((retained == null || retained.isEmpty) && 
metadata.getTotalFilesDeleted > 0 && offset(clean) > start) {
+          fail("the cleaner did not record an earliest retained commit; 
KEEP_LATEST_FILE_VERSIONS "
+            + "cannot prove incremental safety after deleting files")
+        }
+        val boundary = Option(retained).filter(_.nonEmpty)
+        val retainedInstant = completedByRequestedTime.get(retained)
+        val expired = boundary.exists { floor =>
+          requestedStart.exists(_ < floor) || earliestUnread.exists(_ < floor) 
||
+            (requestedStart.isEmpty && retainedInstant.exists(i => start < 
offset(i))) ||
+            (requestedStart.isEmpty && retainedInstant.isEmpty && 
firstWrite.isPresent)

Review Comment:
   Fixed in 57883c3ccfe. When the cleaner floor is absent from the active 
timeline, the unknown-start branch now requires start to precede the first 
active write. Added recent explicit-start regressions with an old archived 
cleaner floor for COW and MOR.



-- 
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]

Reply via email to