wombatu-kun commented on code in PR #20516:
URL: https://github.com/apache/hudi/pull/20516#discussion_r4535031178


##########
hudi-spark-datasource/hudi-spark-common/src/main/scala/org/apache/spark/sql/hudi/streaming/HoodieStreamingSourceSafety.scala:
##########
@@ -0,0 +1,204 @@
+/*
+ * 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.{Dependency, OneToOneDependency, Partition, 
TaskContext}
+import org.apache.spark.rdd.RDD
+import org.apache.spark.sql.{DataFrame, SQLContext}
+import org.apache.spark.sql.catalyst.InternalRow
+import org.apache.spark.sql.execution.LogicalRDD
+import org.apache.spark.sql.execution.datasources.{HadoopFsRelation, 
LogicalRelation}
+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)
+
+  // Spark re-plans the last committed batch during restart without executing 
it. Defer both
+  // validation and relation construction until the scheduler asks for this 
batch's partitions.
+  def defer(sqlContext: SQLContext, schema: StructType)(batch: => DataFrame): 
DataFrame = {
+    val rows = new DeferredBatchRDD(sqlContext, () => 
batch.queryExecution.logical match {
+      case relation: LogicalRDD => relation.rdd
+      case plan =>
+        // Keep the file-group reader's required filters and projection while 
executing this
+        // individual micro-batch as a bounded scan beneath the outer 
streaming LogicalRDD.
+        val boundedPlan = plan.transform {
+          case relation: LogicalRelation => relation.copy(isStreaming = false)
+        }
+        
sparkAdapter.getUnsafeUtils.createDataFrameFrom(sqlContext.sparkSession, 
boundedPlan).queryExecution.toRdd
+    })
+    sparkAdapter.internalCreateDataFrame(sqlContext.sparkSession, rows, 
schema, isStreaming = true)
+  }
+
+  private class DeferredBatchRDD(sqlContext: SQLContext,
+                                 @transient private val build: () => 
RDD[InternalRow])
+    extends RDD[InternalRow](sqlContext.sparkContext, Nil) {
+    @transient private lazy val delegate = build()
+
+    override protected def getPartitions: Array[Partition] = 
delegate.partitions
+
+    override protected def getDependencies: Seq[Dependency[_]] = Seq(new 
OneToOneDependency(delegate))
+
+    override def compute(split: Partition, context: TaskContext): 
Iterator[InternalRow] =
+      firstParent[InternalRow].iterator(split, context)
+  }
+
+  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

Review Comment:
   Since this rejection is now thrown while the sink executes the batch rather 
than from `getBatch`, a Hudi-to-Hudi query with 
`hoodie.datasource.write.streaming.ignore.failed.batch=true` swallows it as an 
ignorable failed batch and commits the offset past the unread range. Is that 
acceptable for `FAIL`, or should the strategy docs say it relies on the sink 
propagating the error?



##########
hudi-spark-datasource/hudi-spark-common/src/main/scala/org/apache/spark/sql/hudi/streaming/HoodieStreamingSourceSafety.scala:
##########
@@ -0,0 +1,204 @@
+/*
+ * 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.{Dependency, OneToOneDependency, Partition, 
TaskContext}
+import org.apache.spark.rdd.RDD
+import org.apache.spark.sql.{DataFrame, SQLContext}
+import org.apache.spark.sql.catalyst.InternalRow
+import org.apache.spark.sql.execution.LogicalRDD
+import org.apache.spark.sql.execution.datasources.{HadoopFsRelation, 
LogicalRelation}
+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)
+
+  // Spark re-plans the last committed batch during restart without executing 
it. Defer both
+  // validation and relation construction until the scheduler asks for this 
batch's partitions.
+  def defer(sqlContext: SQLContext, schema: StructType)(batch: => DataFrame): 
DataFrame = {
+    val rows = new DeferredBatchRDD(sqlContext, () => 
batch.queryExecution.logical match {
+      case relation: LogicalRDD => relation.rdd
+      case plan =>
+        // Keep the file-group reader's required filters and projection while 
executing this
+        // individual micro-batch as a bounded scan beneath the outer 
streaming LogicalRDD.
+        val boundedPlan = plan.transform {
+          case relation: LogicalRelation => relation.copy(isStreaming = false)
+        }
+        
sparkAdapter.getUnsafeUtils.createDataFrameFrom(sqlContext.sparkSession, 
boundedPlan).queryExecution.toRdd
+    })
+    sparkAdapter.internalCreateDataFrame(sqlContext.sparkSession, rows, 
schema, isStreaming = true)

Review Comment:
   Wrapping every batch in `DeferredBatchRDD` hides the scan from the outer 
query's optimizer, so on the default file-group-reader path healthy batches 
lose column pruning and filter/partition pushdown and read every column of 
every file in the range. Was that cost intended, or could `getBatch` validate 
eagerly and defer only a rejection, so healthy batches keep the 
`LogicalRelation` in the plan?



-- 
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