[SPARK-18337] Complete mode memory sinks should be able to recover from checkpoints

## What changes were proposed in this pull request?

It would be nice if memory sinks can also recover from checkpoints. For correctness reasons, the only time we should support it is in `Complete` OutputMode. We can support this in CompleteMode, because the output of the StateStore is already persisted in the checkpoint directory.

## How was this patch tested?

Unit test

Author: Burak Yavuz <brkyvz@gmail.com>

Closes #15801 from brkyvz/mem-stream.

(cherry picked from commit 2afdaa98)
Signed-off-by: Tathagata Das <tathagata.das1565@gmail.com>

sql/core/src/main/scala/org/apache/spark/sql/streaming/DataStreamWriter.scala

@@ -222,14 +222,16 @@ final class DataStreamWriter[T] private[sql](ds: Dataset[T]) {
 
       val sink = new MemorySink(df.schema, outputMode)
       val resultDf = Dataset.ofRows(df.sparkSession, new MemoryPlan(sink))
+      val chkpointLoc = extraOptions.get("checkpointLocation")
+      val recoverFromChkpoint = chkpointLoc.isDefined && outputMode == OutputMode.Complete()
       val query = df.sparkSession.sessionState.streamingQueryManager.startQuery(
         extraOptions.get("queryName"),
-        extraOptions.get("checkpointLocation"),
+        chkpointLoc,
         df,
         sink,
         outputMode,
         useTempCheckpointLocation = true,
-        recoverFromCheckpointLocation = false,
+        recoverFromCheckpointLocation = recoverFromChkpoint,
         trigger = trigger)
       resultDf.createOrReplaceTempView(query.name)
       query

sql/core/src/test/scala/org/apache/spark/sql/streaming/test/DataStreamReaderWriterSuite.scala

@@ -17,6 +17,7 @@
 
 package org.apache.spark.sql.streaming.test
 
+import java.io.File
 import java.util.concurrent.TimeUnit
 
 import scala.concurrent.duration._
@@ -467,4 +468,68 @@ class DataStreamReaderWriterSuite extends StreamTest with BeforeAndAfter {
     val sq = df.writeStream.format("console").start()
     sq.stop()
   }
+
+  test("MemorySink can recover from a checkpoint in Complete Mode") {
+    import testImplicits._
+    val ms = new MemoryStream[Int](0, sqlContext)
+    val df = ms.toDF().toDF("a")
+    val checkpointLoc = newMetadataDir
+    val checkpointDir = new File(checkpointLoc, "offsets")
+    checkpointDir.mkdirs()
+    assert(checkpointDir.exists())
+    val tableName = "test"
+    def startQuery: StreamingQuery = {
+      df.groupBy("a")
+        .count()
+        .writeStream
+        .format("memory")
+        .queryName(tableName)
+        .option("checkpointLocation", checkpointLoc)
+        .outputMode("complete")
+        .start()
+    }
+    // no exception here
+    val q = startQuery
+    ms.addData(0, 1)
+    q.processAllAvailable()
+    q.stop()
+
+    checkAnswer(
+      spark.table(tableName),
+      Seq(Row(0, 1), Row(1, 1))
+    )
+    spark.sql(s"drop table $tableName")
+    // verify table is dropped
+    intercept[AnalysisException](spark.table(tableName).collect())
+    val q2 = startQuery
+    ms.addData(0)
+    q2.processAllAvailable()
+    checkAnswer(
+      spark.table(tableName),
+      Seq(Row(0, 2), Row(1, 1))
+    )
+
+    q2.stop()
+  }
+
+  test("append mode memory sink's do not support checkpoint recovery") {
+    import testImplicits._
+    val ms = new MemoryStream[Int](0, sqlContext)
+    val df = ms.toDF().toDF("a")
+    val checkpointLoc = newMetadataDir
+    val checkpointDir = new File(checkpointLoc, "offsets")
+    checkpointDir.mkdirs()
+    assert(checkpointDir.exists())
+
+    val e = intercept[AnalysisException] {
+      df.writeStream
+        .format("memory")
+        .queryName("test")
+        .option("checkpointLocation", checkpointLoc)
+        .outputMode("append")
+        .start()
+    }
+    assert(e.getMessage.contains("does not support recovering"))
+    assert(e.getMessage.contains("checkpoint location"))
+  }
 }