[SPARK-6010] [SQL] Merging compatible Parquet schemas before computing splits

sql/core/src/main/scala/org/apache/spark/sql/parquet/ParquetTableOperations.scala

@@ -48,6 +48,7 @@ import org.apache.spark.rdd.RDD
 import org.apache.spark.sql.SQLConf
 import org.apache.spark.sql.catalyst.expressions.{Attribute, Expression, Row, _}
 import org.apache.spark.sql.execution.{LeafNode, SparkPlan, UnaryNode}
+import org.apache.spark.sql.types.{DataType, StructType}
 import org.apache.spark.{Logging, SerializableWritable, TaskContext}
 
 /**
@@ -459,13 +460,30 @@ private[parquet] class FilteringParquetRowInputFormat
     val getGlobalMetaData =
       classOf[ParquetFileWriter].getDeclaredMethod("getGlobalMetaData", classOf[JList[Footer]])
     getGlobalMetaData.setAccessible(true)
-    val globalMetaData = getGlobalMetaData.invoke(null, footers).asInstanceOf[GlobalMetaData]
+    var globalMetaData = getGlobalMetaData.invoke(null, footers).asInstanceOf[GlobalMetaData]
 
     if (globalMetaData == null) {
      val splits = mutable.ArrayBuffer.empty[ParquetInputSplit]
      return splits
     }
 
+    Option(globalMetaData.getKeyValueMetaData.get(RowReadSupport.SPARK_METADATA_KEY)).foreach {
+      schemas =>
+        val mergedSchema = schemas
+          .map(DataType.fromJson(_).asInstanceOf[StructType])
+          .reduce(_ merge _)
+          .json
+
+        val mergedMetadata = globalMetaData
+          .getKeyValueMetaData
+          .updated(RowReadSupport.SPARK_METADATA_KEY, setAsJavaSet(Set(mergedSchema)))
+
+        globalMetaData = new GlobalMetaData(
+          globalMetaData.getSchema,
+          mergedMetadata,
+          globalMetaData.getCreatedBy)
+    }
+
     val readContext = getReadSupport(configuration).init(
       new InitContext(configuration,
         globalMetaData.getKeyValueMetaData,

sql/core/src/main/scala/org/apache/spark/sql/parquet/ParquetTest.scala

@@ -131,6 +131,11 @@ private[sql] trait ParquetTest {
     data.toDF().save(path.getCanonicalPath, "org.apache.spark.sql.parquet", SaveMode.Overwrite)
   }
 
+  protected def makeParquetFile[T <: Product: ClassTag: TypeTag](
+      df: DataFrame, path: File): Unit = {
+    df.save(path.getCanonicalPath, "org.apache.spark.sql.parquet", SaveMode.Overwrite)
+  }
+
   protected def makePartitionDir(
       basePath: File,
       defaultPartitionName: String,

sql/core/src/test/scala/org/apache/spark/sql/parquet/ParquetPartitionDiscoverySuite.scala

@@ -36,6 +36,7 @@ class ParquetPartitionDiscoverySuite extends QueryTest with ParquetTest {
   override val sqlContext: SQLContext = TestSQLContext
 
   import sqlContext._
+  import sqlContext.implicits._
 
   val defaultPartitionName = "__NULL__"
 
@@ -319,4 +320,24 @@ class ParquetPartitionDiscoverySuite extends QueryTest with ParquetTest {
       }
     }
   }
+
+  test("read partitioned table - merging compatible schemas") {
+    withTempDir { base =>
+      makeParquetFile(
+        (1 to 10).map(i => Tuple1(i)).toDF("intField"),
+        makePartitionDir(base, defaultPartitionName, "pi" -> 1))
+
+      makeParquetFile(
+        (1 to 10).map(i => (i, i.toString)).toDF("intField", "stringField"),
+        makePartitionDir(base, defaultPartitionName, "pi" -> 2))
+
+      load(base.getCanonicalPath, "org.apache.spark.sql.parquet").registerTempTable("t")
+
+      withTempTable("t") {
+        checkAnswer(
+          sql("SELECT * FROM t"),
+          (1 to 10).map(i => Row(i, null, 1)) ++ (1 to 10).map(i => Row(i, i.toString, 2)))
+      }
+    }
+  }
 }