Major Description
When reading Parquet files that have been written in legacy mode and schema evolution, we observed that 2-level LIST annotated types are traversed incorrectly.
The root cause is the imprecise check on the underlying element type for Array types (and potentially Map types but I have not checked those yet) that happens here: https://github.com/apache/spark/blob/master/sql/core/src/main/scala/org/apache/spark/sql/execution/datasources/parquet/ParquetRowConverter.scala#L606
The issue is only reproducible with schema evolution with parquet-mr reader and when there are two schemas like this:
File 1:
root |-- col-0: array (nullable = true) | |-- element: struct (containsNull = false) | | |-- col-0: integer (nullable = true)
File 2:
root |-- col-0: array (nullable = true) | |-- element: struct (containsNull = false) | | |-- col-0: integer (nullable = true) | | |-- col-1: integer (nullable = true)
When ParquetRowConverter tries to unwrap ArrayType, it checks if the underlying types between Parquet and Spark match. However, in the case above since the actual schema would include both fields, resulting in mismatch and failure to read File 1:
org.apache.spark.SparkException: Job aborted due to stage failure: Task 1 in stage 11.0 failed 1 times, most recent failure: Lost task 1.0 in stage 11.0 (TID 18) (ip-1-2-3-4.us-west-2.compute.internal executor driver): java.lang.ClassCastException: optional int32 col-0 is not a group at org.apache.parquet.schema.Type.asGroupType(Type.java:248) at org.apache.spark.sql.execution.datasources.parquet.ParquetRowConverter.org$apache$spark$sql$execution$datasources$parquet$ParquetRowConverter$$newConverter(ParquetRowConverter.scala:424) at org.apache.spark.sql.execution.datasources.parquet.ParquetRowConverter$ParquetArrayConverter$ElementConverter.<init>(ParquetRowConverter.scala:633) at org.apache.spark.sql.execution.datasources.parquet.ParquetRowConverter$ParquetArrayConverter.<init>(ParquetRowConverter.scala:616) at org.apache.spark.sql.execution.datasources.parquet.ParquetRowConverter.org$apache$spark$sql$execution$datasources$parquet$ParquetRowConverter$$newConverter(ParquetRowConverter.scala:390) at org.apache.spark.sql.execution.datasources.parquet.ParquetRowConverter.$anonfun$fieldConverters$1(ParquetRowConverter.scala:214) at scala.collection.TraversableLike.$anonfun$map$1(TraversableLike.scala:286) at scala.collection.mutable.ResizableArray.foreach(ResizableArray.scala:62) at scala.collection.mutable.ResizableArray.foreach$(ResizableArray.scala:55) at scala.collection.mutable.ArrayBuffer.foreach(ArrayBuffer.scala:49) at scala.collection.TraversableLike.map(TraversableLike.scala:286) at scala.collection.TraversableLike.map$(TraversableLike.scala:279) at scala.collection.AbstractTraversable.map(Traversable.scala:108) at org.apache.spark.sql.execution.datasources.parquet.ParquetRowConverter.<init>(ParquetRowConverter.scala:210)
This happens due to L606 in ParquetRowConverter:
DataType.equalsIgnoreCompatibleNullability(guessedElementType, elementType)
The code assumes that we are working with 3 level lists and would incorrectly remove the “dummy” level from the Parquet schema.
The actual error varies depending on column names - in this case struct type name matches primitive type name so we end up with "optional int32 col-0 is not a group". In other case, it could fail with IndexOutOfBoundException or NoSuchElementException when the column name is not found in the struct.
The reason it works with 3-level list, that DataType.equalsIgnoreCompatibleNullability(guessedElementType, elementType) always evaluates to false, we remove the “dummy” level and perform struct match which takes into account schema evolution.
Repro:
import org.apache.spark.sql._ import org.apache.spark.sql.types._ val schema1 = StructType( StructField("col-0", ArrayType( StructType( StructField("col-0", IntegerType, true) :: Nil ), containsNull = false )) :: Nil ) val rdd1 = sc.parallelize(Row(Array(Row(1))) :: Nil, 1) val df1 = spark.createDataFrame(rdd1, schema1) df1.write.parquet("/tmp/legacy-parquet") val schema2 = StructType( StructField("col-0", ArrayType( StructType( StructField("col-0", IntegerType, true) :: StructField("col-1", IntegerType, true) :: Nil ), containsNull = false )) :: Nil ) val rdd2 = sc.parallelize(Row(Array(Row(1, 2))) :: Nil, 1) val df2 = spark.createDataFrame(rdd2, schema2) df2.write.mode("append").parquet("/tmp/legacy-parquet") // Fails with: Caused by: ClassCastException: optional int32 col-0 is not a group display(spark.read.schema(schema2).parquet("/tmp/legacy-parquet"))
