Why does Apache Spark read unnecessary Parquet columns within nested structures?

My team is building an ETL process to load raw delimited text files into a Parquet based "data lake" using Spark. One of the promises of the Parquet column store is that a query will only read the necessary "column stripes".

But we're seeing unexpected columns being read for nested schema structures.

To demonstrate, here is a POC using Scala and the Spark 2.0.1 shell:

// Preliminary setup
sc.setLogLevel("INFO")
import org.apache.spark.sql.types._
import org.apache.spark.sql._

// Create a schema with nested complex structures
val schema = StructType(Seq(
    StructField("F1", IntegerType),
    StructField("F2", IntegerType),
    StructField("Orig", StructType(Seq(
        StructField("F1", StringType),
        StructField("F2", StringType))))))

// Create some sample data
val data = spark.createDataFrame(
    sc.parallelize(Seq(
        Row(1, 2, Row("1", "2")),
        Row(3, null, Row("3", "ABC")))),
    schema)

// Save it
data.write.mode(SaveMode.Overwrite).parquet("data.parquet")

Then we read the file back into a DataFrame and project to a subset of columns:

// Read it back into another DataFrame
val df = spark.read.parquet("data.parquet")

// Select & show a subset of the columns
df.select($"F1", $"Orig.F1").show

When this runs we see the expected output:

+---+-------+
| F1|Orig_F1|
+---+-------+
|  1|      1|
|  3|      3|
+---+-------+

But... the query plan shows a slightly different story:

The "optimized plan" shows:

val projected = df.select($"F1", $"Orig.F1".as("Orig_F1"))
projected.queryExecution.optimizedPlan
// Project [F1#18, Orig#20.F1 AS Orig_F1#116]
// +- Relation[F1#18,F2#19,Orig#20] parquet

And "explain" shows:

projected.explain
// == Physical Plan ==
// *Project [F1#18, Orig#20.F1 AS Orig_F1#116]
// +- *Scan parquet [F1#18,Orig#20] Format: ParquetFormat, InputPaths: hdfs://sandbox.hortonworks.com:8020/user/stephenp/data.parquet, PartitionFilters: [], PushedFilters: [], ReadSchema: struct<F1:int,Orig:struct<F1:string,F2:string>>

And the INFO logs produced during execution also confirm that the Orig.F2 column is unexpectedly read:

16/10/21 15:13:15 INFO parquet.ParquetReadSupport: Going to read the following fields from the Parquet file:

Parquet form:
message spark_schema {
  optional int32 F1;
  optional group Orig {
    optional binary F1 (UTF8);
    optional binary F2 (UTF8);
  }
}

Catalyst form:
StructType(StructField(F1,IntegerType,true), StructField(Orig,StructType(StructField(F1,StringType,true), StructField(F2,StringType,true)),true))

According to the Dremel paper and the Parquet documentation, columns for complex nested structures should be independently stored and independently retrievable.

Questions:

Is this behavior a limitation of the current Spark query engine? In other words, does Parquet support optimally executing this query, but Spark's query planner is naive?
Or, is this a limitation of the current Parquet implementation?
Or, am I not using the Spark APIs correctly?
Or, am I misunderstanding how Dremel/Parquet column storage is supposed to work?

Solution

The issue has been fixed since Spark 2.4.0. This applies to struct as well as array of structs.

Before Spark 3.0.0:

Set spark.sql.optimizer.nestedSchemaPruning.enabled to true

See related Jira here: https://issues.apache.org/jira/browse/SPARK-4502

After Spark 3.0.0:

spark.sql.optimizer.nestedSchemaPruning.enabled now default is true

Also related SO question: Efficient reading nested parquet column in Spark