Major Description
nondeterministic expression like monotonically_increasing_id fails column pruning
spark.range(10).selectExpr("id as key", "id * 2 as value").
write.format("parquet").save("/tmp/source")
spark.range(10).selectExpr("id as key", "id * 3 as s1", "id * 5 as s2").
write.format("parquet").save("/tmp/target")
val sourceDF = spark.read.parquet("/tmp/source")
val targetDF = spark.read.parquet("/tmp/target").
withColumn("row_id", monotonically_increasing_id())
sourceDF.join(targetDF, "key").select("key", "row_id").explain()
Spark reads all columns from targetDF, but actually, we only need `key` column.
scala> sourceDF.join(targetDF, "key").select("key", "row_id").explain() == Physical Plan == *(2) Project [key#78L, row_id#88L] +- *(2) BroadcastHashJoin [key#78L], [key#82L], Inner, BuildLeft :- BroadcastExchange HashedRelationBroadcastMode(List(input[0, bigint, true])) : +- *(1) Project [key#78L] : +- *(1) Filter isnotnull(key#78L) : +- *(1) FileScan parquet [key#78L] Batched: true, Format: Parquet, Location: InMemoryFileIndex[file:/tmp/source], PartitionFilters: [], PushedFilters: [IsNotNull(key)], ReadSchema: struct<key:bigint> +- *(2) Filter isnotnull(key#82L) +- *(2) Project [key#82L, monotonically_increasing_id() AS row_id#88L] +- *(2) FileScan parquet [key#82L,s1#83L,s2#84L] Batched: true, Format: Parquet, Location: InMemoryFileIndex[file:/tmp/target], PartitionFilters: [], PushedFilters: [], ReadSchema: struct<key:bigint,s1:bigint,s2:bigint>