Major Description
Simple filters on dataframes joined with joinWith() are missing an opportunity to get pushed into the scan because they're written in terms of named_struct that could be removed by the optimizer.
Given the following simple query over two dataframes:
scala> val df = spark.read.parquet("one_million") df: org.apache.spark.sql.DataFrame = [id: bigint, id2: bigint] scala> val df2 = spark.read.parquet("one_million") df2: org.apache.spark.sql.DataFrame = [id: bigint, id2: bigint] scala> df.joinWith(df2, df2.col("id") === df.col("id2")).filter("_2.id > 30").explain == Physical Plan == *(2) BroadcastHashJoin [_1#94.id2], [_2#95.id], Inner, BuildRight :- *(2) Project [named_struct(id, id#0L, id2, id2#1L) AS _1#94] : +- *(2) FileScan parquet [id#0L,id2#1L] Batched: true, Format: Parquet, Location: InMemoryFileIndex[file:/Users/henry/src/spark/one_million], PartitionFilters: [], PushedFilters: [], ReadSchema: struct<id:bigint,id2:bigint> +- BroadcastExchange HashedRelationBroadcastMode(List(input[0, struct<id:bigint,id2:bigint>, false].id)) +- *(1) Project [named_struct(id, id#90L, id2, id2#91L) AS _2#95] +- *(1) Filter (named_struct(id, id#90L, id2, id2#91L).id > 30) +- *(1) FileScan parquet [id#90L,id2#91L] Batched: true, Format: Parquet, Location: InMemoryFileIndex[file:/Users/henry/src/spark/one_million], PartitionFilters: [], PushedFilters: [], ReadSchema: struct<id:bigint,id2:bigint>
Using joinWith means that the filter is placed on a named_struct, and is then pushed down. When the filter is just above the scan, the wrapping-and-projection of named_struct(id...).id is a no-op and could be removed. Then the filter can be pushed down to Parquet.
