Details
-
Bug
-
Status: Resolved
-
Major
-
Resolution: Fixed
-
3.0.3, 3.1.3, 3.2.3
Description
The following example returns an incorrect result:
import pandas as pd from pyspark.sql import SparkSession, Window from pyspark.sql.functions import col, lit, sum spark = SparkSession \ .builder \ .getOrCreate() ids = 1000 days = 1000 parts = 10 id_df = spark.range(ids) day_df = spark.range(days).withColumnRenamed("id", "day") id_day_df = id_df.join(day_df) left_df = id_day_df.select(col("id").alias("id"), col("day").alias("day"), lit("left").alias("side")).repartition(parts).cache() right_df = id_day_df.select(col("id").alias("id"), col("day").alias("day"), lit("right").alias("side")).repartition(parts).cache() #.withColumnRenamed("id", "id2") # note the column order is different to the groupBy("id", "day") column order below window = Window.partitionBy("day", "id") left_grouped_df = left_df.groupBy("id", "day") right_grouped_df = right_df.withColumn("day_sum", sum(col("day")).over(window)).groupBy("id", "day") def cogroup(left: pd.DataFrame, right: pd.DataFrame) -> pd.DataFrame: return pd.DataFrame([{ "id": left["id"][0] if not left.empty else (right["id"][0] if not right.empty else None), "day": left["day"][0] if not left.empty else (right["day"][0] if not right.empty else None), "lefts": len(left.index), "rights": len(right.index) }]) df = left_grouped_df.cogroup(right_grouped_df) \ .applyInPandas(cogroup, schema="id long, day long, lefts integer, rights integer") df.explain() df.show(5)
Output is
== Physical Plan == AdaptiveSparkPlan isFinalPlan=false +- FlatMapCoGroupsInPandas [id#8L, day#9L], [id#29L, day#30L], cogroup(id#8L, day#9L, side#10, id#29L, day#30L, side#31, day_sum#54L), [id#64L, day#65L, lefts#66, rights#67] :- Sort [id#8L ASC NULLS FIRST, day#9L ASC NULLS FIRST], false, 0 : +- Exchange hashpartitioning(id#8L, day#9L, 200), ENSURE_REQUIREMENTS, [plan_id=117] : +- ... +- Sort [id#29L ASC NULLS FIRST, day#30L ASC NULLS FIRST], false, 0 +- Project [id#29L, day#30L, id#29L, day#30L, side#31, day_sum#54L] +- Window [sum(day#30L) windowspecdefinition(day#30L, id#29L, specifiedwindowframe(RowFrame, unboundedpreceding$(), unboundedfollowing$())) AS day_sum#54L], [day#30L, id#29L] +- Sort [day#30L ASC NULLS FIRST, id#29L ASC NULLS FIRST], false, 0 +- Exchange hashpartitioning(day#30L, id#29L, 200), ENSURE_REQUIREMENTS, [plan_id=112] +- ... +---+---+-----+------+ | id|day|lefts|rights| +---+---+-----+------+ | 0| 3| 0| 1| | 0| 4| 0| 1| | 0| 13| 1| 0| | 0| 27| 0| 1| | 0| 31| 0| 1| +---+---+-----+------+ only showing top 5 rows
The first child is hash-partitioned by id and day, while the second child is hash-partitioned by day and id (required by the window function). Therefore, rows end up in different partitions.
This has been fixed in Spark 3.3 by #32875:
== Physical Plan == AdaptiveSparkPlan isFinalPlan=false +- FlatMapCoGroupsInPandas [id#8L, day#9L], [id#29L, day#30L], cogroup(id#8L, day#9L, side#10, id#29L, day#30L, side#31, day_sum#54L)#63, [id#64L, day#65L, lefts#66, rights#67] :- Sort [id#8L ASC NULLS FIRST, day#9L ASC NULLS FIRST], false, 0 : +- Exchange hashpartitioning(id#8L, day#9L, 200), ENSURE_REQUIREMENTS, [plan_id=117] : +- ... +- Sort [id#29L ASC NULLS FIRST, day#30L ASC NULLS FIRST], false, 0 +- Exchange hashpartitioning(id#29L, day#30L, 200), ENSURE_REQUIREMENTS, [plan_id=118] +- Project [id#29L, day#30L, id#29L, day#30L, side#31, day_sum#54L] +- Window [sum(day#30L) windowspecdefinition(day#30L, id#29L, specifiedwindowframe(RowFrame, unboundedpreceding$(), unboundedfollowing$())) AS day_sum#54L], [day#30L, id#29L] +- Sort [day#30L ASC NULLS FIRST, id#29L ASC NULLS FIRST], false, 0 +- Exchange hashpartitioning(day#30L, id#29L, 200), ENSURE_REQUIREMENTS, [plan_id=112] +- ... +---+---+-----+------+ | id|day|lefts|rights| +---+---+-----+------+ | 0| 13| 1| 1| | 0| 63| 1| 1| | 0| 89| 1| 1| | 0| 95| 1| 1| | 0| 96| 1| 1| +---+---+-----+------+ only showing top 5 rows
Only PySpark is to be affected (FlatMapCoGroupsInPandas }}), as Scala API uses {{CoGroup. FlatMapCoGroupsInPandas reports required child distribution ClusteredDistribution, while CoGroup reports HashClusteredDistribution. The EnsureRequirements rule correctly recognizes a HashClusteredDistribution(id, day) as not compatible with hashpartitioning(day, id), while ClusteredDistribution(id, day) is compatible with hashpartitioning(day, id).
