Hi Shyam Thanks for your reply. You mean after knowing the partition number of column_a, column_b, column_c, the sequence of column in partitionBy should be same to the order of partitions number of column a, b and c? But the sequence of columns in partitionBy decides the directory hierarchy structure. I hope the sequence of columns not change.
And I found one more strange things, some tasks in write hdfs stage cost much more time than others, where the amount of writing data is similar. How to solve it? Regard, Junfeng Chen On Tue, Mar 5, 2019 at 3:05 PM Shyam P <shyamabigd...@gmail.com> wrote: > Hi JF , > Try to execute it before df.write.... > > //count by partition_id > import org.apache.spark.sql.functions.spark_partition_id > df.groupBy(spark_partition_id).count.show() > > You will come to know how data has been partitioned inside df. > > Small trick we can apply here while partitionBy(column_a, column_b, > column_c) > Makes sure you should have ( column_a partitions) > ( column_b > partitions) > ( column_c partitions) . > > Try this. > > Regards, > Shyam > > On Mon, Mar 4, 2019 at 4:09 PM JF Chen <darou...@gmail.com> wrote: > >> I am trying to write data in dataset to hdfs via >> df.write.partitionBy(column_a, >> column_b, column_c).parquet(output_path) >> However, it costs several minutes to write only hundreds of MB data to >> hdfs. >> From this article >> <https://stackoverflow.com/questions/45269658/spark-df-write-partitionby-run-very-slow>, >> adding repartition method before write should work. But if there is data >> skew, some tasks may cost much longer time than average, which still cost >> much time. >> How to solve this problem? Thanks in advance ! >> >> >> Regard, >> Junfeng Chen >> >