I see the behavior - so it always goes with min total tasks possible on your settings ( num-executors * num-cores ) - however if you use a huge amount of data then you will see more tasks - that means it has some kind of lower bound on num-tasks.. It may require some digging. other formats did not seem to have this issue.
On Sun, May 8, 2016 at 12:10 AM, Johnny W. <jzw.ser...@gmail.com> wrote: > The file size is very small (< 1M). The stage launches every time i call: > -- > sqlContext.read.parquet(path_to_file) > > These are the parquet specific configurations I set: > -- > spark.sql.parquet.filterPushdown: true > spark.sql.parquet.mergeSchema: true > > Thanks, > J. > > On Sat, May 7, 2016 at 4:20 PM, Ashish Dubey <ashish....@gmail.com> wrote: > >> How big is your file and can you also share the code snippet >> >> >> On Saturday, May 7, 2016, Johnny W. <jzw.ser...@gmail.com> wrote: >> >>> hi spark-user, >>> >>> I am using Spark 1.6.0. When I call sqlCtx.read.parquet to create a >>> dataframe from a parquet data source with a single parquet file, it yields >>> a stage with lots of small tasks. It seems the number of tasks depends on >>> how many executors I have instead of how many parquet files/partitions I >>> have. Actually, it launches 5 tasks on each executor. >>> >>> This behavior is quite strange, and may have potential issue if there is >>> a slow executor. What is this "parquet" stage for? and why it launches 5 >>> tasks on each executor? >>> >>> Thanks, >>> J. >>> >> >