> I was caching it because I didn't want to re-execute the DAG when I > ran the count query. If you have a spark application with multiple > actions, Spark reexecutes the entire DAG for each action unless there > is a cache in between. I was trying to avoid reloading 1/2 a terabyte > of data. Also, cache should use up executor memory, not driver > memory. why not counting the parquet file instead? writing/reading a parquet files is more efficients than caching in my experience. if you really need caching you could choose a better strategy such DISK.
Lalwani, Jayesh <jlalw...@amazon.com.INVALID> writes: > I was caching it because I didn't want to re-execute the DAG when I ran the > count query. If you have a spark application with multiple actions, Spark > reexecutes the entire DAG for each action unless there is a cache in between. > I was trying to avoid reloading 1/2 a terabyte of data. Also, cache should > use up executor memory, not driver memory. > > As it turns out cache was the problem. I didn't expect cache to take Executor > memory and spill over to disk. I don't know why it's taking driver memory. > The input data has millions of partitions which results in millions of tasks. > Perhaps the high memory usage is a side effect of caching the results of lots > of tasks. > > On 10/19/20, 1:27 PM, "Nicolas Paris" <nicolas.pa...@riseup.net> wrote: > > CAUTION: This email originated from outside of the organization. Do not > click links or open attachments unless you can confirm the sender and know > the content is safe. > > > > > Before I write the data frame to parquet, I do df.cache. After writing > > the file out, I do df.countDistinct(“a”, “b”, “c”).collect() > if you write the df to parquet, why would you also cache it ? caching by > default loads the memory. this might affect later use, such > collect. the resulting GC can be explained by both caching and collect > > > Lalwani, Jayesh <jlalw...@amazon.com.INVALID> writes: > > > I have a Dataframe with around 6 billion rows, and about 20 columns. > First of all, I want to write this dataframe out to parquet. The, Out of the > 20 columns, I have 3 columns of interest, and I want to find how many > distinct values of the columns are there in the file. I don’t need the actual > distinct values. I just need the count. I knoe that there are around > 10-16million distinct values > > > > Before I write the data frame to parquet, I do df.cache. After writing > the file out, I do df.countDistinct(“a”, “b”, “c”).collect() > > > > When I run this, I see that the memory usage on my driver steadily > increases until it starts getting future time outs. I guess it’s spending > time in GC. Does countDistinct cause this behavior? Does Spark try to get all > 10 million distinct values into the driver? Is countDistinct not recommended > for data frames with large number of distinct values? > > > > What’s the solution? Should I use approx._count_distinct? > > > -- > nicolas paris > > --------------------------------------------------------------------- > To unsubscribe e-mail: user-unsubscr...@spark.apache.org > > > > --------------------------------------------------------------------- > To unsubscribe e-mail: user-unsubscr...@spark.apache.org -- nicolas paris --------------------------------------------------------------------- To unsubscribe e-mail: user-unsubscr...@spark.apache.org