spark-user mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From Venkat Subramanian <>
Subject Spark SQL table Join, one task is taking long
Date Mon, 01 Dec 2014 22:35:14 GMT
Environment: Spark 1.1, 4 Node Spark and Hadoop Dev cluster - 6 cores, 32 GB
Ram each. Default serialization, Standalone, no security

Data was sqooped from relational DB to HDFS and Data is partitioned across
HDFS uniformly. I am reading a  fact table about 8 GB in size and one small
dim table from HDFS and then doing a join on them based on a criteria. .
Running the Driver on Spark shell on Spark master.

ContactDetail and DAgents are read as RDD and registered as table already.
Each of these tables have 60 to 90 fields and I am using Product class.

val CDJoinQry= sqlContext.sql("SELECT  * FROM ContactDetail, DAgents  WHERE
ContactDetail.f6 = DAgents.f1 and DAgents.f1 = 902") => ta(4)).count   // result is a small number

This works fine and returns the result fine. Hadoop mapPartition reads and
creation of RDDs are all fine But in the Count stage, I see that one of 
task (out of 200 ) does a huge amount of Shuffle Write (some 1 GB or more)
and takes about 1.1 seconds to complete out of the 1.2 seconds of total
execution time. This task is usually around in the 3/4 th (say 160/200) of
the total tasks. At the time of that task running, one of the CPU in one
worker node goes to 100% for the duration of the task. Rest of the tasks
take few ms and does only < 5 MBs of Shuffle write.  I have run it
repeatedly and this happens regardless of which worker node this particular
task is running on. I turned on Spark debug on all nodes to understand, but
it was difficult to figure out where the delay is from the logs. There are
no errors or re-trys in the logs. 

Not sure if I can post logs here for someone to look at, if so I can (about
10 Mb). Also, not sure if this normal in such a table join that one task
would take most amount of time. Let me know if you have any suggestions.



View this message in context:
Sent from the Apache Spark User List mailing list archive at

To unsubscribe, e-mail:
For additional commands, e-mail:

View raw message