[
https://issues.apache.org/jira/browse/SPARK-27648?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16836241#comment-16836241
]
tommy duan commented on SPARK-27648:
------------------------------------
[~gsomogyi] yes,The agg opreation is used in my program
{code:java}
var streamDs=spark.readStream.format("kafka")...load()
streamDs.withWaterMark("timestamp",120
minutes).createOrReplaceTempView("tv_abc")
var streamAgDs=spark.sql(s"""
|select a,b,c,sum(e) as sum_e,sum(xx) as sum_xx,....,avg(ss) as
avg_ss,...,udf_max(h1,h2,h3) as max_h,..
|from tv_abc
|group by window(timestamp,'$windowSize$ minutes','$slide$ minutes'),a,b,c
""".stripMargin)
{code}
> In Spark2.4 Structured Streaming：The executor storage memory increasing over
> time
> ---------------------------------------------------------------------------------
>
> Key: SPARK-27648
> URL: https://issues.apache.org/jira/browse/SPARK-27648
> Project: Spark
> Issue Type: Bug
> Components: Structured Streaming
> Affects Versions: 2.4.0
> Reporter: tommy duan
> Priority: Major
>
> *Spark Program Code Business:*
> Read the topic on kafka, aggregate the stream data sources, and then output
> it to another topic line of kafka.
> *Problem Description:*
> *1) Using spark structured streaming in CDH environment (spark 2.2)*, memory
> overflow problems often occur (because of too many versions of state stored
> in memory, this bug has been modified in spark 2.4).
> {code:java}
> /spark-submit \
> --conf “spark.yarn.executor.memoryOverhead=4096M”
> --num-executors 15 \
> --executor-memory 3G \
> --executor-cores 2 \
> --driver-memory 6G \{code}
> {code}
> Executor memory exceptions occur when running with this submit resource under
> SPARK 2.2 and the normal running time does not exceed one day.
> The solution is to set the executor memory larger than before
> {code:java}
> My spark-submit script is as follows:
> /spark-submit\
> conf "spark. yarn. executor. memoryOverhead = 4096M"
> num-executors 15\
> executor-memory 46G\
> executor-cores 3\
> driver-memory 6G\
> ...{code}
> In this case, the spark program can be guaranteed to run stably for a long
> time, and the executor storage memory is less than 10M (it has been running
> stably for more than 20 days).
> *2) From the upgrade information of Spark 2.4, we can see that the problem of
> large memory consumption of state storage has been solved in Spark 2.4.*
> So we upgraded spark to SPARK 2.4 under CDH, tried to run the spark program,
> and found that the use of memory was reduced.
> But a problem arises, as the running time increases, the storage memory of
> executor is growing (see Executors - > Storage Memory from the Spark on Yarn
> Resource Manager UI).
> This program has been running for 14 days (under SPARK 2.2, running with
> this submit resource, the normal running time is not more than one day,
> Executor memory abnormalities will occur).
> The script submitted by the program under spark2.4 is as follows:
> {code:java}
> /spark-submit \
> --conf “spark.yarn.executor.memoryOverhead=4096M”
> --num-executors 15 \
> --executor-memory 3G \
> --executor-cores 2 \
> --driver-memory 6G
> {code}
> Under Spark 2.4, I counted the size of executor memory as time went by during
> the running of the spark program:
> |Run-time(hour)|Storage Memory size（MB）|Memory growth rate(MB/hour)|
> |23.5H|41.6MB/1.5GB|1.770212766|
> |108.4H|460.2MB/1.5GB|4.245387454|
> |131.7H|559.1MB/1.5GB|4.245254366|
> |135.4H|575MB/1.5GB|4.246676514|
> |153.6H|641.2MB/1.5GB|4.174479167|
> |219H|888.1MB/1.5GB|4.055251142|
> |263H|1126.4MB/1.5GB|4.282889734|
> |309H|1228.8MB/1.5GB|3.976699029|
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
