Open IanMeta opened 4 months ago
Please try 2.1.2 because there is an known memory leak in 2.1.1 during aggregation.
Please try 2.1.2 because there is an known memory leak in 2.1.1 during aggregation.
We just upgraded to Doris 2.1.2 and still observed this trend (see image below).
Previously, we used version 2.0.0 and memory was never above 60% at idle. Is it intended to use more RAM or is this still a memory leak? Are there any configuration options we can try?
@yiguolei I have the same issue. version: 2.1.2
Pasting my BE config java options here, just in case it's relevant.
JAVA_OPTS="-Xmx4096m -XX:-UseGCOverheadLimit -DlogPath=$DORIS_HOME/log/jni.log -Xloggc:$DORIS_HOME/log/be.gc.log.$CUR_DATE -Djavax.security.auth.useSubjectCredsOnly=false -Dsun.security.krb5.debug=true -Dsun.java.command=DorisBE -XX:-CriticalJNINatives -DJDBC_MIN_POOL=1 -DJDBC_MAX_POOL=100 -DJDBC_MAX_IDLE_TIME=300000 -DJDBC_MAX_WAIT_TIME=5000"
@yiguolei I have the same issue. version: 2.1.2
me too
me too. version 2.1.3 version 2.1.4
This bug should be fixed in 2.1.4. If you find memory leak still exists in 2.1.4, please provide more information.
Hi @Gabriel39, My case is different from this issue. And I create a pr to resolve my issue, please help to review it: https://github.com/apache/doris/pull/36966
Hi @kingsylin Have you upgraded to 2.1.4? Has the problem improved?
Doris version 2.1.5 is released Have anyone test the new version to see if the problem is resolved?
Has the problem improved
Yes, I have upgraded to 2.1.4, but problem is still there
Search before asking
Version
2.1.1
What's Wrong?
The memory usage in BE nodes is not going down despite having no queries, eventually causing OOM errors.
According to https://doris.apache.org/docs/admin-manual/maint-monitor/memory-management/memory-tracker, the process memory should be the sum of other types of memory, but we see from mem_tracker that there is 20+GB of memory unaccounted for comparing between process resident memory and sum of other types of memory (can anyone explain how to interpret this?)
From our Grafana monitoring BE statistics , we can see that some memory is released after a burst of queries, but the RAM usage returns to a level slightly higher than before, and over time it trends upwards until it runs out of memory.
What You Expected?
Memory levels should return to normal when there are no queries.
How to Reproduce?
We currently do not have the server resources to reproduce the exact setup, but I believe the issue can be produced by doing some stress tests in a production environment in Doris 2.1.1.
Anything Else?
We also tried setting memory_mode=compact in be.conf, but the problem persists. Currently, we restart the BE nodes periodically to clear previously unreleased cache.
Are you willing to submit PR?
Code of Conduct