[ 
https://issues.apache.org/jira/browse/CASSANDRA-13663?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16115628#comment-16115628
 ] 

Jeff Jirsa commented on CASSANDRA-13663:
----------------------------------------

It's likely being killed by the system on-killer - you'll probably see traces 
in smear

Cassandra allocates a fixed amount of memory for the heap, and a variable 
amount of RAM offheap for things like bloom filters and compression offsets

Both bloom filters and compression offsets scale linearly with disk space (and 
can be much higher with various table specific settings), and are allocated in 
chunks during compaction.

It's very likely you're seeing a compaction start, and the bloom filter 
allocation triggers the Pom-killer

How much ram is in the system? How big is the heap? How much data is on each 
node?

> Cassandra 3.10 crashes without dump
> -----------------------------------
>
>                 Key: CASSANDRA-13663
>                 URL: https://issues.apache.org/jira/browse/CASSANDRA-13663
>             Project: Cassandra
>          Issue Type: Bug
>            Reporter: Matthias Otto
>            Priority: Minor
>         Attachments: 2017-07-04 10_48_34-CloudWatch Management Console.png, 
> cassandra debug.log, cassandra system.log, RamUsageExamle1.png, 
> RamUsageExample2.png
>
>
> Hello. My company runs a 5 node Cassandra cluster. For the last few weeks, we 
> have had a sporadic issue where one of the servers crashes without creating a 
> dump file and without any error messages in the logs. If one restarts the 
> service (which we have by now scripted to happen automatically), the servers 
> resumes work with no complaint.
> Log files of the time of the last crash are attached, thou again they do not 
> log any crash happening.
> Regarding out setup, we are running these servers on AMazon AWS, with 3 
> volumes per server, one for the system, one for data and one for the 
> commitlog. When a crash happens, we can observe a sudden spike of read 
> activity on the commitlog volume. All of these have ample free space. 
> Aspecially the system volume has more then enough free space so that a dump 
> could be written.
> The servers are Ubuntu 16.04 servers and Cassandra is installed from the 
> apt-get packet for version 3.10.
> It is worth noting that these crashes happen more often when nodetool is 
> running either repair job or a backup job, but this is by no means always the 
> case. As for frequency, we have had about 1-2 crashes per week for the last 
> month.



--
This message was sent by Atlassian JIRA
(v6.4.14#64029)

---------------------------------------------------------------------
To unsubscribe, e-mail: commits-unsubscr...@cassandra.apache.org
For additional commands, e-mail: commits-h...@cassandra.apache.org

Reply via email to