[
https://issues.apache.org/jira/browse/CASSANDRA-11082?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15143809#comment-15143809
]
fujian edited comment on CASSANDRA-11082 at 2/12/16 1:37 AM:
-------------------------------------------------------------
I check the netty's code, only the message be sent can decrease the water mark,
so if the message response is big such as > 64M. we can't make one simple
decision to set one low mark so that drop it. So it will involved more
discussion.
according to my understand. the datastax driver's: one message's max size is
256M, and cassadra is 2G. so it is hard to set the low/high water mark. I don't
know is there any other limit can be used?
so involved another question, if one response's column is 2G, how to avoid OOM
when executing select before sent to network.
was (Author: fujian1115):
I check the netty's code, only the message be sent can decrease the water mark,
so if the message response is big such as > 64M. we can't make one simple
decision to set one low mark so that drop it. So it will involved more
discussion.
according to my understand. the datastax driver's: one message's max size is
256M, and cassadra is 2G. so it is hard to set the low/high water mark. I don't
know is there any other limit can be used?
so involved another question, if one column is 2G, how to avoid OOM before sent
to network.
> netty's level hadn't prevent OOM when receiver handle slow.
> -----------------------------------------------------------
>
> Key: CASSANDRA-11082
> URL: https://issues.apache.org/jira/browse/CASSANDRA-11082
> Project: Cassandra
> Issue Type: Bug
> Components: Streaming and Messaging
> Reporter: fujian
>
> as we know, netty will OOM when received client is slow.
> due to if receiver can't handle response fast so that cassandra server can't
> flush data. it will cause channeloutbuffer with big size.
> we see the cassandra had configure write high/low water level, but I can't
> found any code to judge iswritable. so it will have possible OOM
> why cassandra hadn't handle this case?
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
