cassandra-commits mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Alexander Piavlo (JIRA)" <j...@apache.org>
Subject [jira] [Updated] (CASSANDRA-9720) half open tcp connections to cassandra cluster nodes cause 100% cpu load
Date Thu, 02 Jul 2015 20:44:05 GMT

     [ https://issues.apache.org/jira/browse/CASSANDRA-9720?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]

Alexander Piavlo updated CASSANDRA-9720:
----------------------------------------
    Description: 
cassandra 2.1.5

We spotted that few of the nodes in our cluster got sudden cpu 100% spike which never ended.
It's not a GC not increased reads/writes nodes.
What we saw is that those nodes that have 100% cpu load
all have some connections (file descriptios) with "can't identify protocol"
which indicate those must be unprolery handled abrupt connections by cassandra process.
http://stackoverflow.com/questions/7911840/seeing-too-many-lsof-cant-identify-protocol

We are pretty sure what triggered this is the spark cassandra connector
which sudenly started to get stuck in early discovery of cassandra nodes before running any
stages

We had to restart the affected cassandra processes to get the cpu back to normal

ps. we had similar issues some time ago with earlier version, of 2.1.x cassandra branch, and
ended up solving the problerm by upgrading from spark1.2.1 to spark1.3.1 and also upgrading
spark datastax connecor accordingly. Now looks like the problem is back with 99.9% same symptoms

ps2. We have observed previously several java/cassandra unrelated processes (mainly in php-cli)
go crazy with cpu then they had "can't identify protocol" symphoms

  was:
cassandra 2.1.5

We spotted that few of the nodes in our cluster got sudden cpu 100% spike which never ended.
It's not a GC not increased reads/writes nodes.
What we saw is that those nodes that have 100% cpu load
all have some connections (file descriptios) with "can't identify protocol"
which indicate those must be unprolery handled abrupt connections by cassandra process.
http://stackoverflow.com/questions/7911840/seeing-too-many-lsof-cant-identify-protocol

We are pretty sure what triggered this is the spark cassandra connector
which sudenly started to get stuck in early discovery of cassandra nodes before running any
stages

ps. we had similar issues some time ago with earlier version, of 2.1.x cassandra branch, and
ended up solving the problerm by upgrading from spark1.2.1 to spark1.3.1 and also upgrading
spark datastax connecor accordingly. Now looks like the problem is back with 99.9% same symptoms

ps2. We have observed previously several java/cassandra unrelated processes (mainly in php-cli)
go crazy with cpu then they had "can't identify protocol" symphoms


> half open tcp connections to cassandra cluster nodes cause 100% cpu load
> ------------------------------------------------------------------------
>
>                 Key: CASSANDRA-9720
>                 URL: https://issues.apache.org/jira/browse/CASSANDRA-9720
>             Project: Cassandra
>          Issue Type: Bug
>          Components: Core
>            Reporter: Alexander Piavlo
>
> cassandra 2.1.5
> We spotted that few of the nodes in our cluster got sudden cpu 100% spike which never
ended. It's not a GC not increased reads/writes nodes.
> What we saw is that those nodes that have 100% cpu load
> all have some connections (file descriptios) with "can't identify protocol"
> which indicate those must be unprolery handled abrupt connections by cassandra process.
> http://stackoverflow.com/questions/7911840/seeing-too-many-lsof-cant-identify-protocol
> We are pretty sure what triggered this is the spark cassandra connector
> which sudenly started to get stuck in early discovery of cassandra nodes before running
any stages
> We had to restart the affected cassandra processes to get the cpu back to normal
> ps. we had similar issues some time ago with earlier version, of 2.1.x cassandra branch,
and ended up solving the problerm by upgrading from spark1.2.1 to spark1.3.1 and also upgrading
spark datastax connecor accordingly. Now looks like the problem is back with 99.9% same symptoms
> ps2. We have observed previously several java/cassandra unrelated processes (mainly in
php-cli) go crazy with cpu then they had "can't identify protocol" symphoms



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

Mime
View raw message