[
https://issues.apache.org/jira/browse/CASSANDRA-10194?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Benedict updated CASSANDRA-10194:
---------------------------------
Attachment: aggregated thread dump
Attaching a cleaned up version of the thread dump.
I cannot say exactly why the following is happening, but I can say that it is:
* An sstablereader is being released; during release it persists its read
meter, and this is a durable write that requires commit log space
* The commit log segment manager is not allocating one, and is awaiting work. I
_assume_ this means reserve segments are not yet enabled, i.e. it is too early
in startup to perform a durable write. However, I cannot see
{{CassandraDaemon}} anywhere in the log, so it's possible there is some other
reason they cannot be allocated
* An sstable for the same keyspace is being flushed, and this
Full system logs, and perhaps a heap dump, would aid greatly in pinning down
the cause, however we should really make ourselves at least robust to this
deadlock by, for instance, throwing an exception if we attempt to write a
durable mutation while the commit log is disabled.
It's also possible this is another manifestation of the Linux FUTEX_WAIT bug.
If you could rule out your system being affected by this, it would help a great
deal.
> Deadlock on startup
> -------------------
>
> Key: CASSANDRA-10194
> URL: https://issues.apache.org/jira/browse/CASSANDRA-10194
> Project: Cassandra
> Issue Type: Bug
> Reporter: Jeff Jirsa
> Attachments: aggregated thread dump, cassandra-thread
>
>
> Possible deadlock on startup. Seen on 2.1.5 and 2.1.7. Thread dump attached.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
