Yeah. the heap is huge, need to optimize the caches. It was 8Gb previously,
had to increase because there were out of memory errors. Using
ConcMarkSweepGC, which is supposed to not lock the world.
Had to disable optimize (previously we did so by a cron task) because the
index is big and optimize has bad impact on performance and resources usage.
We're using auto and soft commits only:
<autoCommit>
<maxDocs>25000</maxDocs>
<maxTime>300000</maxTime>
<openSearcher>false</openSearcher>
</autoCommit>
<autoSoftCommit>
<maxTime>10000</maxTime>
</autoSoftCommit>
I was thinking we may reach some system limits, but netstat doesn't show
anything suspicious:
alex@solr1:~$ netstat -an|awk '/tcp/ {print $6}'|sort|uniq -c
137 CLOSE_WAIT
24 ESTABLISHED
9 LISTEN
77 TIME_WAIT
It is also not clear where did those errors happen. It would be useful for
users (I mean for those not familiar with Solr development) if instead of
putting the entire backtrace (or in addition to it) Solr logged user
readable messages. Like: "Error while sending response to the client
#{client_ip:client_port}", or "when sending updates to replica
#{replica_ip:replica_port}". Because now those errors are pretty confusing.
Best,
Alex
--
View this message in context:
http://lucene.472066.n3.nabble.com/java-net-SocketException-Connection-reset-tp4145519p4145894.html
Sent from the Solr - User mailing list archive at Nabble.com.
