[ https://issues.apache.org/jira/browse/CASSANDRA-7567?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14081040#comment-14081040 ]
Brandon Williams commented on CASSANDRA-7567: --------------------------------------------- I applied CASSANDRA-7644 and finally got meaningful traces, but none were fruitful - everything always completed quickly and exactly as it should. So I went for the sure-fire method of reproducing, and suspended the jvm on the 'dd' node so it could respond at all. Still nothing suspicious in the trace, however now I _did_ see stress report 12s latencies. Suspecting that stress wasn't actually doing what I told it to (only connect to node1 and node2, not node3 which I'm trying to beat up with dd) I discovered it was actually connecting to all three nodes, and that's why suspending the 'unused' node caused the latencies. I don't think anything is wrong in Cassandra itself here any longer, but something is wrong with stress, probably doing ring discovery and connecting to everything when it shouldn't. > when the commit_log disk for a single node is overwhelmed the entire cluster > slows down > --------------------------------------------------------------------------------------- > > Key: CASSANDRA-7567 > URL: https://issues.apache.org/jira/browse/CASSANDRA-7567 > Project: Cassandra > Issue Type: Bug > Components: Core > Environment: debian 7.5, bare metal, 14 nodes, 64CPUs, 64GB RAM, > commit_log disk sata, data disk SSD, vnodes, leveled compaction strategy > Reporter: David O'Dell > Assignee: Brandon Williams > Fix For: 2.1.0 > > Attachments: 7567.logs.bz2, write_request_latency.png > > > We've run into a situation where a single node out of 14 is experiencing high > disk io. This can happen when a node is being decommissioned or after it > joins the ring and runs into the bug cassandra-6621. > When this occurs the write latency for the entire cluster spikes. > From 0.3ms to 170ms. > To simulate this simply run dd on the commit_log disk (dd if=/dev/zero > of=/tmp/foo bs=1024) and you will see that instantly all nodes in the cluster > have slowed down. > BTW overwhelming the data disk does not have this same effect. > Also I've tried this where the overwhelmed node isn't being connected > directly from the client and it still has the same effect. -- This message was sent by Atlassian JIRA (v6.2#6252)