For the error, you can see http://www.scriptscoop.net/t/3bac9a3307ac/cassandra-lost-notification-from-nodetool-repair.html
Lost notification should not be a problem.please see https://issues.apache.org/jira/browse/CASSANDRA-7909 Infact, we are also currently facing an issue where merkle tree is not received from one or more nodes in remote dc and repair hangs for ever. We would be turning on debug logging as some important TCP messages are being logged at debug. Also we would be monitoring netstats and tcpdump while repair is running. You can try similar things to troubleshoot. May be more experieced guys can comment on this to help u :) Thanks Anuj Sent from Yahoo Mail on Android From:"Badrjan" <badr...@tuta.io> Date:Sun, 15 Nov, 2015 at 6:14 pm Subject:Re: Repair time comparison for Cassandra 2.1.11 Repairs are parallel. The only error-ish message I see in the log of nodetool is "Lost notification. You should check server log for repair status of keyspace" During the repair most of the time was spent in the process of waiting for merkel tree from other nodes. I checked, the streaming was not the issue. So apparently the issue is somewhere in the part of merkle tree generation during validation, and most probably the part where disk is being read. Sequential repairs are off, so no anticompaction is being done. B. 15. Nov 2015 16:22 by anujw_2...@yahoo.co.in: Ok. I dont have much experience with 2.1 as we are on 2.0.x. Are you using sequential repair? If yes, parallel repair can be faster but you need to make sure that your application has sufficient room to run when cluster is running repair. Are you observing any WARN or ERROR messages in logs while repair is running? 50 hours seems too much considering your cluster is stable and you dont have any dropped mutations on any of the nodes. Thanks Anuj Sent from Yahoo Mail on Android From:"Badrjan" <badr...@tuta.io> Date:Sun, 15 Nov, 2015 at 5:39 pm Subject:Re: Repair time comparison for Cassandra 2.1.11 Nothing is being dropped plus the processor is busy around 60%. B. 15. Nov 2015 15:58 by anujw_2...@yahoo.co.in: Repair can take long time if you have lota of inconaistent data. If you havent restarted nodes yet, you can run nodetool tpstats command on all nodes to make sure that there no mutation drops. Thanks Anuj Sent from Yahoo Mail on Android From:"badr...@tuta.io" <badr...@tuta.io> Date:Sun, 15 Nov, 2015 at 4:20 pm Subject:Repair time comparison for Cassandra 2.1.11 Hi, I have cluster of 4 machines With Cassandra 2.1.11, SSD drives, 600 gb data on each node (replication factor 3). When I run partial repair on one node, it takes 50 hours to finish. Is that normal? B.
