On Thu, Sep 15, 2011 at 9:21 AM, Jonathan Ellis <jbel...@gmail.com> wrote:
> Where did the data loss come in? > The outcome of the analytical jobs run overnight while some of these repairs were (not) running is consistent with what I would expect if perhaps 20-30% of the source data was missing. Given the strong consistency model we're using, this is surprising to me, since the jobs did not report any read or write failures. I wonder if this is a consequence of the dead node missing and the new node being operational but having received basically none of its hinted handoff streams. Perhaps with streaming fixed the data will reappear, which would be a happy outcome, but if not, I can reimport the critical stuff from files. Scrub is safe to run in parallel. > Is it somewhat analogous to a major compaction in terms of I/O impact, with perhaps less greedy use of disk space? > On Thu, Sep 15, 2011 at 8:08 AM, Ethan Rowe <et...@the-rowes.com> wrote: > > After further review, I'm definitely going to scrub all the original > nodes > > in the cluster. > > We've lost some data as a result of this situation. It can be restored, > but > > the question is what to do with the problematic new node first. I don't > > particularly care about the data that's on it, since I'm going to > re-import > > the critical data from files anyway, and then I can recreate derivative > data > > afterwards. So it's purely a matter of getting the cluster healthy again > as > > quickly as possible so I can begin that import process. > > Any issue with running scrubs on multiple nodes at a time, provided they > > aren't replication neighbors? > > On Thu, Sep 15, 2011 at 8:18 AM, Ethan Rowe <et...@the-rowes.com> wrote: > >> > >> I just noticed the following from one of Jonathan Ellis' messages > >> yesterday: > >>> > >>> Added to NEWS: > >>> > >>> - After upgrading, run nodetool scrub against each node before > running > >>> repair, moving nodes, or adding new ones. > >> > >> > >> We did not do this, as it was not indicated as necessary in the news > when > >> we were dealing with the upgrade. > >> So perhaps I need to scrub everything before going any further, though > the > >> question is what to do with the problematic node. Additionally, it > would be > >> helpful to know if scrub will affect the hinted handoffs that have > >> accumulated, as these seem likely to be part of the set of failing > streams. > >> On Thu, Sep 15, 2011 at 8:13 AM, Ethan Rowe <et...@the-rowes.com> > wrote: > >>> > >>> Here's a typical log slice (not terribly informative, I fear): > >>>> > >>>> INFO [AntiEntropyStage:2] 2011-09-15 05:41:36,106 > >>>> AntiEntropyService.java (l > >>>> ine 884) Performing streaming repair of 1003 ranges with /10.34.90.8for > >>>> (299 > >>>> > >>>> > 90798416657667504332586989223299634,54296681768153272037430773234349600451] > >>>> INFO [AntiEntropyStage:2] 2011-09-15 05:41:36,427 StreamOut.java > (line > >>>> 181) > >>>> Stream context metadata > >>>> [/mnt/cassandra/data/events_production/FitsByShip-g-1 > >>>> 0-Data.db sections=88 progress=0/11707163 - 0%, > >>>> /mnt/cassandra/data/events_pr > >>>> oduction/FitsByShip-g-11-Data.db sections=169 progress=0/6133240 - 0%, > >>>> /mnt/c > >>>> assandra/data/events_production/FitsByShip-g-6-Data.db sections=1 > >>>> progress=0/ > >>>> 6918814 - 0%, > >>>> /mnt/cassandra/data/events_production/FitsByShip-g-12-Data.db s > >>>> ections=260 progress=0/9091780 - 0%], 4 sstables. > >>>> INFO [AntiEntropyStage:2] 2011-09-15 05:41:36,428 > StreamOutSession.java > >>>> (lin > >>>> e 174) Streaming to /10.34.90.8 > >>>> ERROR [Thread-56] 2011-09-15 05:41:38,515 AbstractCassandraDaemon.java > >>>> (line > >>>> 139) Fatal exception in thread Thread[Thread-56,5,main] > >>>> java.lang.NullPointerException > >>>> at > >>>> org.apache.cassandra.net.IncomingTcpConnection.stream(IncomingTcpC > >>>> onnection.java:174) > >>>> at > >>>> org.apache.cassandra.net.IncomingTcpConnection.run(IncomingTcpConn > >>>> ection.java:114) > >>> > >>> Not sure if the exception is related to the outbound streaming above; > >>> other nodes are actively trying to stream to this node, so perhaps it > comes > >>> from those and temporal adjacency to the outbound stream is just > >>> coincidental. I have other snippets that look basically identical to > the > >>> above, except if I look at the logs to which this node is trying to > stream, > >>> I see that it has concurrently opened a stream in the other direction, > which > >>> could be the one that the exception pertains to. > >>> > >>> On Thu, Sep 15, 2011 at 7:41 AM, Sylvain Lebresne < > sylv...@datastax.com> > >>> wrote: > >>>> > >>>> On Thu, Sep 15, 2011 at 1:16 PM, Ethan Rowe <et...@the-rowes.com> > wrote: > >>>> > Hi. > >>>> > > >>>> > We've been running a 7-node cluster with RF 3, QUORUM reads/writes > in > >>>> > our > >>>> > production environment for a few months. It's been consistently > >>>> > stable > >>>> > during this period, particularly once we got out maintenance > strategy > >>>> > fully > >>>> > worked out (per node, one repair a week, one major compaction a > week, > >>>> > the > >>>> > latter due to the nature of our data model and usage). While this > >>>> > cluster > >>>> > started, back in June or so, on the 0.7 series, it's been running > >>>> > 0.8.3 for > >>>> > a while now with no issues. We upgraded to 0.8.5 two days ago, > having > >>>> > tested the upgrade in our staging cluster (with an otherwise > identical > >>>> > configuration) previously and verified that our application's > various > >>>> > use > >>>> > cases appeared successful. > >>>> > > >>>> > One of our nodes suffered a disk failure yesterday. We attempted to > >>>> > replace > >>>> > the dead node by placing a new node at OldNode.initial_token - 1 > with > >>>> > auto_bootstrap on. A few things went awry from there: > >>>> > > >>>> > 1. We never saw the new node in bootstrap mode; it became available > >>>> > pretty > >>>> > much immediately upon joining the ring, and never reported a > "joining" > >>>> > state. I did verify that auto_bootstrap was on. > >>>> > > >>>> > 2. I mistakenly ran repair on the new node rather than removetoken > on > >>>> > the > >>>> > old node, due to a delightful mental error. The repair got nowhere > >>>> > fast, as > >>>> > it attempts to repair against the down node which throws an > exception. > >>>> > So I > >>>> > interrupted the repair, restarted the node to clear any pending > >>>> > validation > >>>> > compactions, and... > >>>> > > >>>> > 3. Ran removetoken for the old node. > >>>> > > >>>> > 4. We let this run for some time and saw eventually that all the > nodes > >>>> > appeared to be done various compactions and were stuck at streaming. > >>>> > Many > >>>> > streams listed as open, none making any progress. > >>>> > > >>>> > 5. I observed an Rpc-related exception on the new node (where the > >>>> > removetoken was launched) and concluded that the streams were broken > >>>> > so the > >>>> > process wouldn't ever finish. > >>>> > > >>>> > 6. Ran a "removetoken force" to get the dead node out of the mix. > No > >>>> > problems. > >>>> > > >>>> > 7. Ran a repair on the new node. > >>>> > > >>>> > 8. Validations ran, streams opened up, and again things got stuck in > >>>> > streaming, hanging for over an hour with no progress. > >>>> > > >>>> > 9. Musing that lingering tasks from the removetoken could be a > factor, > >>>> > I > >>>> > performed a rolling restart and attempted a repair again. > >>>> > > >>>> > 10. Same problem. Did another rolling restart and attempted a fresh > >>>> > repair > >>>> > on the most important column family alone. > >>>> > > >>>> > 11. Same problem. Streams included CFs not specified, so I guess > they > >>>> > must > >>>> > be for hinted handoff. > >>>> > > >>>> > In concluding that streaming is stuck, I've observed: > >>>> > - streams will be open to the new node from other nodes, but the new > >>>> > node > >>>> > doesn't list them > >>>> > - streams will be open to the other nodes from the new node, but the > >>>> > other > >>>> > nodes don't list them > >>>> > - the streams reported may make some initial progress, but then they > >>>> > hang at > >>>> > a particular point and do not move on for an hour or more. > >>>> > - The logs report repair-related activity, until NPEs on incoming > TCP > >>>> > connections show up, which appear likely to be the culprit. > >>>> > >>>> Can you send the stack trace from those NPE. > >>>> > >>>> > > >>>> > I can provide more exact details when I'm done commuting. > >>>> > > >>>> > With streaming broken on this node, I'm unable to run repairs, which > >>>> > is > >>>> > obviously problematic. The application didn't suffer any > operational > >>>> > issues > >>>> > as a consequence of this, but I need to review the overnight results > >>>> > to > >>>> > verify we're not suffering data loss (I doubt we are). > >>>> > > >>>> > At this point, I'm considering a couple options: > >>>> > 1. Remove the new node and let the adjacent node take over its range > >>>> > 2. Bring the new node down, add a new one in front of it, and > properly > >>>> > removetoken the problematic one. > >>>> > 3. Bring the new node down, remove all its data except for the > system > >>>> > keyspace, then bring it back up and repair it. > >>>> > 4. Revert to 0.8.3 and see if that helps. > >>>> > > >>>> > Recommendations? > >>>> > > >>>> > Thanks. > >>>> > - Ethan > >>>> > > >>> > >> > > > > > > > > -- > Jonathan Ellis > Project Chair, Apache Cassandra > co-founder of DataStax, the source for professional Cassandra support > http://www.datastax.com >
