[CASSANDRA-6651] Repair hanging

• Type: Bug
• Status: Resolved
• Priority: Normal
• Resolution: Duplicate
• Severity: Normal
• Assignee: Yuki Morishita
• Reporter: Eitan Eibschutz

Authors

Yuki Morishita

• Created: 05/Feb/14 05:43
• Updated: 16/Apr/19 09:31
• Resolved: 29/Jul/14 22:12

Description

Hi,

We have a 12 node cluster in PROD environment and we've noticed that repairs are never finishing. The behavior that we've observed is that a repair process will run until at some point it hangs and no other processing is happening.

For example, at the moment, I have a repair process that has been running for two days and not finishing:

nodetool tpstats is showing 2 active and 2 pending AntiEntropySessions

nodetool compactionstats is showing:
pending tasks: 0
Active compaction remaining time : n/a

nodetools netstats is showing:
Mode: NORMAL
Not sending any streams.
Read Repair Statistics:
Attempted: 0
Mismatch (Blocking): 142110
Mismatch (Background): 0
Pool Name Active Pending Completed
Commands n/a 0 107589657
Responses n/a 0 116430785

The last entry that I see in the log is:
INFO [AntiEntropySessions:18] 2014-02-03 04:01:39,145 RepairJob.java (line 116) repair #ae78c6c0-8c2b-11e3-b950-c3b81a36bc9b requesting merkle trees for MyCF (to [/x.x.x.x, /y.y.y.y, /z.z.z.z])

The repair started at 4am so it stopped after 1:40 minute.

On node y.y.y.y I can see this in the log:
INFO [MiscStage:1] 2014-02-03 04:01:38,985 ColumnFamilyStore.java (line 740) Enqueuing flush of Memtable-MyCF@1290890489(2176/5931 serialized/live bytes, 32 ops)
INFO [FlushWriter:411] 2014-02-03 04:01:38,986 Memtable.java (line 333) Writing Memtable-MyCF@1290890489(2176/5931 serialized/live bytes, 32 ops)
INFO [FlushWriter:411] 2014-02-03 04:01:39,048 Memtable.java (line 373) Completed flushing /var/lib/cassandra/main-db/data/MyKS/MyCF/MyKS-MyCF-jb-518-Data.db (1789 bytes) for commitlog position ReplayPosition(segmentId=1390437013339, position=21868792)
INFO [ScheduledTasks:1] 2014-02-03 05:00:04,794 ColumnFamilyStore.java (line 740) Enqueuing flush of Memtable-compaction_history@1649414699(1635/17360 serialized/live bytes, 42 ops)

So for some reason the merkle tree for this CF is never sent back to the node being repaired and it's hanging.

I've also noticed that sometimes, restarting node y.y.y.y will cause the repair to resume.

Another observation is that sometimes when restarting y.y.y.y it will not start with these errors:
ERROR 16:34:18,485 Exception encountered during startup
java.lang.IllegalStateException: Unfinished compactions reference missing sstables. This should never happen since compactions are marked finished before we start removing the old sstables.
at org.apache.cassandra.db.ColumnFamilyStore.removeUnfinishedCompactionLeftovers(ColumnFamilyStore.java:495)
at org.apache.cassandra.service.CassandraDaemon.setup(CassandraDaemon.java:264)
at org.apache.cassandra.service.CassandraDaemon.activate(CassandraDaemon.java:461)
at org.apache.cassandra.service.CassandraDaemon.main(CassandraDaemon.java:504)
java.lang.IllegalStateException: Unfinished compactions reference missing sstables. This should never happen since compactions are marked finished before we start removing the old sstables.
at org.apache.cassandra.db.ColumnFamilyStore.removeUnfinishedCompactionLeftovers(ColumnFamilyStore.java:495)
at org.apache.cassandra.service.CassandraDaemon.setup(CassandraDaemon.java:264)
at org.apache.cassandra.service.CassandraDaemon.activate(CassandraDaemon.java:461)
at org.apache.cassandra.service.CassandraDaemon.main(CassandraDaemon.java:504)
Exception encountered during startup: Unfinished compactions reference missing sstables. This should never happen since compactions are marked finished before we start removing the old sstables.

And it will only restart after manually cleaning the compactions_in-progress folder.

I'm not sure if these two issues are related but we've seen both on all the nodes in our cluster.

I'll be happy to provide more info if needed as we are not sure what could cause this behavior.

Another thing in our environment is that some of the Cassandra nodes have more than one network interface and RPC is listening on 0.0.0.0, not sure if it has anything to do with this.

Thanks,
Eitan

Issue Links

• duplicates: CASSANDRA-7560 'nodetool repair -pr' leads to indefinitely hanging AntiEntropySession
• is related to: CASSANDRA-7904 Repair hangs