[CASSANDRA-12860] Nodetool repair fragile: cannot properly recover from single node failure. Has to restart all nodes in order to repair again

• Type: Bug
• Status: Resolved
• Priority: Urgent
• Resolution: Duplicate

Environment

CentOS 6.7, Java HotSpot(TM) 64-Bit Server VM (build 25.102-b14, mixed mode), Cassandra 3.5.0, fresh install

• Severity: Critical
• Assignee: Unassigned
• Reporter: Bing Wu
• Created: 28/Oct/16 22:00
• Updated: 23/Apr/19 12:32
• Resolved: 29/Oct/16 17:34

Description

Summary of symptom:

• Set up is a multi-region cluster in AWS (5 regions). Each region has at least 4 hosts with RF=1/2 number of nodes, using V-nodes (256)

• How to reproduce:
  • On node A, start this repair job (again we are running fresh 3.5.0):

    nohup sudo nodetool repair -j 2 -pr -full myks > /tmp/repair.log 2>&1 &

  • Job starts fine, reporting progress like

    [2016-10-28 22:37:52,692] Starting repair command #1, repairing keyspace myks with repair options (parallelism: parallel, primary range: true, incremental: false, job threads: 2, ColumnFamilies: [], dataCenters: [], hosts: [], # of ranges: 256)
    [2016-10-28 22:38:35,099] Repair session 36f13450-9d5f-11e6-8bf7-a9f47ff986a9 for range [(4029874034937227774,4033949979656106020]] finished (progress: 1%)
    [2016-10-28 22:38:38,769] Repair session 36f30910-9d5f-11e6-8bf7-a9f47ff986a9 for range [(-2395606719402271267,-2394525508513518837]] finished (progress: 1%)
    [2016-10-28 22:38:48,521] Repair session 36f3f370-9d5f-11e6-8bf7-a9f47ff986a9 for range [(-5223108861718702793,-5221117649630514419]] finished (progress: 2%)
    

  • Then manually shutdown another node (node B) in the same region (haven't tried with other region yet but expect the same behavior from past experience)
  • Shortly after that seeing this message from job log (as well as in system.log) on node A:

    [2016-10-28 22:41:46,268] Repair session 37088ce1-9d5f-11e6-8bf7-a9f47ff986a9 for range [(-928974038666914990,-927967994563261540]] failed with error Endpoint /node_B_ip died (progress: 51%)
    

  • From this point on, repair job seems to hang:
    • no further messages from job log
    • nor any related messages in system.log
    • CPU stayed low (low single digit percent of 1 CPU)
  • After an hour (1hr), manually kill the repair jobs using "ps -eaf | grep repair"
  • Restart C* on node A
    • Verified system is up and no error messages in system.log
    • Also verified that there is no error messages from node B
  • After node A settles down (e.g. no new messages from system.log), restart the same repair job:

    nohup sudo nodetool repair -j 2 -pr -full myks > /tmp/repair.log 2>&1 &

  • Job failes pretty quickly, reporting error from more nodes B and K:

     <production>[[e-mail] ~]$ tail -f /tmp/repair.log 
    [2016-10-28 22:49:52,965] Starting repair command #1, repairing keyspace myks with repair options (parallelism: parallel, primary range: true, incremental: false, job threads: 2, ColumnFamilies: [], dataCenters: [], hosts: [], # of ranges: 256)
    [2016-10-28 22:50:15,839] Repair session e4180720-9d60-11e6-b2f9-cb9524b3c536 for range [(4029874034937227774,4033949979656106020]] failed with error [repair #e4180720-9d60-11e6-b2f9-cb9524b3c536 on myks/rtable, [(4029874034937227774,4033949979656106020]]] Validation failed in /node_K_ip (progress: 1%)
    [2016-10-28 22:50:17,158] Repair session e419dbe0-9d60-11e6-b2f9-cb9524b3c536 for range [(-2395606719402271267,-2394525508513518837]] failed with error [repair #e419dbe0-9d60-11e6-b2f9-cb9524b3c536 on myks/rtable, [(-2395606719402271267,-2394525508513518837]]] Validation failed in /node_B_ip (progress: 1%)
    [2016-10-28 22:50:18,256] Repair session e41b1460-9d60-11e6-b2f9-cb9524b3c536 for range [(-5223108861718702793,-5221117649630514419]] failed with error [repair #e41b1460-9d60-11e6-b2f9-cb9524b3c536 on myks/rtable, [(-5223108861718702793,-5221117649630514419]]] Validation failed in /node_B_ip (progress: 2%)
    

  • On the said nodes (B and K), seeing similar errors:

    ERROR [ValidationExecutor:5] 2016-10-28 22:58:45,307 CompactionManager.java:1320 - Cannot start multiple repair sessions over the same sstables
    ERROR [ValidationExecutor:5] 2016-10-28 22:58:45,307 Validator.java:261 - Failed creating a merkle tree for [repair #14378ec0-9d62-11e6-ab75-cd4d64a01b02 on myks/atable, [(4029874034937227774,4033949979656106020]]], /52.220.127.190 (see log for details)
    INFO  [AntiEntropyStage:1] 2016-10-28 22:58:45,307 Validator.java:274 - [repair #14378ec0-9d62-11e6-ab75-cd4d64a01b02] Sending completed merkle tree to /52.220.127.190 for myks.xtable
    ERROR [ValidationExecutor:5] 2016-10-28 22:58:45,308 CassandraDaemon.java:195 - Exception in thread Thread[ValidationExecutor:5,1,main]
    java.lang.RuntimeException: Cannot start multiple repair sessions over the same sstables
            at org.apache.cassandra.db.compaction.CompactionManager.getSSTablesToValidate(CompactionManager.java:1321) ~[apache-cassandra-3.5.0.jar:3.5.0]
            at org.apache.cassandra.db.compaction.CompactionManager.doValidationCompaction(CompactionManager.java:1211) ~[apache-cassandra-3.5.0.jar:3.5.0]
            at org.apache.cassandra.db.compaction.CompactionManager.access$700(CompactionManager.java:81) ~[apache-cassandra-3.5.0.jar:3.5.0]
            at org.apache.cassandra.db.compaction.CompactionManager$11.call(CompactionManager.java:841) ~[apache-cassandra-3.5.0.jar:3.5.0]
            at java.util.concurrent.FutureTask.run(FutureTask.java:266) ~[na:1.8.0_102]
            at java.util.concurrent.ThreadPoolExecutor.runWorker(ThreadPoolExecutor.java:1142) ~[na:1.8.0_102]
            at java.util.concurrent.ThreadPoolExecutor$Worker.run(ThreadPoolExecutor.java:617) [na:1.8.0_102]
            at java.lang.Thread.run(Thread.java:745) [na:1.8.0_102]
    INFO  [AntiEntropyStage:1] 2016-10-28 22:58:45,318 Validator.java:274 - [repair #14378ec0-9d62-11e6-ab75-cd4d64a01b02] Sending completed merkle tree to /52.220.127.190 for myks.ytable
    

  • At this point, we are back to where we were: kill the repair job on node A, then restart C* on BOTH nodes A and K, but still seeing the same exceptions except sometimes they are on other servers all over the ring.
• Business impact: I am in the process of launch a Cassandra based production system but I have to hold back now because how fragile repair is. And I am told by many sources that I have to rely on periodical repair jobs to fix data inconsistencies.
• The only work around was to rolling restart the Cassandra server on ALL nodes in the entire cluster
  • Then the repair job can proceed without any error

Issue Links

• duplicates: CASSANDRA-11824 If repair fails no way to run repair again
• is a clone of: CASSANDRA-10519 RepairException: [repair #... on .../..., (...,...]] Validation failed in /w.x.y.z
• is related to: CASSANDRA-10302 Track repair state for more reliable repair