cassandra-commits mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Nick Bailey (JIRA)" <j...@apache.org>
Subject [jira] [Created] (CASSANDRA-7317) Repair range validation and calculation is off
Date Thu, 29 May 2014 03:00:06 GMT
Nick Bailey created CASSANDRA-7317:
--------------------------------------

             Summary: Repair range validation and calculation is off
                 Key: CASSANDRA-7317
                 URL: https://issues.apache.org/jira/browse/CASSANDRA-7317
             Project: Cassandra
          Issue Type: Bug
            Reporter: Nick Bailey
             Fix For: 1.2.17, 2.0.8, 2.1 rc1


>From what I can tell the calculation (using the -pr option) and validation of tokens for
repairing ranges is broken. Or at least should be improved. Using an example with ccm:

Nodetool ring:

{noformat}
Datacenter: dc1
==========
Address    Rack        Status State   Load            Owns                Token
                                                                          -10
127.0.0.1  r1          Up     Normal  188.96 KB       50.00%              -9223372036854775808
127.0.0.2  r1          Up     Normal  194.77 KB       50.00%              -10

Datacenter: dc2
==========
Address    Rack        Status State   Load            Owns                Token
                                                                          0
127.0.0.4  r1          Up     Normal  160.58 KB       0.00%               -9223372036854775798
127.0.0.3  r1          Up     Normal  139.46 KB       0.00%               0
{noformat}

Schema:

{noformat}
CREATE KEYSPACE system_traces WITH replication = {
  'class': 'NetworkTopologyStrategy',
  'dc2': '2',
  'dc1': '2'
};
{noformat}

Repair -pr:
{noformat}
[Nicks-MacBook-Pro:21:35:58 cassandra-2.0] cassandra$ bin/nodetool -p 7100 repair -pr system_traces
[2014-05-28 21:36:01,977] Starting repair command #12, repairing 1 ranges for keyspace system_traces
[2014-05-28 21:36:02,207] Repair session f984d290-e6d9-11e3-9edc-5f8011daec21 for range (0,-9223372036854775808]
finished
[2014-05-28 21:36:02,207] Repair command #12 finished
[Nicks-MacBook-Pro:21:36:02 cassandra-2.0] cassandra$ bin/nodetool -p 7200 repair -pr system_traces
[2014-05-28 21:36:14,086] Starting repair command #1, repairing 1 ranges for keyspace system_traces
[2014-05-28 21:36:14,406] Repair session 00bd45b0-e6da-11e3-98fc-5f8011daec21 for range (-9223372036854775798,-10]
finished
[2014-05-28 21:36:14,406] Repair command #1 finished
{noformat}

Note that repairing both nodes in dc1, leaves very small ranges unrepaired. For example (-10,0].
Repairing the 'primary range' in dc2 will repair those small ranges. Maybe that is the behavior
we want but it seems counterintuitive.

The behavior when manually trying to repair the full range of 127.0.0.01 definitely needs
improvement though.

Repair command:
{noformat}
[Nicks-MacBook-Pro:21:50:44 cassandra-2.0] cassandra$ bin/nodetool -p 7100 repair -st -10
-et -9223372036854775808 system_traces
[2014-05-28 21:50:55,803] Starting repair command #17, repairing 1 ranges for keyspace system_traces
[2014-05-28 21:50:55,804] Starting repair command #17, repairing 1 ranges for keyspace system_traces
[2014-05-28 21:50:55,804] Repair command #17 finished
[Nicks-MacBook-Pro:21:50:56 cassandra-2.0] cassandra$ echo $?
1
{noformat}

system.log:
{noformat}
ERROR [Thread-96] 2014-05-28 21:40:05,921 StorageService.java (line 2621) Repair session failed:
java.lang.IllegalArgumentException: Requested range intersects a local range but is not fully
contained in one; this would lead to imprecise repair
{noformat}

* The actual output of the repair command doesn't really indicate that there was an issue.
Although the command does return with a non zero exit status.
* The error here is invisible if you are using the synchronous jmx repair api. It will appear
as though the repair completed successfully.
* Personally, I believe that should be a valid repair command. For the system_traces keyspace,
127.0.0.1 is responsible for this range (and I would argue the 'primary range' of the node.



--
This message was sent by Atlassian JIRA
(v6.2#6252)

Mime
View raw message