cassandra-commits mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Stu Hood (JIRA)" <>
Subject [jira] Created: (CASSANDRA-1961) Some kinds of membership changes will still cause overcounts
Date Tue, 11 Jan 2011 02:28:48 GMT
Some kinds of membership changes will still cause overcounts

                 Key: CASSANDRA-1961
             Project: Cassandra
          Issue Type: Bug
          Components: Core
            Reporter: Stu Hood
             Fix For: 0.8

Assume replicas A, B, C, and a joining node D, where D is joining between A and B. The join
process will remove C from the replica set, but C will still be holding counts for those replicas
(unless cleanup is run, which we can't safely assume). If a second membership change occurs
such that any of A, B or D leave the ring, C will be acting as a new member of the replica
set, but it will still be holding its old counts.

The join will:
 * BOOTSTRAP - D will bootstrap from the nearest replica, possibly C (but not necessarily)

The leave will either: 
 * UNBOOTSTRAP - D will send to C
 * RESTORE_REPLICA_COUNT - since D is assumed dead, C will stream from the nearest replica

Only the AES stream task performs fixups of counters: in all other cases I think we assume
that 'nodetool cleanup' has run, so it is possible to overcount.

This message is automatically generated by JIRA.
You can reply to this email to add a comment to the issue online.

View raw message