cassandra-commits mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Stu Hood (JIRA)" <j...@apache.org>
Subject [jira] [Updated] (CASSANDRA-2455) Improve counter disk usage
Date Mon, 11 Apr 2011 22:17:06 GMT

     [ https://issues.apache.org/jira/browse/CASSANDRA-2455?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]

Stu Hood updated CASSANDRA-2455:
--------------------------------

    Description: 
Counter values currently use a huge amount of space on disk:
{code}
(header + length + RF * (nodeid + count + clock))
==
(2 + 2 + RF * (16 + 8 + 8)) bytes
{code}

Type specific compression (as on CASSANDRA-2398) is a long term solution to this problem,
but we need a short term fix to make a large volume of counters possible.

The largest and most redundant part of the counter is the nodeid, which is now 16 bytes per
replica. One proposed improvement would be keep a per-sstable dictionary of all replica sets,
and to assume the replicas are sorted by nodeid in the counter value. This would allow us
to encode the replica as a single integer in the counter value, and to use it to look up the
replica set in the dictionary. Assuming an integer replica set id, you could allow for 2^32
replica changes with 4 total bytes of overhead in each counter:
{code}
(header + length + replicasetid + RF (count + clock))
==
(2 + 2 + 4 + RF * (8 + 8)) bytes
{code}

  was:
Counter values currently use a huge amount of space on disk:
{{(header + length + RF * (nodeid + count + clock)) bytes}}
or
{{(2 + 2 + RF * (16 + 8 + 8)) bytes}}

Type specific compression (as on CASSANDRA-2398) is a long term solution to this problem,
but we need a short term fix to make a large volume of counters possible.

The largest and most redundant part of the counter is the nodeid, which is now 16 bytes per
replica. One proposed fix would be keep a per-sstable dictionary of all replica sets, and
to assume the replicas are sorted by nodeid in the counter value. This would allow us to encode
the replica as a single integer in the counter value, and to use it to look up the replica
set in the dictionary. Assuming an integer replica set id, you could allow for 2^32 replica
changes with 4 total bytes of overhead in each counter:
{{(header + length + replicasetid + RF (count + clock)) bytes}}
or
{{(2 + 2 + 4 + RF * (8 + 8)) bytes}}


> Improve counter disk usage
> --------------------------
>
>                 Key: CASSANDRA-2455
>                 URL: https://issues.apache.org/jira/browse/CASSANDRA-2455
>             Project: Cassandra
>          Issue Type: Improvement
>            Reporter: Stu Hood
>
> Counter values currently use a huge amount of space on disk:
> {code}
> (header + length + RF * (nodeid + count + clock))
> ==
> (2 + 2 + RF * (16 + 8 + 8)) bytes
> {code}
> Type specific compression (as on CASSANDRA-2398) is a long term solution to this problem,
but we need a short term fix to make a large volume of counters possible.
> The largest and most redundant part of the counter is the nodeid, which is now 16 bytes
per replica. One proposed improvement would be keep a per-sstable dictionary of all replica
sets, and to assume the replicas are sorted by nodeid in the counter value. This would allow
us to encode the replica as a single integer in the counter value, and to use it to look up
the replica set in the dictionary. Assuming an integer replica set id, you could allow for
2^32 replica changes with 4 total bytes of overhead in each counter:
> {code}
> (header + length + replicasetid + RF (count + clock))
> ==
> (2 + 2 + 4 + RF * (8 + 8)) bytes
> {code}

--
This message is automatically generated by JIRA.
For more information on JIRA, see: http://www.atlassian.com/software/jira

Mime
View raw message