cassandra-commits mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Stu Hood (JIRA)" <>
Subject [jira] Commented: (CASSANDRA-1577) Split gossiped load into 'primary' and 'total'
Date Thu, 07 Oct 2010 19:31:30 GMT


Stu Hood commented on CASSANDRA-1577:

After a lot more discussion with Nick, it would appear that this issue is invalid if we make
certain guarantees about replication strategies. As long as the replication strategy chooses
nodes that are _clustered near the official position of the token in the ring_, whether the
load is caused by being the first or nth replica doesn't matter. An adversary replication
strategy would choose nodes at essentially random positions in the ring, and then return those
positions deterministically: for that case, load balancing would fail.

> Split gossiped load into 'primary' and 'total'
> ----------------------------------------------
>                 Key: CASSANDRA-1577
>                 URL:
>             Project: Cassandra
>          Issue Type: Improvement
>          Components: Core
>            Reporter: Stu Hood
>            Assignee: Stu Hood
>             Fix For: 0.8
> Bootstrap and load balancing decisions are currently made using the total load on a node,
as opposed to data the node is primarily responsible for due to a token hashing to its local
range. The latter number is the only reasonable number to be load balancing on, since ReplicationStrategies
can make arbitrary numbers of hops on the ring to find additional places to put replicas.
> To find the primary load efficiently, we can use the SSTableReader.getPositionsForRanges
method, although we should probably move the load calculation outside of the atomic swap in
> The format of the serialized 'LOAD' gossip state will need to change in a backwards incompatible

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

View raw message