cassandra-commits mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Brandon Williams (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (CASSANDRA-7667) Improve migration
Date Fri, 01 Aug 2014 20:57:38 GMT

    [ https://issues.apache.org/jira/browse/CASSANDRA-7667?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14082952#comment-14082952
] 

Brandon Williams commented on CASSANDRA-7667:
---------------------------------------------

I'm not sure this makes a lot of sense.  Adding DC optimization like writes _might_ save on
latency, but it's also adding an extra hop so maybe not, and the point of dc write optimization
is bandwidth, of which little is used here and schema changes aren't that frequent.  Remember
that broadcast is just best effort, if it fails it doesn't matter because schema propagation
is two-pronged: broadcast if possible, but also passively announce via gossip and let the
other side pull.

> Improve migration
> -----------------
>
>                 Key: CASSANDRA-7667
>                 URL: https://issues.apache.org/jira/browse/CASSANDRA-7667
>             Project: Cassandra
>          Issue Type: Improvement
>            Reporter: Robert Stupp
>
> Idea is to optimize migration especially for large clusters with multiple DCs.
>  
> Currently {{MigrationManager.announce()}} simply broadcasts to all nodes.
>  
> Idea is to not announce schema migration to every node.
>  
> *Against remote DCs*: only contact a single node in each remote DC. This could make migration
faster, because network latency comes into play. And it lowers cross-DC network requirements
(since 1-to-all communication creates a huge amount of TCP connections). Drawback: remote-DC-announcement
must be reliable (receiver must ACK - if it doesn't, sender must retry another node).
>  
> *Within the local DC*: use some gossip approach - e.g. only contact N nodes at once.
Not sure about this one  - it might be nice, if there are really many (~100) nodes in one
DC. But it's a bit difficult to guarantee, that all nodes will receive the announcement without
giving up the "fire-and-forget" approach.



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

Mime
View raw message