nodes aren't talking
justin at basho.com
Mon Nov 9 14:21:30 EST 2009
On Mon, Nov 9, 2009 at 1:40 PM, B. Todd Burruss <bburruss at real.com> wrote:
> thx dan. your explanation of vnodes makes sense, but isn't what i would
> expect. for HA reasons i would expect the riak server to understand the
> physical location of 2 vnodes (maybe use IP address) and only replicate
> to a vnode that is on another physical server. if riak works as you
> describe it doesn't seem very viable for an HA system.
Riak attempts to give you the best of both worlds here. The "ideal"
vnodes to send requests to will always be on different actual nodes as
long as there are >= N such nodes to choose from.
When some nodes are down, the hinted handoff mechanism will send the
relevant messages to other vnodes where they will be managed until the
ideal nodes regain connectivity. When possible, the handoff nodes
will be on separate actual nodes, but if enough nodes are down so as
to make less than N actual nodes available, then some of the handoff
vnodes may be on the same actual node.
This provides both long-term availability by spreading the data across
as many physical nodes as possible and also short-term availability by
allowing that data to be captured even when a sufficient number of
nodes are not reachable.
More information about the riak-users