Having N replicas more than nodes in the cluster
grant at basho.com
Tue May 11 14:02:24 EDT 2010
On May 11, 2010, at 12:43 PM, Randall Leeds wrote:
> On Tue, May 11, 2010 at 08:20, Grant Schofield <grant at basho.com> wrote:
>>> I have a two nodes cluster and a bucket with n_val = 3.
>>> I put more than 1 600 000 documents in the bucket and it takes less than 200 GB with Innostore.
>>> I shut down the second node with "q()." in a "riak console".
>>> Firstly, there was no synchronisation of data on the first node, i assume it's better because very similar to a node crash.
>>> Secondly, when i did a read request on a document with :
>>> - ?r=1 -> success
>>> - ?r=2 -> success
>>> - ?r=3 -> fail
>> This is due to the fact that 2 of the 3 vnodes were able to respond with your data, but not the third because your took down the node that had 1 of the 3 partitions.
> After the next round of gossip, will the remaining Riak node attempt
> to reclaim these vnodes or does the ring structure persist until the
> node has explicitly left?
The remaining nodes won't try to take over a partition unless the missing node is removed from the cluster explicitly. However, other nodes will handle writes so when the node comes back online it will go through the hinted handoff process.
More information about the riak-users