[
https://issues.apache.org/jira/browse/IGNITE-4378?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15742451#comment-15742451
]
Alexei Scherbakov commented on IGNITE-4378:
-------------------------------------------
Dmitriy, thanks for the comment.
* I think this is a good idea. Abstract affinity can implement the logic to
combine filters (or some other building blocks) together to produce various
effects. Currently, if I use excludeNeighbors property, I basically can't
further filter down backups. Such scenarios should be implemented by
combination of filters.
I need to do more research on current affinity implementations to provide the
sample base code.
* {{AffinityPrimaryFilter}} filters only primary nodes. But they can be put
together with {{AffinityBackupFilter}} in one interface.
* Agreed.
Also I think we should provide test framework to verify AF feasibiltiy and
quality.
> Affinity function should support assigning partition to subset of cluster
> nodes
> -------------------------------------------------------------------------------
>
> Key: IGNITE-4378
> URL: https://issues.apache.org/jira/browse/IGNITE-4378
> Project: Ignite
> Issue Type: New Feature
> Components: cache
> Reporter: Dmitriy Setrakyan
> Assignee: Alexei Scherbakov
> Fix For: 2.0
>
>
> Currently both default affinity function(AF) implementations randomly choose
> primary node among all topology nodes.
> This may not be enough to handle complex data placement scenarios without
> implementing own AF.
> On example, some partitions can be assigned to more powerful hardware, or
> limited to subset of cluster nodes due to ease of management or fault
> tolerance scenarios.
> We should implement node filter, which will allow to choose subset of cluster
> nodes to place primary and backup partitions.
> With already existing ability to filter backup nodes (using
> {{AffinityBackupFilter}}) it will allow to implement different approaches to
> data placement with Ignite without resorting to custom AF.
> It's also desirable to include a practical example of both topology filters
> based on node attribute values.
> Proposed primary filter interface is below.
> {noformat}
> /**
> * Allows partition placement to subset of cluster node.
> *
> * Backup nodes also will be assigned from the subset.
> */
> public interface AffinityPrimaryFilter extends IgniteBiClosure<Integer,
> List<ClusterNode>, List<ClusterNode>> {
> /**
> * Return nodes allowed to contain given partition.
> * @param partition Partition.
> * @param currentTopologyNodes All nodes from current topology.
> * @return Subset of nodes.
> */
> @Override public List<ClusterNode> apply(Integer partition,
> List<ClusterNode> currentTopologyNodes);
> }
> {noformat}
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
