[
https://issues.apache.org/jira/browse/USERGRID-527?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14389467#comment-14389467
]
Todd Nine commented on USERGRID-527:
------------------------------------
See the email from Rafal. I think we should mock this up via Junit, so that we
can assert our routing works correctly.
Hello Todd,
As promised, I'm sending the hashing. Elasticsearch 1.x uses DjbHash:
https://github.com/elastic/elasticsearch/blob/1.5/src/main/java/org/elasticsearch/cluster/routing/operation/hash/djb/DjbHashFunction.java
The routing logic is in the PlainOperationRouting:
https://github.com/elastic/elasticsearch/blob/1.5/src/main/java/org/elasticsearch/cluster/routing/operation/plain/PlainOperationRouting.java
The interesting method is starting on line 265:
private int shardId(ClusterState clusterState, String index, String type,
@Nullable String id, @Nullable String routing)
In Elasticsearch 2.0 things will be a bit different:
https://github.com/elastic/elasticsearch/blob/master/src/main/java/org/elasticsearch/cluster/routing/OperationRouting.java
You are interested in method:
private int shardId(ClusterState clusterState, String index, String type,
String id, @Nullable String routing)
starting at line 250. In 2.0 Elasticsearch does hash calculations
using implementation of
org.elasticsearch.cluster.routing.HashFunction. The available
implementations will be:
* org.elasticsearch.cluster.routing.SimpleHashFunction
* org.elasticsearch.cluster.routing.Murmur3HashFunction
* org.elasticsearch.cluster.routing.DjbHashFunction
The default hash function will be switched to Murmur hash.
> Write external ES test to ensure we get even routing across the ES cluster
> --------------------------------------------------------------------------
>
> Key: USERGRID-527
> URL: https://issues.apache.org/jira/browse/USERGRID-527
> Project: Usergrid
> Issue Type: Story
> Reporter: Jeffrey
> Priority: Critical
>
> One of the problems we have faced is uneven distribution in the ES cluster.
> We have made changes for document routing that seemed to help, but we are
> still getting hot nodes in the cluster. We need to have a test that hits ES
> directly with load as UG would do to test our cluster distribution.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
