[
https://issues.apache.org/jira/browse/SOLR-12088?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16410958#comment-16410958
]
Cao Manh Dat edited comment on SOLR-12088 at 3/23/18 11:31 AM:
---------------------------------------------------------------
I wrote some test to benchmark the indexing throughput in case of
# High indexing + delete replica (on 7.2)
# High indexing + delete replica (after the fix SOLR-12087)
# High indexing + down node (on 7.2)
I do see a downgrade in update per second for case 1. So it seems the problem
relates to the implementation of old LIR.
I modified {{DeleteReplicaTest.deleteReplicaOnIndexing}} to count the number of
LIR threads get started and the number of time
{{ZkController.ensureReplicaInLeaderInitiatedRecovery}} get called
||Time (in sec)||LIR threads get started (times)||ZKController get called
(times)||
|100|96132|1504|
|500|6513|509690|
|1000|6444|594742|
As we can see after 500 seconds, there are no more LIR threads get called and I
do see the indexing get recovered. [~jerry.bao]: are you sure that the
downgrade indexing last forever?
was (Author: caomanhdat):
I wrote some test to benchmark the indexing throughput in case of
# High indexing + delete replica (on 7.2)
# High indexing + delete replica (after the fix SOLR-12087)
# High indexing + down node (on 7.2)
I do see a downgrade in update per second for case 1. So it seems the problem
relate to the implementation of old LIR.
> Shards with dead replicas cause increased write latency
> -------------------------------------------------------
>
> Key: SOLR-12088
> URL: https://issues.apache.org/jira/browse/SOLR-12088
> Project: Solr
> Issue Type: Bug
> Security Level: Public(Default Security Level. Issues are Public)
> Components: SolrCloud
> Affects Versions: 7.2
> Reporter: Jerry Bao
> Priority: Major
>
> If a collection's shard contains dead replicas, write latency to the
> collection is increased. For example, if a collection has 10 shards with a
> replication factor of 3, and one of those shards contains 3 replicas and 3
> downed replicas, write latency is increased in comparison to a shard that
> contains only 3 replicas.
> My feeling here is that downed replicas should be completely ignored and not
> cause issues to other alive replicas in terms of write latency.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
