hbase-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Ted Yu (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (HBASE-15102) HeapMemoryTuner can "overtune" memstore size and suddenly drop it into blocking zone
Date Wed, 13 Jan 2016 23:55:40 GMT

    [ https://issues.apache.org/jira/browse/HBASE-15102?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15097313#comment-15097313
] 

Ted Yu commented on HBASE-15102:
--------------------------------

{code}
280	        newMemstoreSize = context.getCurMemStoreSize() - step;
281	        if (newMemstoreSize < minMemstoreSize) {
282	          newMemstoreSize = minMemstoreSize;
283	          step = context.getCurMemStoreSize() - newMemstoreSize;
284	        }
285	        newBlockCacheSize = context.getCurBlockCacheSize() + step;
{code}
The value of step is modified inside the if block. This potentially would result in mismatch
between adjusts to memstore size and blockcache size.

> HeapMemoryTuner can "overtune" memstore size and suddenly drop it into blocking zone
> ------------------------------------------------------------------------------------
>
>                 Key: HBASE-15102
>                 URL: https://issues.apache.org/jira/browse/HBASE-15102
>             Project: HBase
>          Issue Type: Bug
>          Components: regionserver
>    Affects Versions: 2.0.0, 1.2.0, 1.3.0, 1.2.1
>            Reporter: Ashu Pachauri
>            Assignee: Ashu Pachauri
>            Priority: Critical
>         Attachments: HBASE-15102-V0.patch
>
>
> DefaultHeapMemoryTuner often resets the maximum step size for tuning to 8% of total heap
size. Often, when the size of memstore is to be decreased while tuning, the 8% tuning can
suddenly drop the memstore size below the low water mark of the previous memstore size (which
could potentially be  the used size of the memstore)
> This is problematic because suddenly it blocks all the updates by suddenly causing a
situation where memstore used size is above high water mark. This has a very bad performance
impact on an otherwise fine HBase cluster. 



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

Mime
View raw message