hbase-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "stack (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (HBASE-16650) Wrong usage of BlockCache eviction stat for heap memory tuning
Date Tue, 20 Sep 2016 19:37:20 GMT

    [ https://issues.apache.org/jira/browse/HBASE-16650?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15507567#comment-15507567
] 

stack commented on HBASE-16650:
-------------------------------

Looking at the patch, can we update the description for the metric to explain we DO NOT include
evictions because an hfile was removed. Can do on commit.

Is it right putting code under this check: 	    if (evictedByEvictionProcess) {

We are setting some flags in this block that are used later. Do we need those flags when an
eviction because an hfile has been removed?

This is ugly but we are already keeping blockcache in a global.... when we fix this, we can
fix this new addition:

  private static LruBlockCache GLOBAL_L1_CACHE_INSTANCE;


Otherwise, +1.

> Wrong usage of BlockCache eviction stat for heap memory tuning
> --------------------------------------------------------------
>
>                 Key: HBASE-16650
>                 URL: https://issues.apache.org/jira/browse/HBASE-16650
>             Project: HBase
>          Issue Type: Sub-task
>    Affects Versions: 1.0.0
>            Reporter: Anoop Sam John
>            Assignee: Anoop Sam John
>             Fix For: 2.0.0, 1.4.0
>
>         Attachments: HBASE-16650.patch
>
>
> 1. We use the stat evictedBlocksCount - A block can get evicted because of eviction thread
due to lack of space or because of removal of an HFile itself (After a compaction). We should
not consider the latter in the tune decision at all. These are actually invalidation of blocks.
Should the stat counter itself not use this count of evicted blocks? I think yes. This will
give wrong message to users that there are lot of real eviction happening.
> 2. In case L1+ L2 combined block cache, what we use is the sum of evictions from both.
But we will be tuning L1 size alone. Eviction count from L2 should not affect the tuning of
L1



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

Mime
View raw message