hbase-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Esteban Gutierrez (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (HBASE-17755) CellBasedKeyBlockIndexReader#midkey should exhaust search of the target middle key on skewed regions
Date Wed, 08 Mar 2017 00:05:38 GMT

    [ https://issues.apache.org/jira/browse/HBASE-17755?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15900424#comment-15900424
] 

Esteban Gutierrez commented on HBASE-17755:
-------------------------------------------

I worked on this with [~saint.ack@gmail.com] and we should follow up with additional improvements.

> CellBasedKeyBlockIndexReader#midkey should exhaust search of the target middle key on
skewed regions
> ----------------------------------------------------------------------------------------------------
>
>                 Key: HBASE-17755
>                 URL: https://issues.apache.org/jira/browse/HBASE-17755
>             Project: HBase
>          Issue Type: Bug
>          Components: HFile
>            Reporter: Esteban Gutierrez
>            Assignee: Esteban Gutierrez
>
> We have always been returning the middle key of the the block index regardless the distribution
of the data on an HFile. A side effect of that approach is that when millions of rows share
the same key its quite easy to run into a situation when the start key is equal to the middle
key or when the end key is equal to the middle key making that HFile nearly impossible to
split until enough data is written into the region and the middle key shifts to another row
or when an operator uses a custom split point in order to split that region. 
> Instead we should exhaust the search of the middle key in the block index in order to
be able to split an HFile earlier when possible even if our edge case is to serve a region
that could hold a single key with millions of versions of a row or with millions of qualifiers
on the same row.



--
This message was sent by Atlassian JIRA
(v6.3.15#6346)

Mime
View raw message