[
https://issues.apache.org/jira/browse/HBASE-10451?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Anoop Sam John updated HBASE-10451:
-----------------------------------
Attachment: HBASE-10451_V5.patch
Sorry.. I have to again -1 my own patch and have to change the approach.
Dictionary impl is LRU and so when find index of an entry the entry will be
moved to head. So we can not assume the head entry is the one we have to
delete on move previous. [KV can have a new tag and an already added tag as
second tag] Also based on the counter we can not know the position of entry to
be deleted, as when the size of the dictionary reaches capacity, we will add
new entries in random locations (pointed by current tail).
A simpler approach to solve the problem can be like
When we do movePrevious, we cache the tag bytes into the current State and also
a boolean state set as not to do uncompress of tags in the next time
decodeNext() call. Also we will cache the total compressed length of the tags
bytes. When next time decode of next entry happens, we will not try
uncompressinf tags. We will just use the cached tags bytes and just skip
'tagsCompressedLength' bytes of source stream.
With this patch the YCSB test scenario what Ram tested passes.(Scenario
reproduced in a UT) Also the Integration test passes.
Ram can you run YCSB in your env also once pls and confirm this is fine ?
> Enable back Tag compression on HFiles
> -------------------------------------
>
> Key: HBASE-10451
> URL: https://issues.apache.org/jira/browse/HBASE-10451
> Project: HBase
> Issue Type: Bug
> Affects Versions: 0.98.0
> Reporter: Anoop Sam John
> Assignee: Anoop Sam John
> Priority: Critical
> Fix For: 0.98.1, 0.99.0
>
> Attachments: HBASE-10451.patch, HBASE-10451_V2.patch,
> HBASE-10451_V3.patch, HBASE-10451_V4.patch, HBASE-10451_V5.patch
>
>
> HBASE-10443 disables tag compression on HFiles. This Jira is to fix the
> issues we have found out in HBASE-10443 and enable it back.
--
This message was sent by Atlassian JIRA
(v6.1.5#6160)