[
https://issues.apache.org/jira/browse/HBASE-15368?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15176063#comment-15176063
]
Dave Latham commented on HBASE-15368:
-------------------------------------
This still is not going to work well. Using the min timestamp does not fix it.
With relative windows, there's nothing to prevent a fully compacted HFile for
one window from sliding across the window boundary, being compacted into that
window, and sliding quickly to the next again. There would be no reliable
behavior of actual data staying partitioned in a well defined way. For
example, suppose you have windows of .2 days, 1 day, 5 days, 25 days, each with
a single HFile containing well patterned data spread across most of their
windows. A few seconds later the .2 day HFile could be in the 1 day window,
then be compacted with it, then that HFile slide to the 5 day window, be
compacted, then to the 25 day window. Now quickly all of your data winds up in
a single window and HFile. This is just one example of how the relative
windows could go wrong. To make it work you would probably need to more
rigidly control which files within a window could be allowed to be compacted
together - I'm not sure a reasonable scheme exists that doesn't wind up
approximating the fixed windows anyway.
> Add relative window support
> ---------------------------
>
> Key: HBASE-15368
> URL: https://issues.apache.org/jira/browse/HBASE-15368
> Project: HBase
> Issue Type: Sub-task
> Components: Compaction
> Reporter: Duo Zhang
> Assignee: Duo Zhang
> Attachments: HBASE-15368-v1.patch, HBASE-15368.patch
>
>
> To better determine 'hot' data.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
