hbase-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Andrew Purtell (JIRA)" <j...@apache.org>
Subject [jira] [Reopened] (HBASE-14397) PrefixFilter doesn't filter all remaining rows if the prefix is longer than rowkey being compared
Date Tue, 22 Sep 2015 16:51:05 GMT

     [ https://issues.apache.org/jira/browse/HBASE-14397?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]

Andrew Purtell reopened HBASE-14397:
------------------------------------

Sorry I'm late. I have some concerns with this patch. 

Why skip only some rows where the prefix is longer than the key being compared? Why not all?
Why make it a fixed number? Why is there no logging or indication that the max skip limit
was reached?  This hasn't fixed the problem, its added a different type of surprising behavior.


I think we should amend or revert and recommit the current patch with an update that removes
the "skippedCompareRows" stuff. Just skip the rows if the keys aren't long enough to match
the prefix, and document this _consistent_ behavior. 

> PrefixFilter doesn't filter all remaining rows if the prefix is longer than rowkey being
compared
> -------------------------------------------------------------------------------------------------
>
>                 Key: HBASE-14397
>                 URL: https://issues.apache.org/jira/browse/HBASE-14397
>             Project: HBase
>          Issue Type: Improvement
>          Components: Filters
>    Affects Versions: 2.0.0
>            Reporter: Jianwei Cui
>            Assignee: Jianwei Cui
>            Priority: Minor
>             Fix For: 2.0.0, 1.3.0
>
>         Attachments: HBASE-14397-trunk-v1.patch
>
>
> The PrefixFilter will filter rowkey as:
> {code}
>   public boolean filterRowKey(Cell firstRowCell) {
>     ...
>     int length = firstRowCell.getRowLength();
>     if (length < prefix.length) return true; // ===> return directly if the prefix
is longer
>     ....
>     if ((!isReversed() && cmp > 0) || (isReversed() && cmp < 0))
{
>       passedPrefix = true;
>     }
>     filterRow = (cmp != 0);
>     return filterRow;
>   }
> {code}
> If the prefix is longer than the current rowkey, PrefixFilter#filterRowKey will filter
the rowkey directly without comparing, so that won't set 'passedPrefix' flag even the current
row is larger than the prefix.
> For example, if there are three rows 'a', 'b' and 'c' in the table, and we issue a scan
request as:
> {code}
> hbase(main):001:0> scan 'test_table', {STARTROW => 'a', FILTER => "(PrefixFilter
('aa'))"}
> {code}
> The region server will check the three rows before returning.  In our production, the
user issue a scan with a PrefixFilter. The prefix is longer than the rowkeys of following
millions of rows, so the region server will continue to check rows until hit a rowkey longer
than the prefix. This make the client easily timeout. To fix this case, it seems we need to
compare the prefix with the rowkey every serveral rows even when the prefix is longer.



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

Mime
View raw message