[
https://issues.apache.org/jira/browse/HBASE-22448?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16846396#comment-16846396
]
Zheng Hu edited comment on HBASE-22448 at 5/23/19 3:07 AM:
-----------------------------------------------------------
To show the desgin, I draw a picture (see
[^filter-list-with-or-internal-2.png]) for this FilterListWithOR and why this
case is time-consuming, assume we have a FilterList and all the subFilters are
ColumnPrefixFilter:
{code:java}
filterList = subFilter1 OR subFilter2 OR subFilter3
{code}
For cell-0, subFilter1#nextHint is cell-6 , subFilter2#nextHint is cell-1,
subFilter3#nextHint is cell-7, then the filterList will choose the minimal next
hint between all those subFilters , which will be cell-1. That's correct.
For cell-1, sbufFilter1 need a comparison to see whether current cell has
reached its previous next hint cell ( cell-6). also the subFilter3 need the
comparision..
For cell-2. both the subFilter1 and subFilter2 need the same comparision as
cell-1 did.
...
All the black dots means the extra comparisions compared branch-1.2*, while it
was designed in a wrong way for branch-1.2 so seems more efficient.
was (Author: openinx):
To show the desgin, I draw a picture (see
https://issues.apache.org/jira/secure/attachment/12969462/filter-list-with-or-internal-2.png)
for this FilterListWithOR and why this case is time-consuming, assume we have
a FilterList and the subFilter is a ColumnPrefixFilter:
{code}
filterList = subFilter1 OR subFilter2 OR subFilter3
{code}
For cell-0, subFilter1#nextHint is cell-6 , subFilter2#nextHint is cell-1,
subFilter3#nextHint is cell-7, then the filterList will choose the minimal next
hint between all those subFilters , which will be cell-1. That's correct.
For cell-1, sbufFilter1 need a comparison to see whether current cell has
reached its previous next hint cell ( cell-6). also the subFilter3 need the
comparision..
For cell-2. both the subFilter1 and subFilter2 need the same comparision as
cell-1 did.
...
All the black dots means the extra comparisions compared branch-1.2*, while it
was designed in a wrong way for branch-1.2 so seems more efficient.
> Scan is slow for Multiple Column prefixes
> -----------------------------------------
>
> Key: HBASE-22448
> URL: https://issues.apache.org/jira/browse/HBASE-22448
> Project: HBase
> Issue Type: Bug
> Components: Scanners
> Affects Versions: 1.4.8, 1.4.9
> Reporter: Karthick
> Assignee: Zheng Hu
> Priority: Critical
> Labels: prefix, scan, scanner
> Fix For: 1.5.0, 1.4.10
>
> Attachments: 0001-benchmark-UT.patch, HBaseFileImport.java,
> filter-list-with-or-internal-2.png,
> org.apache.hadoop.hbase.filter.TestSlowColumnPrefix-output.zip,
> qualifiers.txt, scanquery.txt
>
>
> While scanning a row (around 10 lakhs columns) with 100 column prefixes, it
> takes around 4 seconds in hbase-1.2.5 and when the same query is executed in
> hbase-1.4.9 it takes around 50 seconds.
> Is there any way to optimise this?
>
> *P.S:*
> We have applied the patch provided in
> [-HBASE-21620-|https://jira.apache.org/jira/browse/HBASE-21620] and
> [-HBASE-21734-|https://jira.apache.org/jira/browse/HBASE-21734] . Attached
> *qualifiers*.*txt* file which contains the column keys. Use the
> *HBaseFileImport.java* file provided to populate in your table and use
> *scanquery.txt* to query.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
