[ 
https://issues.apache.org/jira/browse/HBASE-20636?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16490097#comment-16490097
 ] 

Allan Yang commented on HBASE-20636:
------------------------------------

I think bump up the HFile version  for a new type of bloom filter is too much. 
Maybe we can go around, like:
Set the BLOOM_FILTER_TYPE_KEY to NONE if it is a prefix filter. Add another 
file info e.g. PREFIX_BLOOM_FILTER_FLAG. The type of the new BF can be set 
here. The old version of RegionServer will regards the  BF as NONE while the 
new server has enough information. 
I don't know if it is a good idea, [~apurtell], [~andrewcheng] 

> Introduce two bloom filter type : ROWPREFIX and ROWPREFIX_DELIMITED
> -------------------------------------------------------------------
>
>                 Key: HBASE-20636
>                 URL: https://issues.apache.org/jira/browse/HBASE-20636
>             Project: HBase
>          Issue Type: New Feature
>          Components: regionserver
>            Reporter: Guangxu Cheng
>            Assignee: Guangxu Cheng
>            Priority: Major
>         Attachments: HBASE-20636.master.001.patch, 
> HBASE-20636.master.002.patch
>
>
> As we all know, HBase uses BloomFilter(ROW and ROWCOL) to filter unnecessary 
> files to improve read performance. But they only support Get and do not 
> support Scan.
> In our company(Tencent), many users need to scan all rows with the same 
> prefix, such as Tencent Game. Game user's some operational record will be 
> written into HBase, each game user will have a lot of records, the rowkey is 
> constructed as userid+'#'+timestamps. So we can scan all records for a given 
> user for a specified period.
> For this scenario, we designed the prefix Bloom filter. If the startRow and 
> stopRow of the Scan has a valid common prefix, the scan will be allowed to 
> use BloomFilter to filter files which will enhance the performance of the 
> scan.
> Now, this feature has been running on our cluster over a year, and scan 
> performance for this scenario has been improved by more than one times than 
> before.



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)

Reply via email to