[jira] [Commented] (HIVE-3420) Inefficiency in hbase handler when process query including rowkey range scan

["Navis (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/HIVE-3420?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13685109#comment-13685109 ] 

Navis commented on HIVE-3420:
-----------------------------

For a table with two region, executed simple query with PPD-able predicates,
input rows (without patch) : 2,472 rows 
input rows (with patch) : 1,236 rows

For large hbase table, it can make a big difference.
                
> Inefficiency in hbase handler when process query including rowkey range scan
> ----------------------------------------------------------------------------
>
>                 Key: HIVE-3420
>                 URL: https://issues.apache.org/jira/browse/HIVE-3420
>             Project: Hive
>          Issue Type: Improvement
>          Components: HBase Handler
>         Environment: Hive-0.9.0 + HBase-0.94.1
>            Reporter: Gang Deng
>            Assignee: Navis
>            Priority: Critical
>         Attachments: HIVE-3420.D7311.1.patch
>
>   Original Estimate: 2h
>  Remaining Estimate: 2h
>
> When query hive with hbase rowkey range, hive map tasks do not leverage startrow, endrow information in tablesplit. For example, if the rowkeys fit into 5 hbase files, then where will be 5 map tasks. Ideally, each task will process 1 file. But in current implementation, each task processes 5 files repeatedly. The behavior not only waste network bandwidth, but also worse the lock contention in HBase block cache as each task have to access the same block. The problem code is in HiveHBaseTableInputFormat.convertFilte as below:
> ……
>     if (tableSplit != null) {
>       tableSplit = new TableSplit(
>         tableSplit.getTableName(),
>         startRow,
>         stopRow,
>         tableSplit.getRegionLocation());
>     }
>     scan.setStartRow(startRow);
>     scan.setStopRow(stopRow);
> ……
> As tableSplit already include startRow, endRow information of file, the better implementation will be:
>         ……
>         byte[] splitStart = startRow;
>         byte[] splitStop = stopRow;
>     if (tableSplit != null) {
>                 
>            if(tableSplit.getStartRow() != null){
>                         splitStart = startRow.length == 0 ||
>           Bytes.compareTo(tableSplit.getStartRow(), startRow) >= 0 ?
>             tableSplit.getStartRow() : startRow;
>                 }
>                 if(tableSplit.getEndRow() != null){
>                         splitStop = (stopRow.length == 0 ||
>           Bytes.compareTo(tableSplit.getEndRow(), stopRow) <= 0) &&
>           tableSplit.getEndRow().length > 0 ?
>             tableSplit.getEndRow() : stopRow;
>                 }                       
>       tableSplit = new TableSplit(
>         tableSplit.getTableName(),
>         splitStart,
>         splitStop,
>         tableSplit.getRegionLocation());
>     }
>     scan.setStartRow(splitStart);
>     scan.setStopRow(splitStop);
>         ……
> In my test, the changed code will improve performance more than 30%.

--
This message is automatically generated by JIRA.
If you think it was sent incorrectly, please contact your JIRA administrators
For more information on JIRA, see: http://www.atlassian.com/software/jira