[
https://issues.apache.org/jira/browse/HBASE-22980?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Shardul Singh updated HBASE-22980:
----------------------------------
Description:
*Problem:*
Partitioner class HRegionPartitioner in a HBase MapReduce job has a method
getPartition(). In getPartition(), there is a scenario where we have check for
less number of reducers than region. This scenario seems incorrect because for
a rowKey present in last region(let us say nth region) , getPartition() should
return value (n-1). But it is not returning n-1 for the last region as it is
falling in the case where number of reducers < number of regions and returning
some random value.
So if a client uses this class as a partitioner class in HBase MapReduce jobs,
this method incorrectly partitions the regions because rowKeys present in the
last regions does not fall to the last region.
[https://github.com/apache/hbase/blob/fbd5b5e32753104f88600b0f4c803ab5659bce64/hbase-mapreduce/src/main/java/org/apache/hadoop/hbase/mapreduce/HRegionPartitioner.java#L92]
Consider the following scenario:
if there are 5 regions for the table, partitions = 5 and number of reducers is
also 5.
So in this case above check for reducers < regions should not return true.
But for the last region when i=4(last region, 5th region) , getPartition should
return 4 but it returns 2 because it falls in the case of when we have less
reduces than region and returns true for the above condition even though we
have reducers = regions. So the condition is incorrect.
*Solution:*
Instead of
{code} if (i >= numPartitions-1) {code}
It should be
{code} if (i >= numPartitions){ {code}
was:
Partitioner class HRegionPartitioner in a HBase MapReduce job has a method
getPartition(). In getPartition(), there is a scenario where we have check for
less number of reducers than region. This scenario seems incorrect because for
a rowKey present in last region(let us say nth region) , getPartition() should
return value (n-1). But it is not returning n-1 for the last region as it is
falling in the case where number of reducers < number of regions and returning
some random value.
So if a client uses this class as a partitioner class in HBase MapReduce jobs,
this method incorrectly partitions the regions because rowKeys present in the
last regions does not fall to the last region.
[https://github.com/apache/hbase/blob/fbd5b5e32753104f88600b0f4c803ab5659bce64/hbase-mapreduce/src/main/java/org/apache/hadoop/hbase/mapreduce/HRegionPartitioner.java#L92]
Consider the following scenario:
if there are 5 regions for the table, partitions = 5 and number of reducers is
also 5.
So in this case above check for reducers < regions should not return true.
But for the last region when i=4(last region, 5th region) , getPartition should
return 4 but it returns 2 because it falls in the case of when we have less
reduces than region and returns true for the above condition even though we
have reducers = regions. So the condition is incorrect.
Solution:
Instead of
{code} if (i >= numPartitions-1) {code}
It should be
{code} if (i >= numPartitions){ {code}
> HRegionPartioner getPartition() method incorrectly partitions the regions of
> the table.
> ---------------------------------------------------------------------------------------
>
> Key: HBASE-22980
> URL: https://issues.apache.org/jira/browse/HBASE-22980
> Project: HBase
> Issue Type: Bug
> Components: mapreduce
> Reporter: Shardul Singh
> Assignee: Shardul Singh
> Priority: Major
>
> *Problem:*
> Partitioner class HRegionPartitioner in a HBase MapReduce job has a method
> getPartition(). In getPartition(), there is a scenario where we have check
> for less number of reducers than region. This scenario seems incorrect
> because for a rowKey present in last region(let us say nth region) ,
> getPartition() should return value (n-1). But it is not returning n-1 for the
> last region as it is falling in the case where number of reducers < number of
> regions and returning some random value.
> So if a client uses this class as a partitioner class in HBase MapReduce
> jobs, this method incorrectly partitions the regions because rowKeys present
> in the last regions does not fall to the last region.
> [https://github.com/apache/hbase/blob/fbd5b5e32753104f88600b0f4c803ab5659bce64/hbase-mapreduce/src/main/java/org/apache/hadoop/hbase/mapreduce/HRegionPartitioner.java#L92]
> Consider the following scenario:
> if there are 5 regions for the table, partitions = 5 and number of reducers
> is also 5.
> So in this case above check for reducers < regions should not return true.
> But for the last region when i=4(last region, 5th region) , getPartition
> should return 4 but it returns 2 because it falls in the case of when we have
> less reduces than region and returns true for the above condition even though
> we have reducers = regions. So the condition is incorrect.
>
> *Solution:*
> Instead of
> {code} if (i >= numPartitions-1) {code}
> It should be
> {code} if (i >= numPartitions){ {code}
--
This message was sent by Atlassian Jira
(v8.3.2#803003)
