[HBASE-5982] HBase Coprocessor Local Aggregation

• Type: Improvement
• Status: Closed
• Priority: Major
• Resolution: Incomplete
• Affects Version/s: 0.92.1
• Component/s: Coprocessors
• Labels: Coprocessor

Environment

cloudera-cdh3u3,hbase-0.92.1

• Assignee: Unassigned
• Reporter: dengpeng
• Created: 10/May/12 02:48
• Updated: 13/Jun/22 16:22
• Resolved: 11/Nov/13 20:21

Description

In our application, we need to handle the following SQL-like process on hbase. There are very complex processes on each region, and the result of 'top #' from each region will be sent back to the coprocessor client in the current region-based endpoint framework.

Let's take the following SQL as an example. Suppose there are 100 regions in each RS and there are 100 RSs in the cluster, the client will receive 100*100*1M = 10G records from all the region, and then select top 1M records from 10G records. The client need much RAM to handle these data and the network of the cluster maybe the bottleneck.

If we have the RS-based endpoint, each RS will handle parts of result from its regions, the client will receive 100*1M = 0.1G records. The burden of the client and the network will dramatically reduced.

example:
select top 1000000 count(1) as A , sum(intRxlevDL)/count(intRxlevDL) as B , intBscPc as bscPc , intLac as LAC , intCI as CI from ftbMrMsg t1 where ( t1.dtTime >= '2012-03-02 04:00:00.000' and t1.dtTime < '2012-03-02 05:00:00.000' )group by bscPc , LAC , CI having B >= 0.2order by bscPc ASC , LAC ASC , CI ASC

So far, the network is a bottleneck in our application when using coprocessor to handle the above SQL. I think the RS-based Endpoint is worth doing, especially for the 'top #' process. What's your opinion about this? I think we can open a jira.