phoenix-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Ankit Singhal (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (PHOENIX-4018) HashJoin may produce nulls for LHS table columns
Date Fri, 14 Jul 2017 09:20:00 GMT

    [ https://issues.apache.org/jira/browse/PHOENIX-4018?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16087074#comment-16087074
] 

Ankit Singhal commented on PHOENIX-4018:
----------------------------------------

Thanks, [~sergey.soldatov] for the patch. The problem may not be visible with other region
scanner probably because we override BaseRegionScanner#next(List<Cell>,ScannerContext)
to call BaseRegionScanner#next(List<Cell>). 

This might already be impacting scanner Heartbeat, though we have a thread at the client which
is renewing scanners but we should fix this too.

And, if we are already avoiding RPC chunking at the cell level, so IMO to leverage "hbase.client.scanner.max.result.size"
(applied at row level), We should be incrementing limits in scannerContext while we finish
with all the region scanners because there will be other wrapped regionScanner(like OffsetRegionScanner)
which might be skipping rows subsequently. Probably the best place would be BaseScannerRegionObserver.RegionScannerHolder#next(List<Cell>,ScannerContext).


If you think it is of huge effort and testing let's do it in a separate JIRA and we can close
this by just removing scannerContext from next() call for now.



> HashJoin may produce nulls for LHS table columns
> ------------------------------------------------
>
>                 Key: PHOENIX-4018
>                 URL: https://issues.apache.org/jira/browse/PHOENIX-4018
>             Project: Phoenix
>          Issue Type: Bug
>    Affects Versions: 4.11.0
>            Reporter: Sergey Soldatov
>            Assignee: Sergey Soldatov
>            Priority: Critical
>         Attachments: PHOENIX-4018-1.patch
>
>
> Here is the problem: in HashJoinRegionScanner methods (nextRow for example) we are using
the same scanner context that was created in RSRpcServices. It has limits (i.e. 2Mb size).
Let's say that we have 3Mb region and the only key that match the join condition is located
at the end of the region. In HashJoinRegionScanner#nextRow when we iterate through the region
rows once we reached the limit of 2Mb, every region scanner nextRow will  return a single
cell and the scanner context will have SIZE_LIMIT_REACHED_MID_ROW state. But we don't have
any logic that check that, so this single cell is considered as a complete row with all nulls
except one column. 
> How to fix it: 
> 1. for region scanner we may provide NoLimitScannerContext, so we will never get a partial
result.  
> 2. We need to update the scanner context that we got from RSRpcServices with the real
data, basing on the size of results we are going to return. 



--
This message was sent by Atlassian JIRA
(v6.4.14#64029)

Mime
View raw message