[
https://issues.apache.org/jira/browse/HBASE-19047?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16222789#comment-16222789
]
Appy commented on HBASE-19047:
------------------------------
bq. This was done as part of another comment fix. This is rare chance that the
post hook just close the core scanner being passed to it and create a scanner
using region#getScanner() and return back. Even if this happens and we dont
take that as shipper now, what happens, we can not cal shipped after every RPC.
Still there wont be a func issue. We will delay the ref count decr on blocks
until the scanner close happens. So I am ok to remove this extra 'shipper '
assignment.
Your explanation gave me more context. Based on it, I feel like real issue here
is wrong ownership. If you want to keep the "shipper" around even if the
original scanner is destroyed, than the scanner, neither interface nor any
implementation, should "own" it. It should probably be separate.
> CP exposed Scanner types should not extend Shipper
> --------------------------------------------------
>
> Key: HBASE-19047
> URL: https://issues.apache.org/jira/browse/HBASE-19047
> Project: HBase
> Issue Type: Sub-task
> Components: Coprocessors
> Reporter: Anoop Sam John
> Assignee: Anoop Sam John
> Priority: Critical
> Fix For: 2.0.0-alpha-4
>
> Attachments: HBASE-19047.patch, HBASE-19047_V2.patch,
> HBASE-19047_V2.patch, HBASE-19047_V3.patch, HBASE-19047_V4.patch,
> HBASE-19047_V4.patch, HBASE-19047_V4.patch
>
>
> Shipper is a IA.Private interface and very much internal..
> Right now CP exposed RegionScanner is extending this and so exposing the
> shipped() method. This by mistake is called, can harm the correctness of the
> cells in the Results.
> preScannerOpen() allowing to return a new Scanner is also problematic now.
> This can allow users to create a Region scanner from Region and then wrap it
> and return back (Well same can be done by postScannerOpen also), it can so
> happen that the wrapper is not implementing the shipped() properly. In any
> way exposing the shipped () is problematic.
> Solution Steps
> 1. Remove preScannerOpen() , the use case I can think of is wrapping the
> original scanner. The original scanner can be created by Region.getScanner
> way only.. May be no need to remove this hook. Just remove the ability for
> it to return a RegionScanner instance. Call this with the Scan object and
> the CP can change the Scan object if they want.
> 2. Let RegionScanner not extending Shipper but only RegionScannerImpl
> implements this
> 3. We have ref to the RegionScanner created by core and let that be used by
> RegionScannerShippedCallBack when the post hook doing a wrap.
--
This message was sent by Atlassian JIRA
(v6.4.14#64029)
