[
https://issues.apache.org/jira/browse/HBASE-6104?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13860623#comment-13860623
]
James Taylor commented on HBASE-6104:
-------------------------------------
It's about conserving network bandwidth - we don't want to take the hit of
transferring the same data between client and server multiple times. For
example, with secondary indexing, we'd be tacking on data for every Put - if
you have a batch of 10,000, that's a lot of extra data. We could try to figure
out which Put is the "first one" for each region, but what if a split occurs
after we figure this out -- this seems too brittle.
In the case of a Hash Join, we'd be sending over the compressed results of a
scan that ran over the smaller table (which gets joined against in a
coprocessor when the scan over the other table is ran). This can become very
large - imagine you're joining against a table with 10M rows. We would not want
to send this data for every region of the region server (or even multiple times
per region depending on how the scan gets parallelized on the client).
> Require EXEC permission to call coprocessor endpoints
> -----------------------------------------------------
>
> Key: HBASE-6104
> URL: https://issues.apache.org/jira/browse/HBASE-6104
> Project: HBase
> Issue Type: New Feature
> Components: Coprocessors, security
> Reporter: Gary Helmling
> Assignee: Andrew Purtell
> Fix For: 0.99.0
>
> Attachments: 6104-addendum-1.patch, 6104-revert.patch, 6104.patch,
> 6104.patch, 6104.patch, 6104.patch, 6104.patch, 6104.patch
>
>
> The EXEC action currently exists as only a placeholder in access control. It
> should really be used to enforce access to coprocessor endpoint RPC calls,
> which are currently unrestricted.
> How the ACLs to support this would be modeled deserves some discussion:
> * Should access be scoped to a specific table and CoprocessorProtocol
> extension?
> * Should it be possible to grant access to a CoprocessorProtocol
> implementation globally (regardless of table)?
> * Are per-method restrictions necessary?
> * Should we expose hooks available to endpoint implementors so that they
> could additionally apply their own permission checks? Some CP endpoints may
> want to require READ permissions, others may want to enforce WRITE, or READ +
> WRITE.
> To apply these kinds of checks we would also have to extend the
> RegionObserver interface to provide hooks wrapping HRegion.exec().
--
This message was sent by Atlassian JIRA
(v6.1.5#6160)
