[
https://issues.apache.org/jira/browse/ARROW-801?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15968091#comment-15968091
]
Jacques Nadeau commented on ARROW-801:
--------------------------------------
FixedWidthVector is a disjunct set from FieldVector. For example IntVector is a
FixedWidthVector but is not a FieldVector. What're you're asking for could be
done by implementing a new set of interfaces such as FixedFieldVector and
VariableFieldVector. But this begs the question of starting to rationalize why
we even have ValueVector versus FieldVector, and whether ValueVector should be
logical type-independent (defined solely by width). That is why I said what
you're asking for is really an iceberg to a larger problem.
> [JAVA] Provide direct access to underlying buffer memory addresses in
> consistent way without generating garbage or large amount indirections
> --------------------------------------------------------------------------------------------------------------------------------------------
>
> Key: ARROW-801
> URL: https://issues.apache.org/jira/browse/ARROW-801
> Project: Apache Arrow
> Issue Type: Bug
> Components: Java - Vectors
> Reporter: Jacques Nadeau
>
> When working with Arrow vectors recently, we observed a situation where our
> time was dominated by calls to getFieldBuffers() to be able to retrieve
> memory addresses (22s out of 26s total for a piece of code). We should
> provide a direct mechanism to access this data so we can avoid all the extra
> indirection and object creation.
> A proposal:
> getBitAddress();
> getDataAddress();
> getOffsetAddress();
> These interfaces would be made available at the FieldVector interface and
> simply throw UnsupportedOperationException where not supported.
> Unsupported Operations:
> data for list type
> offset for fixed width types
> data and offset for struct type
> data for union type
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
