[
https://issues.apache.org/jira/browse/DRILL-3477?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14619768#comment-14619768
]
Jacques Nadeau commented on DRILL-3477:
---------------------------------------
Hmm... wondering what other ramifications are.
- Need to modify c++ client (I guess)
- Wonder about how it should be exposed in ODBC and JDBC. (or before?)
I'm wondering what issues we'll uncover if we change this. Did testing all go
through with a hiccup? Although, I'm not sure what negative test-cases we have
about selecting null columns. I am also wondering about overhead.
> Using IntVector for null expressions causes problems with implicit cast
> -----------------------------------------------------------------------
>
> Key: DRILL-3477
> URL: https://issues.apache.org/jira/browse/DRILL-3477
> Project: Apache Drill
> Issue Type: Bug
> Reporter: Steven Phillips
> Assignee: Jinfeng Ni
>
> See DRILL-3353, for example.
> A simple example is this:
> {code}
> select * from t where a = 's';
> {code}
> If the first batch scanned from table t does not contain the column a, the
> expression materializer in Project defaults to Nullable Int as the type. The
> Filter then sees an Equals expression between a VarChar and an Int type, so
> it does an implicit cast. Implicit cast rules give Int higher precedence, so
> the literal 's' is cast to Int, which ends up throwing a
> NumberFormatException.
> In the class ResolverTypePrecedence, we see that Null type has the lowest
> precedence, which makes sense. But since we don't actually currently have an
> implementation for NullVector, we should materialize the Null type as the
> Vector with the lowest possible precedence, which is VarBinary.
> My suggestion is that we should use VarBinary as the default type in
> ExpressionMaterializer instead of Int.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
