[
https://issues.apache.org/jira/browse/HIVE-9645?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=14392336#comment-14392336
]
Gopal V commented on HIVE-9645:
-------------------------------
[~ashutoshc]: The latest patch is still breaking type-checks all over the place
when I test it.
We can start using VOID types across the board, but it would effectively drop
all compatibility for hive-1.0 UDFs & UDAFs.
{code}
explain select * from
(select trim(concat("", ss_sold_date_sk)) from store_sales a
UNION ALL
select trim(concat("", ss_sold_date_sk)) from store_sales b where
ss_sold_date_sk is null)
x
limit 2
;
{code}
Every user-deployed UDF out there which is strict-typed would need a rewrite to
support this patch & the type-shifting has the ability to break vectorization
on the reducer side, because the key types won't match between ReduceSinks on a
JOIN.
To keep UDF compatibility, you have to resolve the {{ss_sold_date_sk}} to a
{{cast(null as bigint)}} and refuse to constant-fold casts to VOID, instead
generating a {{new JavaConstantLongObjectInspector(null);}} its place.
That seems to be a good way to keep the schema from changing, but I will defer
to [~jpullokkaran] on how the CBO layer will treat this case, but UDF compat
might be worth considering now that we're already 1.0.
> Constant folding case NULL equality
> -----------------------------------
>
> Key: HIVE-9645
> URL: https://issues.apache.org/jira/browse/HIVE-9645
> Project: Hive
> Issue Type: Bug
> Components: Logical Optimizer
> Affects Versions: 0.14.0, 1.0.0, 1.1.0
> Reporter: Gopal V
> Assignee: Ashutosh Chauhan
> Attachments: HIVE-9645.1.patch, HIVE-9645.2.patch, HIVE-9645.3.patch,
> HIVE-9645.patch
>
>
> Hive logical optimizer does not follow the Null scan codepath when
> encountering a NULL = 1;
> NULL = 1 is not evaluated as false in the constant propogation implementation.
> {code}
> hive> explain select count(1) from store_sales where null=1;
> ...
> TableScan
> alias: store_sales
> filterExpr: (null = 1) (type: boolean)
> Statistics: Num rows: 550076554 Data size: 49570324480
> Basic stats: COMPLETE Column stats: COMPLETE
> Filter Operator
> predicate: (null = 1) (type: boolean)
> Statistics: Num rows: 275038277 Data size: 0 Basic stats:
> PARTIAL Column stats: COMPLETE
> {code}
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
