hive-dev mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Sergey Shelukhin (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (HIVE-6028) Partition predicate literals are not interpreted correctly.
Date Fri, 13 Dec 2013 23:56:08 GMT

    [ https://issues.apache.org/jira/browse/HIVE-6028?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=13848079#comment-13848079
] 

Sergey Shelukhin commented on HIVE-6028:
----------------------------------------

Yeah nm I built 12 and can repro. Yeah it seems to be the root cause of the issue; 12 does
validation on the client and apparently doesn't check that types are compatible, only that
both are supported (this is the root cause). Then it uses getPartitionsByFilter to query partitions
and filter being sent (expression.getExprString) is changed to "hour = 0", which fails.
Of course you could say that SemanticAnalyzer should have caught this earlier, what is 00?
It doesn't look like it should be treated like a valid string to me.

Trunk uses getPartitionsByExpr, and validates on server that column and value type are compatible.
As far as I understand hive doesn't do sub-releases; how should this jira be handled, closed
as already fixed in 13? 
Or do we want to change SemanticAnalyzer to be less lenient?



> Partition predicate literals are not interpreted correctly.
> -----------------------------------------------------------
>
>                 Key: HIVE-6028
>                 URL: https://issues.apache.org/jira/browse/HIVE-6028
>             Project: Hive
>          Issue Type: Bug
>    Affects Versions: 0.12.0
>            Reporter: Pala M Muthaia
>         Attachments: Hive-6028-explain-plan.txt
>
>
> When parsing/analyzing query, hive treats partition predicate value as int instead of
string. This breaks down and leads to incorrect result when the partition predicate value
starts with int 0, e.g: hour=00, hour=05 etc.
> The following repro illustrates the bug:
> -- create test table and partition, populate with some data
> create table test_partition_pred(col1 int) partitioned by (hour STRING);
> insert into table test_partition_pred partition (hour=00) select 21 FROM  some_table
limit 1;
> -- this query returns incorrect results, i.e. just empty set.
> select * from test_partition_pred where hour=00;
> OK
> -- this query returns correct result. Note predicate value is string literal
> select * from test_partition_pred where hour='00';
> OK
> 21	00
> explain plan illustrates how the query was interpreted. Particularly the partition predicate
is pushed down as regular filter clause, with hour=0 as predicate. See attached explain plan
file.
> Note:
> 1. The type of the partition column is defined as string, not int.
> 2. This is a regression in Hive 0.12. This used to work in Hive 0.11
> 3. Not an issue when the partition value starts with integer other than 0, e.g hour=10,
hour=11 etc.
> 4. As seen above, workaround is to use string literal hour='00' etc.
> This should not be too bad if in the failing case hive complains that partition hour=0
is not found, or complains literal type doesn't match column type. Instead hive silently pushes
it down as filter clause, and query succeeds with empty set as result.
> We found this out in our production tables partitioned by hour, only a few days after
it started occurring, when there were empty data sets for partitions hour=00 to hour=09.



--
This message was sent by Atlassian JIRA
(v6.1.4#6159)

Mime
View raw message