[ 
https://issues.apache.org/jira/browse/IMPALA-9699?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
 ]

Tamas Mate resolved IMPALA-9699.
--------------------------------
    Resolution: Resolved

> Skip '-1' values when aggregating num_null incremental statistics
> -----------------------------------------------------------------
>
>                 Key: IMPALA-9699
>                 URL: https://issues.apache.org/jira/browse/IMPALA-9699
>             Project: IMPALA
>          Issue Type: Improvement
>          Components: Catalog
>    Affects Versions: Impala 3.3.0
>            Reporter: Tamas Mate
>            Assignee: Tamas Mate
>            Priority: Major
>              Labels: backwards-compatibility
>
> IMPALA-7659 added the population of NULL counts while computing stats, later 
> IMPALA-8566 fixed an accuracy issue caused by the initialization of 
> statistics. The initial value was changed from '-1' to '0'. The fix also 
> contained a slight change on how the values are being summarized. Earlier the 
> negative values were excluded from the summary:
> {code:java}
> if (num_new_nulls >= 0) num_nulls += num_new_nulls;
> {code}
> while in the new implementation, as these values should not be negative, the 
> condition was removed:
> {code:java}
> num_nulls += num_new_nulls;
> {code}
> This change does not cause any problem for stats created after this fix, 
> however it can make table metadata unavailable between earlier and newer 
> releases. The metadata can be invalid if a compute incremental stats is 
> issued on a partition because the '-1' values can decrease the column level 
> num_nulls under '-1'. Later a smaller than '-1' num_null will fail on a 
> precondition check when CatalogD is trying to fetch the table metadata.
> The condition should not cause any problem and due to backward compatibility 
> reasons we should put it back.



--
This message was sent by Atlassian Jira
(v8.3.4#803005)

---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]

Reply via email to