[
https://issues.apache.org/jira/browse/HIVE-14893?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Sergey Shelukhin updated HIVE-14893:
------------------------------------
Description:
See the results for vectorized in decimal_11 test added in HIVE-14863.
We cast decimal to various int types; the cast is specialized for each type on
non-vectorized side; on vectorized side, it's only specialized for
LongColumnVector, so all the decimals get converted to longs. LongColumnVector
gets converted to a proper type in some other mysterious place later, and
tiny/small/regular ints become truncated at that point.
Logically, I am not sure if every vectorized expression should be aware of the
underlying type for the LongColumnVector (that seems implausible - I am not
sure if type information is even available, and if yes it doesn't look like
it's used in other places), or if the long-to-smaller-type automatic conversion
should be fixed to produce nulls on overflow.
However it seems like a good idea to do the latter in any case, to have a
catch-all for all the vectorized expressions that might treat LongCV as
representing longs at all times.
Update - I see 10s of places in the code where it does something like this:
(int) ((LongColumnVector) batch.cols[projectionColumnNum]).vector[adjustedIndex]
Also for other types. These might all be problematic.
was:
See the results for vectorized in decimal_11 test added in HIVE-14863.
We cast decimal to various int types; the cast is specialized for each type on
non-vectorized side; on vectorized side, it's only specialized for
LongColumnVector, so all the decimals get converted to longs. LongColumnVector
gets converted to a proper type in some other mysterious place later, and
tiny/small/regular ints become truncated at that point.
Logically, I am not sure if every vectorized expression should be aware of the
underlying type for the LongColumnVector (that seems implausible - I am not
sure if type information is even available, and if yes it doesn't look like
it's used in other places), or if the long-to-smaller-type automatic conversion
should be fixed to produce nulls on overflow.
However it seems like a good idea to do the latter in any case, to have a
catch-all for all the vectorized expressions that might treat LongCV as
representing longs at all times.
> vectorized execution may convert LongCV to smaller types incorrectly
> --------------------------------------------------------------------
>
> Key: HIVE-14893
> URL: https://issues.apache.org/jira/browse/HIVE-14893
> Project: Hive
> Issue Type: Bug
> Reporter: Sergey Shelukhin
> Assignee: Matt McCline
> Priority: Critical
>
> See the results for vectorized in decimal_11 test added in HIVE-14863.
> We cast decimal to various int types; the cast is specialized for each type
> on non-vectorized side; on vectorized side, it's only specialized for
> LongColumnVector, so all the decimals get converted to longs.
> LongColumnVector gets converted to a proper type in some other mysterious
> place later, and tiny/small/regular ints become truncated at that point.
> Logically, I am not sure if every vectorized expression should be aware of
> the underlying type for the LongColumnVector (that seems implausible - I am
> not sure if type information is even available, and if yes it doesn't look
> like it's used in other places), or if the long-to-smaller-type automatic
> conversion should be fixed to produce nulls on overflow.
> However it seems like a good idea to do the latter in any case, to have a
> catch-all for all the vectorized expressions that might treat LongCV as
> representing longs at all times.
> Update - I see 10s of places in the code where it does something like this:
> (int) ((LongColumnVector)
> batch.cols[projectionColumnNum]).vector[adjustedIndex]
> Also for other types. These might all be problematic.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
