[
https://issues.apache.org/jira/browse/CALCITE-2645?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16672662#comment-16672662
]
Rui Wang edited comment on CALCITE-2645 at 11/2/18 5:04 PM:
------------------------------------------------------------
Correct me if I am wrong: in physical implementation, session window will only
emit one timestamp, which could be either window start or end, but not both. So
SESSION_END(ts), SESSION_START(ts) will have same value.
I think the Calcite implementation leaves no room for physical implementation
to correct SESSION_END(ts).
was (Author: amaliujia):
Correct me if I am wrong: in physical implementation, session window will only
emit one timestamp, which could be either window start or end, but now both. So
SESSION_END(ts), SESSION_START(ts) will have same value.
I think the Calcite implementation leaves no room for physical implementation
to correct SESSION_END(ts).
> SESSION_END() is transformed to RexInputRef, which is always equal to
> SESSION_START()
> -------------------------------------------------------------------------------------
>
> Key: CALCITE-2645
> URL: https://issues.apache.org/jira/browse/CALCITE-2645
> Project: Calcite
> Issue Type: Bug
> Reporter: Rui Wang
> Assignee: Julian Hyde
> Priority: Major
>
> https://github.com/apache/calcite/blob/master/core/src/main/java/org/apache/calcite/sql2rel/AuxiliaryConverter.java#L56
> In the AuxiliaryConverter implementation, SESSION_END() is transformed to a
> RexInputRef, as the same as SESSION_START(). By doing so, it cannot restored
> when generating physical plan because it cannot be differentiated which
> RexInputRef is session start, and which RexInputRef is session end.
> Ideally, during toRel, SESSION_END can be kept as a RexCall with function
> information in it. So in the following step, the RexCall can be replaced as a
> correct input reference to GROUP BY's field.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
