[
https://issues.apache.org/jira/browse/CALCITE-4125?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17161430#comment-17161430
]
Rui Wang commented on CALCITE-4125:
-----------------------------------
>From relation perspective, yes those proposed TVF produces relations and then
>relations are joined. And each relation has window_start and window_end and
>people can decide to how/when to join based on window metadata.
This JIRA comes from underlying engine perspective. E.g. Apache Beam and Apache
Flink, both more or less following Cloud Dataflow model:
https://research.google/pubs/pub43864/
For engines that follow Cloud Dataflow model, time-varying relations are mapped
to PCollection, and for unbounded PCollection (aka. stream), it has associated
streaming windowing strategy thus two unbounded PCollection joins do have a
streaming windowing strategy compatible problem. E.g. TUMBLE unbounded
PCollection join HOP unbounded PCollection.
> Stream Join
> -----------
>
> Key: CALCITE-4125
> URL: https://issues.apache.org/jira/browse/CALCITE-4125
> Project: Calcite
> Issue Type: Sub-task
> Reporter: Rui Wang
> Priority: Major
>
> Right now stream JOIN is supported in syntax (e.g. in planner), however,
> there are more areas needs a enhancement:
> 1. Does windowing strategy in JOIN have a constraint? For example, is a
> TUMBLE stream allowed to join a HOP stream?
> 2. If there is any constraints on windowing strategy in JOIN, how will that
> affect JOIN reordering?
> 3. How to optimize stream join stream or stream join table, which, by nature,
> should be different from table join table?
--
This message was sent by Atlassian Jira
(v8.3.4#803005)
