[
https://issues.apache.org/jira/browse/FLINK-6216?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15957917#comment-15957917
]
ASF GitHub Bot commented on FLINK-6216:
---------------------------------------
Github user fhueske commented on a diff in the pull request:
https://github.com/apache/flink/pull/3646#discussion_r110041174
--- Diff:
flink-libraries/flink-table/src/test/scala/org/apache/flink/table/api/scala/stream/sql/WindowAggregateTest.scala
---
@@ -299,27 +299,6 @@ class WindowAggregateTest extends TableTestBase {
}
@Test
- def testGroupWithFloorExpression() = {
- val sql = "SELECT COUNT(*) FROM MyTable GROUP BY FLOOR(localTimestamp
TO HOUR)"
--- End diff --
I think we need to discuss with the community how to distinguish
non-windowed and windowed group aggregates. At the moment `GROUP BY
FLOOR(rowtime() TO HOUR)` is translated into a tumbling event time window. By
adding support for non-windows aggregates, it could also be executed as a such,
i.e., with early firing and late update instead of a final result when the
window is closed. The final result should be the same, but the behavior during
execution would be different.
I think a good approach would be to only translate group window functions
(TUMBLE, HOP, SESSION) into group windows and treat everything else as
non-windowed aggregation.
We should move this discussion to the mailing list though. I don't think we
have to wait for a decision for this issue to continue.
> DataStream unbounded groupby aggregate with early firing
> --------------------------------------------------------
>
> Key: FLINK-6216
> URL: https://issues.apache.org/jira/browse/FLINK-6216
> Project: Flink
> Issue Type: Sub-task
> Components: Table API & SQL
> Reporter: Shaoxuan Wang
> Assignee: Shaoxuan Wang
>
> Groupby aggregate results in a replace table. For infinite groupby aggregate,
> we need a mechanism to define when the data should be emitted (early-fired).
> This task is aimed to implement the initial version of unbounded groupby
> aggregate, where we update and emit aggregate value per each arrived record.
> In the future, we will implement the mechanism and interface to let user
> define the frequency/period of early-firing the unbounded groupby aggregation
> results.
> The limit space of backend state is one of major obstacles for supporting
> unbounded groupby aggregate in practical. Due to this reason, we suggest two
> common (and very useful) use-cases of this unbounded groupby aggregate:
> 1. The range of grouping key is limit. In this case, a new arrival record
> will either insert to state as new record or replace the existing record in
> the backend state. The data in the backend state will not be evicted if the
> resource is properly provisioned by the user, such that we can provision the
> correctness on aggregation results.
> 2. When the grouping key is unlimited, we will not be able ensure the 100%
> correctness of "unbounded groupby aggregate". In this case, we will reply on
> the TTL mechanism of the RocksDB backend state to evicted old data such that
> we can provision the correct results in a certain time range.
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
