[
https://issues.apache.org/jira/browse/FLINK-6650?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16018620#comment-16018620
]
Fabian Hueske commented on FLINK-6650:
--------------------------------------
Thanks for starting this discussion. I think the current behavior is correct.
The result of a query should only depend on the operators of a query
({{select('a.sum)}}) and not on the method to convert the result into a stream.
The choice of {{toAppendStream()}} and {{toRetractStream()}} should only affect
the representation of the result but not the result itself.
Moreover, the result of a query on a streaming table must be identical to a
batch query on the same input data. Hence, the example in case 1 would not be
correct, because it should be a single row with a single value {{56}} as in
case 2 and not multiple rows.
The correct query to specify the desired of case 1 is the OVER query as shown
in the description. The result of an OVER query can be converted into an append
stream (unless we add / enable support for late data).
To summarize: The drawback of implementing this would be incorrect semantics
(batch != streaming) and unclear behavior (result of a query depends on the
conversion method).
> Fix Non-windowed group-aggregate error when using append-table mode.
> --------------------------------------------------------------------
>
> Key: FLINK-6650
> URL: https://issues.apache.org/jira/browse/FLINK-6650
> Project: Flink
> Issue Type: Sub-task
> Components: Table API & SQL
> Reporter: sunjincheng
> Assignee: sunjincheng
>
> When I test Non-windowed group-aggregate with {{stream.toTable(tEnv, 'a, 'b,
> 'c).select('a.sum, weightAvgFun('a, 'b)).toAppendStream[Row].addSink(new
> StreamITCase.StringSink)}}, I got the error as follows:
> {code}
> org.apache.flink.table.api.TableException: Table is not an append-only table.
> Output needs to handle update and delete changes.
> at
> org.apache.flink.table.api.StreamTableEnvironment.translate(StreamTableEnvironment.scala:631)
> at
> org.apache.flink.table.api.StreamTableEnvironment.translate(StreamTableEnvironment.scala:607)
> at
> org.apache.flink.table.api.scala.StreamTableEnvironment.toAppendStream(StreamTableEnvironment.scala:219)
> at
> org.apache.flink.table.api.scala.StreamTableEnvironment.toAppendStream(StreamTableEnvironment.scala:195)
> at
> org.apache.flink.table.api.scala.TableConversions.toAppendStream(TableConversions.scala:121)
> {code}
> The reason is {{DataStreamGroupAggregate#producesUpdates}} as follows:
> {code}
> override def producesUpdates = true
> {code}
> I think in the view of the user, what user want are(for example):
> Data:
> {code}
> val data = List(
> (1L, 1, "Hello"),
> (2L, 2, "Hello"),
> (3L, 3, "Hello"),
> (4L, 4, "Hello"),
> (5L, 5, "Hello"),
> (6L, 6, "Hello"),
> (7L, 7, "Hello World"),
> (8L, 8, "Hello World"),
> (20L, 20, "Hello World"))
> {code}
> * Case1:
> TableAPI
> {code}
> stream.toTable(tEnv, 'a, 'b, 'c).select('a.sum).toRetractStream[Row]
> .addSink(new StreamITCase.RetractingSink)
> {code}
> Result
> {code}
> 1
> 3
> 6
> 10
> 15
> 21
> 28
> 36
> 56
> {code}
> * Case 2:
> TableAPI
> {code}
> stream.toTable(tEnv, 'a, 'b, 'c).select('a.sum).toRetractStream[Row]
> .addSink(new StreamITCase.RetractingSink)
> {code}
> Result:
> {code}
> 56
> {code}
> In fact about #Case 1,we can using unbounded OVER windows, as follows:
> TableAPI
> {code}
> stream.toTable(tEnv, 'a, 'b, 'c, 'proctime.proctime)
> .window(Over orderBy 'proctime preceding UNBOUNDED_ROW as 'w)
> .select('a.sum over 'w)
> .toAppendStream[Row].addSink(new StreamITCase.StringSink)
> {code}
> Result
> {code}
> Same as #Case1
> {code}
> But after the [FLINK-6649 | https://issues.apache.org/jira/browse/FLINK-6649]
> OVER can not express the #Case1 with earlyFiring.
> So I still think that Non-windowed group-aggregate not always update-table,
> user can decide which mode to use.
> Is there any drawback to this improvement? Welcome anyone feedback?
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
