[
https://issues.apache.org/jira/browse/FLINK-6075?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15948809#comment-15948809
]
Fabian Hueske commented on FLINK-6075:
--------------------------------------
Hi [~rtudoran],
Calcite added support for group window functions (TUMBLE, HOP, SESSION) with
its latest 1.12 release. We will upgrade our dependency soon FLINK-5829.
Regarding the semantics of the queries we have to stick to the batch semantics.
As you noticed, some queries are not accepted by Calcite (Q1: does not group on
{{a}}, Q3: does not allow LIMIT in the OVER clause).
I think we can support the following queries:
ORDER BY:
{code}
SELECT * FROM t ORDER BY t.rowtime; // order stream by event-time
SELECT * FROM t ORDER BY t.rowtime, a; // order stream primarily by event-time,
then by a
SELECT * FROM t ORDER BY t.proctime; // order stream by proc-time
{code}
ORDER BY does only make sense for append-only tables. The primary sort
attribute must be the timestamp and must be ASC.
LIMIT / FETCH FIRST / OFFSET
{code}
SELECT * FROM t ORDER BY t.rowtime LIMIT 10; // the last 10 rows in event-time.
Requires support for retraction to remove older rows.
SELECT * FROM t ORDER BY t.proctime LIMIT 10; // the last 10 rows in proc-time.
Requires support for retraction to remove older rows.
SELECT MIN(a) AS minA FROM t GROUP BY b ORDER BY minA ASC LIMIT 10; //
Maintains the rows with the 10 smallest minA values. This requires retraction
and support for aggregations with out windows.
{code}
Each query with a LIMIT / FETCH FIRST / OFFSET returns an update table, because
the query needs to retract previously emitted results. Hence these features
depends on FLINK-6047.
> Support Limit/Top(Sort) for Stream SQL
> --------------------------------------
>
> Key: FLINK-6075
> URL: https://issues.apache.org/jira/browse/FLINK-6075
> Project: Flink
> Issue Type: New Feature
> Components: Table API & SQL
> Reporter: radu
> Labels: features
> Attachments: sort.png
>
>
> These will be split in 3 separated JIRA issues. However, the design is the
> same only the processing function differs in terms of the output. Hence, the
> design is the same for all of them.
> Time target: Proc Time
> **SQL targeted query examples:**
> *Sort example*
> Q1)` SELECT a FROM stream1 GROUP BY HOP(proctime, INTERVAL '1' HOUR, INTERVAL
> '3' HOUR) ORDER BY b`
> Comment: window is defined using GROUP BY
> Comment: ASC or DESC keywords can be placed to mark the ordering type
> *Limit example*
> Q2) `SELECT a FROM stream1 WHERE rowtime BETWEEN current_timestamp - INTERVAL
> '1' HOUR AND current_timestamp ORDER BY b LIMIT 10`
> Comment: window is defined using time ranges in the WHERE clause
> Comment: window is row triggered
> *Top example*
> Q3) `SELECT sum(a) OVER (ORDER BY proctime RANGE INTERVAL '1' HOUR PRECEDING
> LIMIT 10) FROM stream1`
> Comment: limit over the contents of the sliding window
> General Comments:
> -All these SQL clauses are supported only over windows (bounded collections
> of data).
> -Each of the 3 operators will be supported with each of the types of
> expressing the windows.
> **Description**
> The 3 operations (limit, top and sort) are similar in behavior as they all
> require a sorted collection of the data on which the logic will be applied
> (i.e., select a subset of the items or the entire sorted set). These
> functions would make sense in the streaming context only in the context of a
> window. Without defining a window the functions could never emit as the sort
> operation would never trigger. If an SQL query will be provided without
> limits an error will be thrown (`SELECT a FROM stream1 TOP 10` -> ERROR).
> Although not targeted by this JIRA, in the case of working based on event
> time order, the retraction mechanisms of windows and the lateness mechanisms
> can be used to deal with out of order events and retraction/updates of
> results.
> **Functionality example**
> We exemplify with the query below for all the 3 types of operators (sorting,
> limit and top). Rowtime indicates when the HOP window will trigger – which
> can be observed in the fact that outputs are generated only at those moments.
> The HOP windows will trigger at every hour (fixed hour) and each event will
> contribute/ be duplicated for 2 consecutive hour intervals. Proctime
> indicates the processing time when a new event arrives in the system. Events
> are of the type (a,b) with the ordering being applied on the b field.
> `SELECT a FROM stream1 HOP(proctime, INTERVAL '1' HOUR, INTERVAL '2' HOUR)
> ORDER BY b (LIMIT 2/ TOP 2 / [ASC/DESC] `)
> ||Rowtime|| Proctime|| Stream1|| Limit 2|| Top 2|| Sort
> [ASC]||
> | |10:00:00 |(aaa, 11) | | |
> |
> | |10:05:00 |(aab, 7) | | | |
> |10-11 |11:00:00 | | aab,aaa |aab,aaa | aab,aaa
> |
> | |11:03:00 |(aac,21) | | | |
>
> |11-12 |12:00:00 | | aab,aaa |aab,aaa | aab,aaa,aac|
> | |12:10:00 |(abb,12) | | | |
>
> | |12:15:00 |(abb,12) | | | |
>
> |12-13 |13:00:00 | | abb,abb | abb,abb |
> abb,abb,aac|
> |...|
> **Implementation option**
> Considering that the SQL operators will be associated with window boundaries,
> the functionality will be implemented within the logic of the window as
> follows.
> * Window assigner – selected based on the type of window used in SQL
> (TUMBLING, SLIDING…)
> * Evictor/ Trigger – time or count evictor based on the definition of the
> window boundaries
> * Apply – window function that sorts data and selects the output to trigger
> (based on LIMIT/TOP parameters). All data will be sorted at once and result
> outputted when the window is triggered
> An alternative implementation can be to use a fold window function to sort
> the elements as they arrive, one at a time followed by a flatMap to filter
> the number of outputs.
> !sort.png!
> **General logic of Join**
> ```
> inputDataStream.window(new [Slide/Tumble][Time/Count]Window())
> //.trigger(new [Time/Count]Trigger()) – use default
> //.evictor(new [Time/Count]Evictor()) – use default
> .apply(SortAndFilter());
> ```
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
