flink-issues mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Fabian Hueske (JIRA)" <j...@apache.org>
Subject [jira] [Commented] (FLINK-5047) Add sliding group-windows for batch tables
Date Thu, 10 Nov 2016 12:49:59 GMT

    [ https://issues.apache.org/jira/browse/FLINK-5047?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15653968#comment-15653968

Fabian Hueske commented on FLINK-5047:

Right, the third approach does only work for combinable aggregates and reduces the amount
of replicated data because only pre-aggregates are replicated. I'd prefer it over approach
2 because it is easier to implement (it extends approach 1) than approach 2 which would require
an implementation for SQL Window.

> Add sliding group-windows for batch tables
> ------------------------------------------
>                 Key: FLINK-5047
>                 URL: https://issues.apache.org/jira/browse/FLINK-5047
>             Project: Flink
>          Issue Type: Sub-task
>          Components: Table API & SQL
>            Reporter: Jark Wu
> Add Slide group-windows for batch tables as described in [FLIP-11|https://cwiki.apache.org/confluence/display/FLINK/FLIP-11%3A+Table+API+Stream+Aggregations].
> There are two ways to implement sliding windows for batch:
> 1. replicate the output in order to assign keys for overlapping windows. This is probably
the more straight-forward implementation and supports any aggregation function but blows up
the data volume.
> 2. if the aggregation functions are combinable / pre-aggregatable, we can also find the
largest tumbling window size from which the sliding windows can be assembled. This is basically
the technique used to express sliding windows with plain SQL (GROUP BY + OVER clauses). For
a sliding window Slide(10 minutes, 2 minutes) this would mean to first compute aggregates
of non-overlapping (tumbling) 2 minute windows and assembling consecutively 5 of these into
a sliding window (could be done in a MapPartition with sorted input). The implementation could
be done as an optimizer rule to split the sliding aggregate into a tumbling aggregate and
a SQL WINDOW operator. Maybe it makes sense to implement the WINDOW clause first and reuse
this for sliding windows.
> 3. There is also a third, hybrid solution: Doing the pre-aggregation on the largest non-overlapping
windows (as in 2) and replicating these results and processing those as in the 1) approach.
The benefits of this is that it a) is based on the implementation that supports non-combinable
aggregates (which is required in any case) and b) that it does not require the implementation
of the SQL WINDOW operator. Internally, this can be implemented again as an optimizer rule
that translates the SlidingWindow into a pre-aggregating TublingWindow and a final SlidingWindow
(with replication).
> see FLINK-4692 for more discussion

This message was sent by Atlassian JIRA

View raw message