[
https://issues.apache.org/jira/browse/FLINK-26004?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Lijie Wang updated FLINK-26004:
-------------------------------
Description:
If there are multiple consecutive the same groupBy(i.e. keyBy), SQL planner
will change them except the first one to use forward partitioner, so that these
operators can be chained to reduce unnecessary shuffles.
However, sometimes the consecutive hash operators are not chained (e.g.
multiple inputs), and this kind of forward partitioners will turn into forward
job edges. These forward edges still have the consecutive hash assumption, so
that they cannot be changed into rescale/rebalance edges, otherwise it can lead
to incorrect results. This prevents the adaptive batch scheduler from
determining parallelism for other forward edge downstream job vertices (see
FLINK-25046).
To solve it, I propose to introduce a new
{{{}ForwardForConsecutiveHashPartitioner{}}}. When SQL planner optimizes the
case of multiple consecutive the same groupBy, it should use the proposed
partitioner, so that the runtime framework can further decide whether the
partitioner can be changed to hash or not.
h4.
was:
If there are multiple consecutive the same groupBy(i.e. keyBy), SQL planner
will change them except the first one to use forward partitioner, so that these
operators can be chained to reduce unnecessary shuffles.
However, sometimes the consecutive hash operators are not chained (e.g.
multiple inputs), and this kind of forward partitioners will turn into forward
job edges. These forward edges still have the consecutive hash assumption, so
that they cannot be changed into rescale/rebalance edges, otherwise it can lead
to incorrect results. This prevents the adaptive batch scheduler from
determining parallelism for other forward edge downstream job vertices (see
FLINK-25046).
To solve it, I propose to introduce a new
{{{}ForwardForLocalKeyByPartitioner{}}}. When SQL planner optimizes the case of
multiple consecutive the same groupBy, it should use the proposed partitioner,
so that the runtime framework can further decide whether the partitioner can be
changed to hash or not.
h4.
> Introduce ForwardForConsecutiveHashPartitioner
> ----------------------------------------------
>
> Key: FLINK-26004
> URL: https://issues.apache.org/jira/browse/FLINK-26004
> Project: Flink
> Issue Type: Sub-task
> Reporter: Lijie Wang
> Priority: Major
> Labels: pull-request-available
>
> If there are multiple consecutive the same groupBy(i.e. keyBy), SQL planner
> will change them except the first one to use forward partitioner, so that
> these operators can be chained to reduce unnecessary shuffles.
> However, sometimes the consecutive hash operators are not chained (e.g.
> multiple inputs), and this kind of forward partitioners will turn into
> forward job edges. These forward edges still have the consecutive hash
> assumption, so that they cannot be changed into rescale/rebalance edges,
> otherwise it can lead to incorrect results. This prevents the adaptive batch
> scheduler from determining parallelism for other forward edge downstream job
> vertices (see FLINK-25046).
> To solve it, I propose to introduce a new
> {{{}ForwardForConsecutiveHashPartitioner{}}}. When SQL planner optimizes the
> case of multiple consecutive the same groupBy, it should use the proposed
> partitioner, so that the runtime framework can further decide whether the
> partitioner can be changed to hash or not.
> h4.
--
This message was sent by Atlassian Jira
(v8.20.1#820001)
