[
https://issues.apache.org/jira/browse/FLINK-5342?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15762989#comment-15762989
]
godfrey he edited comment on FLINK-5342 at 12/20/16 8:51 AM:
-------------------------------------------------------------
I think it's difficult and unintuitive about tuning the parallelism of all
operators. User does not know the target operators translated from a query. And
when more calcite rules added, the execution graph will be different with
before for the same query. So as a first step, I think we can give
coarse-grained control. After have statistics about the input and more correct
cost of each operator, we can set the parallelism for each operator base on
cost model.
was (Author: godfreyhe):
I think it's difficult and unintuitive about tuning the parallelism of all
operators. User does not know the target operators translated from a query.
When more calcite rules added, the execution graph will be different with
before for the same query. So as a first step, I think we can give
coarse-grained control. After we have statistics about the input and more
correct cost of each operator, we can set the parallelism for each operator
base on cost model.
> Setting the parallelism automatically for operators base on cost model
> ----------------------------------------------------------------------
>
> Key: FLINK-5342
> URL: https://issues.apache.org/jira/browse/FLINK-5342
> Project: Flink
> Issue Type: Improvement
> Components: Table API & SQL
> Reporter: godfrey he
>
> On Flink table API, a query will be translated to operators without
> parallelism. And user do not know even do not care the target operators
> translated from query. So it's better to set the parallelism automatically
> for each operator base on cost model.
--
This message was sent by Atlassian JIRA
(v6.3.4#6332)
