[jira] [Comment Edited] (FLINK-11974) Introduce StreamOperatorFactory to help table perform the whole Operator CodeGen

["Piotr Nowojski (JIRA)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/FLINK-11974?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16826040#comment-16826040 ] 

Piotr Nowojski edited comment on FLINK-11974 at 4/25/19 12:58 PM:
------------------------------------------------------------------

Yes, I think so. I would see two questions/problems to your poc, for which I don't have a clear answer yet:
 1. Those {{if (node.getOperatorFactory() instanceof SimpleOperatorFactory)}} are not pretty and they prevent from using the combination of non {{SimpleOperatorFactory}} with for example {{AbstractUdfStreamOperator}} or {{OutputTypeConfigurable}}. Could we express the same logic within {{StreamOperatorFactory}} framework? For example replacing
{code:java}
if (taskOperatorFactory instanceof SimpleOperatorFactory) {
	StreamOperator operatorObject = ((SimpleOperatorFactory) taskOperatorFactory).getOperator();
	if (operatorObject instanceof OutputTypeConfigurable && outTypeInfo != null) {
		...
	}
}
{code}
with:
{code:java}
if (taskOperatorFactory.isOutputTypeConfigurable() && outTypeInfo != null) {
}
{code}
or
{code:java}
if (taskOperatorFactory instanceof OutputTypeConfigurableFactory && outTypeInfo != null) {
}
{code}
2. {{StreamOperatorFactory#createStreamOperator(...)}} could return already "setup" operator. We could for example move the {{StreamOperator#setup(...)}} method to a new interface {{SetupableStreamOperator}} and {{SetupableStreamOperator}} would be only used in {{SimpleOperatorFactory}}. However I'm not sure what would be the next clean up step - for example how much effort and how to actually get rid of {{#setup()}} method from {{AbstractStreamOperator}}?


was (Author: pnowojski):
Yes, I think so. I would see two questions/problems to your poc, for which I don't have a clear answer yet:
 1. Those {{if (node.getOperatorFactory() instanceof SimpleOperatorFactory)}} are not pretty and they prevent from using the combination of non {{SimpleOperatorFactory}} with for example {{AbstractUdfStreamOperator}} or {{OutputTypeConfigurable}}. Could we express the same logic withing {{StreamOperatorFactory}} framework? For example replacing
{code:java}
if (taskOperatorFactory instanceof SimpleOperatorFactory) {
	StreamOperator operatorObject = ((SimpleOperatorFactory) taskOperatorFactory).getOperator();
	if (operatorObject instanceof OutputTypeConfigurable && outTypeInfo != null) {
		...
	}
}
{code}
with:
{code:java}
if (taskOperatorFactory.isOutputTypeConfigurable() && outTypeInfo != null) {
}
{code}
or
{code:java}
if (taskOperatorFactory instanceof OutputTypeConfigurableFactory && outTypeInfo != null) {
}
{code}
2. {{StreamOperatorFactory#createStreamOperator(...)}} could return already "setup" operator. We could for example move the {{StreamOperator#setup(...)}} method to a new interface {{SetupableStreamOperator}} and {{SetupableStreamOperator}} would be only used in {{SimpleOperatorFactory}}. However I'm not sure what would be the next clean up step - for example how much effort and how to actually get rid of {{#setup()}} method from {{AbstractStreamOperator}}?

> Introduce StreamOperatorFactory to help table perform the whole Operator CodeGen
> --------------------------------------------------------------------------------
>
>                 Key: FLINK-11974
>                 URL: https://issues.apache.org/jira/browse/FLINK-11974
>             Project: Flink
>          Issue Type: New Feature
>          Components: Runtime / Operators
>            Reporter: Jingsong Lee
>            Assignee: Jingsong Lee
>            Priority: Major
>              Labels: pull-request-available
>          Time Spent: 10m
>  Remaining Estimate: 0h
>
> If we need CodeGen an entire Operator, one possible solution is to introduce an OperatorWrapper, then generate a CodeGen sub-Operator in OperatorWrapper's open, and then proxy all methods to the sub-Operator. But introduce OperatorWrapper results in multiple virtual function calls.
> The another way is to introduce a StreamOperatorFactory. In runtime, we get the StreamOperatorFactory and create real operator to invoke. In this way, there is no redundant virtual call, the test results show that the performance improves by about 10% after the introduction of StreamOperatorFactory. (Benchmark for simple query: [https://github.com/JingsongLi/flink/blob/benchmarkop/flink-table/flink-table-planner-blink/src/test/java/org/apache/flink/table/benchmark/batch/CalcBenchmark.java])



--
This message was sent by Atlassian JIRA
(v7.6.3#76005)