[
https://issues.apache.org/jira/browse/FLINK-3850?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15902057#comment-15902057
]
ASF GitHub Bot commented on FLINK-3850:
---------------------------------------
Github user fhueske commented on a diff in the pull request:
https://github.com/apache/flink/pull/3040#discussion_r105023568
--- Diff:
flink-libraries/flink-table/src/main/scala/org/apache/flink/table/plan/nodes/dataset/DataSetCalc.scala
---
@@ -124,8 +127,26 @@ class DataSetCalc(
body,
returnType)
+ def getForwardIndices = {
+ // get (input, output) indices of operands,
+ // filter modified operands and specify forwarding
+ val inputFields = extractRefInputFields(calcProgram)
+ calcProgram.getProjectList
+ .map(_.getIndex)
+ .zipWithIndex
+ .filter(tup => inputFields.contains(tup._1))
--- End diff --
This position based mapping is making some assumptions about the internal
organization of a `RexProgram`, i.e., that the first `n` fields of the
expressions list are filled in order by the `n` fields of the input.
Can we change this to iterate over the projection list and looking up the
expression and check whether it is a `RexInputRef`. Basically something like:
```
calcProgram.getProjectList.zipWithIndex.map { case (p, out) =>
val expr = calcProgram.getExprList.get(p.getIndex)
expr match {
case i: RexInputRef => Some((i.getIndex, out))
case _ => None
}
}.flatten
```
> Add forward field annotations to DataSet operators generated by the Table API
> -----------------------------------------------------------------------------
>
> Key: FLINK-3850
> URL: https://issues.apache.org/jira/browse/FLINK-3850
> Project: Flink
> Issue Type: Improvement
> Components: Table API & SQL
> Reporter: Fabian Hueske
> Assignee: Nikolay Vasilishin
>
> The DataSet API features semantic annotations [1] to hint the optimizer which
> input fields an operator copies. This information is valuable for the
> optimizer because it can infer that certain physical properties such as
> partitioning or sorting are not destroyed by user functions and thus generate
> more efficient execution plans.
> The Table API is built on top of the DataSet API and generates DataSet
> programs and code for user-defined functions. Hence, it knows exactly which
> fields are modified and which not. We should use this information to
> automatically generate forward field annotations and attach them to the
> operators. This can help to significantly improve the performance of certain
> jobs.
> [1]
> https://ci.apache.org/projects/flink/flink-docs-release-1.0/apis/batch/index.html#semantic-annotations
--
This message was sent by Atlassian JIRA
(v6.3.15#6346)
