[ 
https://issues.apache.org/jira/browse/FLINK-4469?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=15711937#comment-15711937
 ] 

ASF GitHub Bot commented on FLINK-4469:
---------------------------------------

Github user twalthr commented on a diff in the pull request:

    https://github.com/apache/flink/pull/2653#discussion_r90442225
  
    --- Diff: 
flink-libraries/flink-table/src/main/scala/org/apache/flink/api/table/expressions/ExpressionParser.scala
 ---
    @@ -451,6 +452,28 @@ object ExpressionParser extends JavaTokenParsers with 
PackratParsers {
         }
       }
     
    +  lazy val tableFunctionCall: PackratParser[LogicalNode] =
    +    functionIdent ~ "(" ~ repsep(expression, ",") ~ ")" ^^ {
    +    case name ~ _ ~ args ~ _ => 
UnresolvedTableFunctionCall(name.toUpperCase, args)
    --- End diff --
    
    I think it would not hurt if we support `split(c) as (f1, f2)` for all 
expressions. This could also help in cases where we expand the projection like 
`pojo.flatten() as (f1, f2)`. I don't like the idea of a logical parser. It is 
already confusing enough for the user that we have Java API expression strings 
and SQL strings, we don't need another string syntax.
    @fhueske what do you think?


> Add support for user defined table function in Table API & SQL
> --------------------------------------------------------------
>
>                 Key: FLINK-4469
>                 URL: https://issues.apache.org/jira/browse/FLINK-4469
>             Project: Flink
>          Issue Type: New Feature
>          Components: Table API & SQL
>            Reporter: Jark Wu
>            Assignee: Jark Wu
>
> Normal user-defined functions, such as concat(), take in a single input row 
> and output a single output row. In contrast, table-generating functions 
> transform a single input row to multiple output rows. It is very useful in 
> some cases, such as look up in HBase by rowkey and return one or more rows.
> Adding a user defined table function should:
> 1. inherit from UDTF class with specific generic type T
> 2. define one or more evel function. 
> NOTE: 
> 1. the eval method must be public and non-static.
> 2. the generic type T is the row type returned by table function. Because of 
> Java type erasure, we can’t extract T from the Iterable.
> 3. use {{collect(T)}} to emit table row
> 4. eval method can be overload. Blink will choose the best match eval method 
> to call according to parameter types and number.
> {code}
> public class Word {
>   public String word;
>   public Integer length;
> }
> public class SplitStringUDTF extends UDTF<Word> {
>     public Iterable<Word> eval(String str) {
>         if (str != null) {
>             for (String s : str.split(",")) {
>                 collect(new Word(s, s.length()));
>             }
>         }
>     }
> }
> // in SQL
> tableEnv.registerFunction("split", new SplitStringUDTF())
> tableEnv.sql("SELECT a, b, t.* FROM MyTable, LATERAL TABLE(split(c)) AS 
> t(w,l)")
> // in Java Table API
> tableEnv.registerFunction("split", new SplitStringUDTF())
> // rename split table columns to “w” and “l”
> table.crossApply("split(c) as (w, l)")        
>      .select("a, b, w, l")
> // without renaming, we will use the origin field names in the POJO/case/...
> table.crossApply("split(c)")
>      .select("a, b, word, length")
> // in Scala Table API
> val split = new SplitStringUDTF()
> table.crossApply(split('c) as ('w, 'l))
>      .select('a, 'b, 'w, 'l)
> // outerApply for outer join to a UDTF
> table.outerApply(split('c))
>      .select('a, 'b, 'word, 'length)
> {code}
> See [1] for more information about UDTF design.
> [1] 
> https://docs.google.com/document/d/15iVc1781dxYWm3loVQlESYvMAxEzbbuVFPZWBYuY1Ek/edit#



--
This message was sent by Atlassian JIRA
(v6.3.4#6332)

Reply via email to