[
https://issues.apache.org/jira/browse/SPARK-23986?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16610330#comment-16610330
]
Dmitry Zanozin commented on SPARK-23986:
----------------------------------------
Well, it's a bit hard to provide the exact code we use when this failure
occurs, many aggregated attributes with several custom UDFs. The same code
works fine with a different dataframe which differs from the first one by
having 1 string attribute instead of 4 int ones (among other 22 attributes).
Here is a DF schema which fails:
{noformat}
scala> df.printSchema()
root
|-- attr1: timestamp (nullable = true)
|-- attr2: string (nullable = true)
|-- attr3: integer (nullable = true)
|-- attr4: integer (nullable = true)
|-- attr5: string (nullable = true)
|-- attr6: integer (nullable = true)
|-- attr7: byte (nullable = true)
|-- attr8: integer (nullable = true)
|-- attr9: timestamp (nullable = true)
|-- attr10: string (nullable = true)
|-- attr11: byte (nullable = true)
|-- attr12: short (nullable = true)
|-- attr13: short (nullable = true)
|-- attr14: integer (nullable = true)
|-- attr15: integer (nullable = true)
|-- attr16: short (nullable = true)
|-- attr17: integer (nullable = true)
|-- attr18: string (nullable = true)
|-- attr19: byte (nullable = true)
|-- attr20: byte (nullable = true)
|-- attr21: integer (nullable = true)
|-- attr22: date (nullable = true)
{noformat}
if we replace attr12-attr15 (which represent the unique record key and are part
of a group by statement) with a single string attr (a primary key of another
object type) the code works fine.
Information about the execution environment:
{noformat}
spark-2.3.1-bin-without-hadoop
hadoop-2.8.4
Scala 2.11.8
Java HotSpot(TM) 64-Bit Server VM, Java 1.8.0_162
{noformat}
> CompileException when using too many avg aggregation after joining
> ------------------------------------------------------------------
>
> Key: SPARK-23986
> URL: https://issues.apache.org/jira/browse/SPARK-23986
> Project: Spark
> Issue Type: Bug
> Components: SQL
> Affects Versions: 2.3.0
> Reporter: Michel Davit
> Assignee: Marco Gaido
> Priority: Major
> Fix For: 2.3.1, 2.4.0
>
> Attachments: spark-generated.java
>
>
> Considering the following code:
> {code:java}
> val df1: DataFrame = sparkSession.sparkContext
> .makeRDD(Seq((0, 1, 2, 3, 4, 5, 6)))
> .toDF("key", "col1", "col2", "col3", "col4", "col5", "col6")
> val df2: DataFrame = sparkSession.sparkContext
> .makeRDD(Seq((0, "val1", "val2")))
> .toDF("key", "dummy1", "dummy2")
> val agg = df1
> .join(df2, df1("key") === df2("key"), "leftouter")
> .groupBy(df1("key"))
> .agg(
> avg("col2").as("avg2"),
> avg("col3").as("avg3"),
> avg("col4").as("avg4"),
> avg("col1").as("avg1"),
> avg("col5").as("avg5"),
> avg("col6").as("avg6")
> )
> val head = agg.take(1)
> {code}
> This logs the following exception:
> {code:java}
> ERROR CodeGenerator: failed to compile:
> org.codehaus.commons.compiler.CompileException: File 'generated.java', Line
> 467, Column 28: Redefinition of parameter "agg_expr_11"
> {code}
> I am not a spark expert but after investigation, I realized that the
> generated {{doConsume}} method is responsible of the exception.
> Indeed, {{avg}} calls several times
> {{org.apache.spark.sql.execution.CodegenSupport.constructDoConsumeFunction}}.
> The 1st time with the 'avg' Expr and a second time for the base aggregation
> Expr (count and sum).
> The problem comes from the generation of parameters in CodeGenerator:
> {code:java}
> /**
> * Returns a term name that is unique within this instance of a
> `CodegenContext`.
> */
> def freshName(name: String): String = synchronized {
> val fullName = if (freshNamePrefix == "") {
> name
> } else {
> s"${freshNamePrefix}_$name"
> }
> if (freshNameIds.contains(fullName)) {
> val id = freshNameIds(fullName)
> freshNameIds(fullName) = id + 1
> s"$fullName$id"
> } else {
> freshNameIds += fullName -> 1
> fullName
> }
> }
> {code}
> The {{freshNameIds}} already contains {{agg_expr_[1..6]}} from the 1st call.
> The second call is made with {{agg_expr_[1..12]}} and generates the
> following names:
> {{agg_expr_[11|21|31|41|51|61|11|12]}}. We then have a parameter name
> conflicts in the generated code: {{agg_expr_11.}}
> Appending the 'id' in s"$fullName$id" to generate unique term name is source
> of conflict. Maybe simply using undersoce can solve this issue :
> $fullName_$id"
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
---------------------------------------------------------------------
To unsubscribe, e-mail: issues-unsubscr...@spark.apache.org
For additional commands, e-mail: issues-h...@spark.apache.org
