[SPARK-30260][SQL] Spark-Shell throw ClassNotFoundException exception for more than one statement to use UDF jar

[southernriver]

What changes were proposed in this pull request?

When we start spark-shell and use the udf for the first statement ,it's ok. But for the other statements it failed to load jar to current classpath and would throw ClassNotFoundException.It seems like that the first ClassLoader is different from the other's. For Spark-shell, the maintained class loader is always IMainsTranslatingClassLoader ,and for addJar Operation, the current classLoader is NonClosableMutuableclassLoader. For the first statement, there jar was loaded to right classLoader,and for other statements, the jar has been registered to functionRegistry and would not reload to NonClosableMutuableclassLoader, we need to reset classloader to active sparkSession's.

Here, I will show difference between the first statement and the second.
First statement is NonClosableMutuableclassLoader:

Second statement is IMainsTranslatingClassLoader:

Why are the changes needed?

The problem can be reproduced as described in the below.

scala> val res = spark.sql("select bigdata_test.Add(1,2)").show()
----------------------
|bigdata_test.Add(1, 2)|
----------------------
| 3|
----------------------
scala> val res = spark.sql("select bigdata_test.Add(1,2)").show()
org.apache.spark.sql.AnalysisException: No handler for UDF/UDAF/UDTF 'scala.didi.udf.Add': java.lang.ClassNotFoundException: scala.didi.udf.Add; line 1 pos 8
at scala.reflect.internal.util.AbstractFileClassLoader.findClass(AbstractFileClassLoader.scala:62)
at java.lang.ClassLoader.loadClass(ClassLoader.java:424)
at java.lang.ClassLoader.loadClass(ClassLoader.java:357)
at org.apache.spark.sql.hive.HiveShim$HiveFunctionWrapper.createFunction(HiveShim.scala:251)
at org.apache.spark.sql.hive.HiveSimpleUDF.function$lzycompute(hiveUDFs.scala:56)
at org.apache.spark.sql.hive.HiveSimpleUDF.function(hiveUDFs.scala:56)
at org.apache.spark.sql.hive.HiveSimpleUDF.method$lzycompute(hiveUDFs.scala:60)
at org.apache.spark.sql.hive.HiveSimpleUDF.method(hiveUDFs.scala:59)
at org.apache.spark.sql.hive.HiveSimpleUDF.dataType$lzycompute(hiveUDFs.scala:77)
at org.apache.spark.sql.hive.HiveSimpleUDF.dataType(hiveUDFs.scala:77)
at org.apache.spark.sql.hive.HiveSessionCatalog$$anonfun$makeFunctionExpression$3.apply(HiveSessionCatalog.scala:79)
at org.apache.spark.sql.hive.HiveSessionCatalog$$anonfun$makeFunctionExpression$3.apply(HiveSessionCatalog.scala:71)
at scala.util.Try.getOrElse(Try.scala:79)
at org.apache.spark.sql.hive.HiveSessionCatalog.makeFunctionExpression(HiveSessionCatalog.scala:71)
at org.apache.spark.sql.catalyst.catalog.SessionCatalog$$anonfun$org$apache$spark$sql$catalyst$catalog$SessionCatalog$$makeFunctionBuilder$1.apply(SessionCatalog.scala:1133)

After fix:

scala> val res = spark.sql("select bigdata_test.Add(1,2)").show()
+----------------------+ |bigdata_test.Add(1, 2)|
+----------------------+
| 3|
+----------------------+
scala> val res = spark.sql("select bigdata_test.Add(1,2)").show()
+----------------------+
|bigdata_test.Add(1, 2)|
+----------------------+
| 3|
+----------------------+

we should resolve this bug!

Does this PR introduce any user-facing change?

No

How was this patch tested?

manual.

[southernriver]

cc @cloud-fan@maropu@dongjoon-hyun

[southernriver]

@AmplabJenkins

[dongjoon-hyun]

ok to test

[maropu]

Is this related to #23921? It seems the #23921 has a different approach from this. cc: @HyukjinKwon

[SparkQA]

Test build #115349 has finished for PR 26888 at commit 7b24673.

• This patch passes all tests.
• This patch merges cleanly.
• This patch adds no public classes.

[HyukjinKwon]

Yeah, seems a duplicate.

[HyukjinKwon]

@southernriver can you explain why/when the class loaders are changed?

[cloud-fan]

should IMainTranslatingClassLoader fallback to spark context class loader?

[HeartSaVioR]

That's duplicated with #23921 and IMHO #23921 is clearer fix. As it is quite old and no test, I've just took over and raised #27025.

[dongjoon-hyun]

According to the above discussion, I'll close this PR and SPARK-30260 . Thank you, @southernriver and all!