[SPARK-10121] [SQL] Thrift server always use the latest class loader provided by the conf of executionHive's state

[yhuai]

https://issues.apache.org/jira/browse/SPARK-10121

Looks like the problem is that if we add a jar through another thread, the thread handling the JDBC session will not get the latest classloader.

[yhuai]

@liancheng I manually tested it and this looks like a right fix. How to write a unit test? Basically, I need to start the thriftserver and add a jar from another thread (e.g. spark shell) and test it in a jdbc connection.

[SparkQA]

Test build #41392 has finished for PR 8368 at commit 2e9279a.

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

[SparkQA]

Test build #41394 has finished for PR 8368 at commit 690d9cd.

• This patch passes all tests.
• This patch merges cleanly.
• This patch adds the following public classes (experimental):
  • public static class AppExecId
  • public static class StoreVersion

[liancheng]

This can be tested using HiveThriftJdbcTest.withMultipleConnectionJdbcStatement(). This method accepts zero or more Statement => Unit functions, and runs them within different JDBC sessions. Although these functions are executed sequentially, all the connections are established at the beginning, so different function are guaranteed to be executed in different thread.

[SparkQA]

Test build #41437 has finished for PR 8368 at commit aac89f4.

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

[yhuai]

@liancheng is this test good?

[liancheng]

Yeah, this LGTM.

[liancheng]

LGTM. Merging to master and branch-1.5.