[
https://issues.apache.org/jira/browse/HIVE-28834?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17937130#comment-17937130
]
Stamatis Zampetakis commented on HIVE-28834:
--------------------------------------------
The best solution would be to rewrite the code to use the try-with-resources
mechanism that can automatically deal with this kind of problems. Below an
extract from the [JVM
specification|https://docs.oracle.com/javase/specs/jls/se8/html/jls-14.html#jls-14.20.3.1]:
In a basic {{{}try{}}}-with-resources statement that manages a single resource:
* If the initialization of the resource completes abruptly because of a
{{throw}} of a value {{{}V{}}}, then the {{{}try{}}}-with-resources statement
completes abruptly because of a {{throw}} of the value {{{}V{}}}.
* If the initialization of the resource completes normally, and the {{try}}
block completes abruptly because of a {{throw}} of a value {{{}V{}}}, then:
** If the automatic closing of the resource completes normally, then the
{{{}try{}}}-with-resources statement completes abruptly because of a {{throw}}
of the value {{{}V{}}}.
** If the automatic closing of the resource completes abruptly because of a
{{throw}} of a value {{{}V2{}}}, then the {{{}try{}}}-with-resources statement
completes abruptly because of a {{throw}} of value {{V}} with {{V2}} added to
the suppressed exception list of {{{}V{}}}.
* If the initialization of the resource completes normally, and the {{try}}
block completes normally, and the automatic closing of the resource completes
abruptly because of a {{throw}} of a value {{{}V{}}}, then the
{{{}try{}}}-with-resources statement completes abruptly because of a {{throw}}
of the value {{{}V{}}}.
An alternative, and possibly simpler solution, would be to catch exceptions
during rollback, log them, and stop their propagation. This would be somewhat
problematic if an exception only occurs during rollback.
> HMS swallows original failure when the rollback of a transaction fails
> ----------------------------------------------------------------------
>
> Key: HIVE-28834
> URL: https://issues.apache.org/jira/browse/HIVE-28834
> Project: Hive
> Issue Type: Bug
> Components: Metastore
> Reporter: Stamatis Zampetakis
> Assignee: Stamatis Zampetakis
> Priority: Major
>
> The problem mainly affects {{HMSHandler}} and {{ObjectStore}} but can be
> generalized to any client that is executing rollback in a {{finally}} block.
> Some very frequent problematic patterns are outlined below.
> +Pattern in ObjectStore+
> {code:java}
> boolean committed = false;
> try {
> openTransaction();
> // some operations
> committed = commitTransaction();
> } finally {
> rollbackAndCleanup(committed, null);
> }
> {code}
>
> +Pattern in HMSHandler+
> {code:java}
> boolean success = false;
> RawStore ms = getMS();
> try {
> ms.openTransaction();
> // some operations
> success = ms.commitTransaction();
> } finally {
> if (!success) {
> ms.rollbackTransaction();
> }
> }
> {code}
> The problem is that whatever failure/exception is raised inside the {{try}}
> block between the open and commit transaction will be completely lost if the
> rollback fails. The exception from the rollback masks the original error.
> Most of the time the initial failure is far more important than the failure
> inside the rollback so losing it is major issue when trying to diagnose a
> problem in HMS.
--
This message was sent by Atlassian Jira
(v8.20.10#820010)
