[DBCP-484] Connection leak during XATransaction in high load

• Type: Bug
• Status: Resolved
• Priority: Major
• Resolution: Fixed
• Affects Version/s: 2.2.0
• Fix Version/s: 2.4.0
• External issue URL: https://github.com/apache/commons-dbcp/pull/10
• External issue ID: 10
• Assignee: Unassigned
• Reporter: Emanuel Freitas
• Created: 25/Jan/18 15:48
• Updated: 08/Jan/19 20:49
• Resolved: 21/May/18 14:44

Description

We're experiencing a connection leak in a distributed transaction when the system is under heavy load. We're using commons-dbcp (latest version) + eclipselink and narayana to perform transaction coordination.

From time to time we can see a stacktrace reporting an abandoned connection. We are trying to figure out what's the root cause and we think that might be some issue in the commons dbcp (not sure) . More specifically, this parte of the code:

ManagedConnection#updateTransactionStatus

if (transactionContext != null) {
    if (transactionContext.isActive()) {
        if (transactionContext != transactionRegistry.getActiveTransactionContext()) {
            throw new SQLException("Connection can not be used while enlisted in another transaction");
        }
        return;
    }
    // transaction should have been cleared up by TransactionContextListener, but in
    // rare cases another lister could have registered which uses the connection before
    // our listener is called.  In that rare case, trigger the transaction complete call now
    transactionComplete();    
}

 

If the transactionContext is different than null but the state is not "active" (ex: STATUS_ROLLEDBACK, STATUS_ROLLING_BACK, etc) it executes the transactionComplete mothod that clears the reference to a shared connection and after that the connection is never closed (returned to the pool). 

 

If we move the transactionComplete(); to an else,(see below), the connection leak does not happen.

if (transactionContext != null) {
    if (transactionContext.isActive()) {
        if (transactionContext != transactionRegistry.getActiveTransactionContext()) {
            throw new SQLException("Connection can not be used while enlisted in another transaction");
        }
        return;
    }
} else {
    transactionComplete();
}

 

After this the dbcp unit tests still pass but I'm not sure about this changes. Can you please check?

Thanks

 

Issue Links

• relates to: DBCP-491 Ensure DBCP ConnectionListener can deal with transaction managers which invoke rollback in a separate thread