[ https://forge.continuent.org/jira/browse/SEQUOIA-955?page=comments#action_14433 ]
Emmanuel Cecchet commented on SEQUOIA-955: ------------------------------------------ There is a bug in RAIDb-1 load balancer in the way commit,rollback to savepoint and release savepoint tasks are posted. We check if the transaction was started on the backend to post the task to its queue. However, a backend can be late and could still have the transaction start in its pending queue. In that case the commit (or rollback/release savepoint) will never be posted to that backend leaving it with uncompleted transactions. A fix for that problem is committed in branch 2.10. > Raidb1 load balancer and WaitForCompletionPolicy=first does not work > -------------------------------------------------------------------- > > Key: SEQUOIA-955 > URL: https://forge.continuent.org/jira/browse/SEQUOIA-955 > Project: Sequoia > Type: Bug > Components: Core > Versions: Sequoia 2.10.9, Sequoia 2.10.8 > Environment: Windows XP, single controller, RAIDb-1, > WaitForCompletionPolicy=first > Reporter: xavier roques > Assignee: Emmanuel Cecchet > Priority: Critical > Fix For: sequoia 2.10.10 > > > If have a VirtualDatabase with two backends db1 and db2, the load balancer is > the Raiddb1 and the waitForCompletionPolicy is First. > DB1 and DB2 are up. My application makes a lot of updates and after some time > one of the Database is disabled. > I don't undertstand why, because the database is up. > After some searches I find out that database was disabled because of a > timeout while trying to get a new connection. > My pool of connection 50 was full ? But I released all my connections ... > So now, I will explain why it does not work. > There is a load balancer with two backends: Db1 and DB2. Each backend has its > own queue and the queue of each backend is managed by different a thread. > A lot of updates are sent to the virtualdatabase. > The load balancer dispatches the requests on each backend. Now let's us > suppose that the two following requests arrive. > Request 1 (update) > Request 2 (commit) > The load balancer sends the request 1 on each backend and wait for the first > completion. > If the thread managing the DB1 executes the request 1 and send the result to > the load balancer. It is possible that the load balancer dispatch the request > 2 even if the Request 1 has not been started on the DB2. > The code of the loadBalancer RAIDDB1.commit is: > {code} > for (int i = 0; i < nbOfThreads; i++) > { > DatabaseBackend backend = (DatabaseBackend) enabledBackends.get(i); > if (backend.isStartedTransaction(lTid)) > commitList.add(backend); > } > {code} > Unfortunately when this code is executed the backend DB2 may not have started > to execute the request 1 so the commit will only be pushed on the backend 1. > So when the request 1 will be executed on DB2, a new connection will be open > and it will never be released because the commit will never be executed on > the backend DB2. -- This message is automatically generated by JIRA. - If you think it was sent incorrectly contact one of the administrators: https://forge.continuent.org/jira/secure/Administrators.jspa - For more information on JIRA, see: http://www.atlassian.com/software/jira _______________________________________________ Sequoia mailing list [email protected] https://forge.continuent.org/mailman/listinfo/sequoia
