[jira] [Commented] (CALCITE-5009) Transparent JDBC connection re-creation may lead to data loss

["Istvan Toth (Jira)" <ji...@apache.org>]

    [ https://issues.apache.org/jira/browse/CALCITE-5009?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=17511719#comment-17511719 ] 

Istvan Toth commented on CALCITE-5009:
--------------------------------------

I've been mulling over this.

In the strictest sense, transparent reconnecting always breaks JDBC semantics, as we have no idea of how the user changed the state of the connection, nor is it possible for us to know that.

For example, even for the simplest non-transactional select statement, the user may have set a locale on the connections via a "SET locale=x" SQL command, and the new connection would have the default locale instead, delivering unexpected results.

> Transparent JDBC connection re-creation may lead to data loss
> -------------------------------------------------------------
>
>                 Key: CALCITE-5009
>                 URL: https://issues.apache.org/jira/browse/CALCITE-5009
>             Project: Calcite
>          Issue Type: Bug
>          Components: avatica
>            Reporter: Istvan Toth
>            Assignee: Istvan Toth
>            Priority: Major
>
> Currently, if the server-side JDBC connection goes away for any reason
>  * Avatica connection cache expiry
>  * LB/HA Failover
>  * Some problem with the "real" connection
> we attempt to create a new "real" JDBC connection, and continue using that instead of the original connection
> [https://github.com/apache/calcite-avatica/blob/fbdcc62745a0e8920db759fb6bdce564d854e407/core/src/main/java/org/apache/calcite/avatica/AvaticaConnection.java#L796]
> This is fine for most read-only connections, but it can break transaction semantics, which is captured in the "real" connection object.
> {noformat}
> conn.setAutocommit(false)
> stmt = conn.createStatement()
> execute(insert A)
> //Connection lost and object recreated which now proxies a new "real" connection
> execute(insert B)
> conn.commit()
> //We have lost "insert A"{noformat}
> I'm not sure if we synchronize autocommit state of the new connection to the lost one or not, but it's bad either way.
>  
> We should either completely drop this feature, add some logic that avoids it if there is an open transaction and/or only allow it for connections that have the readOnly flag set.



--
This message was sent by Atlassian Jira
(v8.20.1#820001)