Blog do projektu Open Source JavaHotel

poniedziałek, 30 lipca 2012

DB2, HADR and C++, make an application failover transparent

When failover is executed the application receives "SQL30108N" exception. The remedy is quit simple, just redo the last transaction. But how to accomplish it in a programmatic way ? If our application contains tens or hundreds transaction scattered through the code it could be complicated. One would have to identify all transaction boundaries and add additional logic to detect "SQL30108N" exception after every commit and replay transaction.
Solution example
There is no universal solution to this problem but in the example provided before  it is quite easy to make this application failover transparent. It can be done by virtue of command pattern implemented there.
To implement all methods exposed in header I designed a simple class with signature like

class SqlCommand {
    hadrmodel &ha;
    virtual SQLRETURN command(hadrmodel &ha,SQLHANDLE  StmtHandle) = 0;
    bool errorHandleSTMT;
    bool directrunCommand(std::string info) {
    SQLRETURN RetCode = command(ha,StmtHandle);
public :
    void runCommand(std::string info) ...
Client should implement 'command' method and reuse some common stuff provided (for instance statement handler created before). Main function (runCommand method) is responsible for preparing environment, launching 'command' and evaluating result. We assume that every instance of SqlCommand is stateless and contains all necessary data. Because main class evaluates result it is quite easy to recognize "SQL30108N" exception (-30108 numerically) and rerun command again. So assuming that 'directrunCommand' return 'true' if replaying is suggested the main method looks like.
    SqlCommand(hadrmodel &par) : ha(par) {}
    void runCommand(std::string info) {
      if (directrunCommand(info)) {
The full source code is available here.
Because the next transaction after failover should be executed correctly there is no loop but simple logic : 'if (repeat suggested) then {run again once}' . It is unlikely to have series of failover and failback one after another so if again "SQL30108N" will be caught it is the other mistake and it is much better to raise an error at once.
Only modifying 'sqlCommand' class was necessary in this example. All 'client' code (meaning 'sqlCommand' implementations) and application code did not require any fixes.
Making application failover transparent was very easy in this simple example. I dare say that even the application were more complicated (meaning more sql related commands) it would be simple also and would involve only fixes in one place.
But it is not an universal solution, it was possible only because a proper design pattern was implemented from the very beginning.

Brak komentarzy:

Publikowanie komentarza