如果数据库通过Castle Windsor和nhibernate工具脱机，如何重试Windows服务启动？

Question

Problem: If the DB is offline when this service is started, this service will not start as it fails inside this line: var container = new BootStrapper().Container; on start.

private static void Main(string[] args)
{
    Logger.Info("Engine Service is bootstrapping...");
    AppDomain.CurrentDomain.UnhandledException += UncaughtExceptions.DomainException;
    Directory.SetCurrentDirectory(AppDomain.CurrentDomain.BaseDirectory);

    var container = new BootStrapper().Container;
    var controller = container.Resolve<EngineController>();
    ServiceBase.Run(controller.MainView as ServiceBase);

    container.Dispose();
}

The reason it fails there is that it runs this code where it adds the nhibernate facility container.AddFacility<NHibernateFacility>(); and fails with a connection timeout.

public void Install(IWindsorContainer container, IConfigurationStore store)
{
    var isAutoTxFacilityRegistered = container.Kernel.GetFacilities().Any(f => f is AutoTxFacility);
    if (!isAutoTxFacilityRegistered) container.AddFacility<AutoTxFacility>();

    container.Register(
        Component.For<INHibernateInstaller>().ImplementedBy<CieFluentInstaller>().IsDefault().LifestyleTransient(),
        Classes.FromThisAssembly().Pick().WithService.DefaultInterfaces().LifestyleTransient()
        );

    var isNHibernateFacilityRegistered = container.Kernel.GetFacilities().Any(f => f is NHibernateFacility);
    if (!isNHibernateFacilityRegistered) container.AddFacility<NHibernateFacility>();
}

If the windows service start up takes longer than 30 seconds (which it may if updates or backups are being conducted on the DB) the app service fails to start.

I'm using FluentNhibernate, NHibernate, Castle Windsor with NHibernateFacility.

Things I've tried:

• Can't do it from the service start event because it fails before it gets to the view or controller. The view and controller have no direct access to the IoC container, only via an injected IoCFactory as per Castle Windsor recommendations.

• I've tried to spawn a thread in the main and start it off there with a retry loop but because the service "waits" inside the ServiceBase.Run method, I can't seem to get the correct returns to make it "fake start" while in a retry loop.

• Investigated lengthening the service start timeout, but can't access the servicebase/view since it fails before then and a system wide change at hundreds of production sites is not an option.

Question: How can I make it so that the windows service "starts" when DB is offline given the design?

Answer 1

You need to divide your startup actions into two categories:

1. Actions that must happen fairly immediately and/or won't fix themselves in case of failure. Things such as a mandatory configuration file missing, for which administrator intervention would be required.

2. Actions that we're OK to delay, or - more importantly - actions that can fail due to transient errors. Such errors can be network failure or that we happened to start somewhat faster than the database server after a reboot.

You service OnStart code should follow this basic structure:

OnStart:
    Perform the immediate category 1 tasks and exit if any of these fail.
    Launch the main application thread.

One approach to the "main application thread" is to follow this basic structure:

ManualResetEvent shutdownRequestedEvent = new ManualResetEvent()

RealMain:
    while (!shutdownRequestedEvent.WaitOne(0) && !bootstrapPerformed)
    {
        try
        {
            PerformBootstrap()
            bootstrapPerformed = true
        }
        catch (Exception ex)
        {
            LogError(ex)
        }

        if (!bootstrapPerformed)
            shutdownRequestedEvent.WaitOne(some timeout)
    }

    Second bootstrap action similar to above, etc.

    Third bootstrap action similar to above, etc.

    Eventually, start performing real work, while listening to 
    the shutdownRequestedEvent.

The services OnShutdown would signal the shutdownRequestedEvent and then wait for the RealMain thread to exit.

If the RealMain thread serves no purpose other then setup, it should perhaps be allowed to exit when it's done with all bootstrap tasks.

Another thing to be careful about is to make sure your service, during normal operation, can withstand the temporary loss of access to a resource due to transient errors. For example, your service shouldn't crash just because someone reboots the database server. It should just wait patiently and retry forever.

An alternative approach that can work in some cases is to handle the bootstrapping as a dependency of whatever the real task is. For instance, launch the real task, the real task will request a database session, to get that we must have the session factory, if we don't yet have the session factory, launch the session factory initialization. If the session factory cannot be created, exception bubbles up and the whole task fails. The remaining work is now to wait a little while and then retry the task. Repeat forever.

Answer 2

Turned out to be a bug in NHibernate that prevented doing any of the above. Between Nibernate 2.0 and 3.0 you have to add the following to the NHibernate v3.0+ config (or in this case the FluentNHibernate):

cfg.SetProperty("hbm2ddl.keywords", "none");

This allows NHibernate to properly bootstrap itself and get to the controller now without error.