I've just run into a problem with a Quartz job that I'm invoking through Spring. My ExecuteInternal method had a [Transaction] attribute (because it does a load of DB calls) but when it runs, I get the 'No NHibernate session bound to thread' error.
Just wondering if that's because Spring.NET doesn't support the [Transaction] attribute in Quartz objects?
If not, that's fine... I can start a transaction manually, but wanted to check that it was the case, and not a silly error in my config somewhere.
[Update]
I figured it out actually. In the API docs, it says the preferable way to do this is use transactions on the service layer. My job was using DAOs to do its work, but my transactions are on my service layer, so I just called service methods from my job instead to do the same work (saving, updating records etc) since they already existed.
It also suggests that if you give the SchedulerFactoryObject a DbProvider, you can use transactions in the job itself, but when I did that it seemed to want to find my triggers configured in a special table in the DB (which I haven't set up since my triggers are all in XML) but that's possibly another way to do it.
Calling service methods works fine for me though.
The transaction attribute works using aop. Spring.NET creates an aop proxy for the decorated object. This proxy creates the session and starts the transaction.
In the ExecuteInternal method, you don't call the method on a proxy, but on the target itself. Therefore spring cannot intercept the call and do its transaction magic.
Your services are injected and therefore the transaction attribute works for them.
There's a good explanation in the spring docs on this subject: http://www.springframework.net/doc-latest/reference/html/transaction.html#tx-understandingimpl
Related
I have a simple question. I'm newer with UnityContainer of Miscrosoft. I'm writing an ASP.NET MVC application with Unity for DI.
Have I a different CONTAINER for each user connected to my web app? Or the CONTAINER is the same for all users?
So if I resolve the life time of an object with ContainerControlledLifetimeManager does it mean that only for one user session this object is always the same?
I hope you understand.
Thanks,
Christian
Lifetime refers to the life of the object created by the DI process. Per request means each request gets its own object. If the object depends on the current user, querystring values on that request or values/presence of Request headers a PerRequest lifetime is appropriate. If you have settings that vary based on location of your service, for example, you saved values from web.config, then a the container is most likely created in global.asa and these objects can live as long as the container lives.
A concrete example:
You have a service as part of your site and you are migrating to vNext of that service. Users can opt-in by clicking a link that includes a parameter like &myService=vNext to see the new behavior. your Factory method uses the value of this parameter to select vNow or vNext for each request.
Here's some pseudo code to get you started:
container.RegisterInstance<IProductFactory>("enterprise", new EnterpriseProductFactory());
container.RegisterInstance<IProductFactory>("retail", new RetailProductFactory());
container.RegisterVersionedServiceFactory<IProductFactorySettings, IProductFactory>();
In this example RegisterVersionedServiceFactory is an extension method that does nothing but decide which of the IProductFactory instances to use for the current request. The factory provides the current instance (there are only two for the life of the service) to use for this request (thousands per seconds).
This pattern is what makes a very large site you probably used recently both very stable and very flexible. New versions of services are rolled out using this exact same pattern to help keep the site very stable.
I'm reading the Grails docs on services which make numerous mention of transactions/transactionality, but without really defining what a transactional service method really is.
Given the nature of services, they frequently require transactional behaviour.
What exactly does this mean? Are transactional methods only those that use JPA/JDBC to communicate with a relational DB, or do they apply to anything covered by JTA?
Is there any reason why I just wouldn't make a service class #Transactional in case it evolves to some day use a transaction? In other words, are there performance concerns to making all service methods transactional?
Grails services are transactional by default - if you don't want a service to be transactional, you need to remove all #Transactional annotations (both Grails' #grails.transaction.Transactional and Spring's #org.springframework.transaction.annotation.Transactional) and add
static transactional = false
If you haven't disabled transactions with the transactional property and have no annotations, the service works the same as if it were annotated with Spring's annotation. That is, at runtime Spring creates a CGLIB proxy of your class and registers an instance of the proxy as the Spring bean, and it delegates to an instance of your actual class to do the database access and your business logic. This lets the proxy intercept all public method calls and start a new transaction, join an existing one, create a new one, etc.
The newer Grails annotation has all of the same settings as the Spring annotation, but it works a bit differently. Instead of triggering the creation of a single proxy, each method is rewritten by an AST transform during compilation, essentially creating a mini proxy for each method (this is obviously a simplification). This is better because the database access and transaction semantics are the same, but if you call one annotated method from another annotated with different settings, the different settings will be respected. But with a proxy, it's a direct call inside the delegate instance, and the proxy is bypassed. Since the proxy has all of the logic to create a new transaction or use other different settings, the two methods will use the first method's settings. With the Grails annotation every method works as expected.
There is a small performance hit involved for transactional methods, and this can accumulate if there are a lot of calls and/or a lot of traffic. Before your code runs, a transaction is started (assuming one isn't active) and to do this, a connection must be retrieved from the pool (DataSource) and configured to turn off autocommit, and make the various transaction settings (isolation, timeout, readonly, etc.) have to be made. But the Grails DataSource is actually a smart wrapper around the "real" one. It doesn't get a real JDBC Connection until you start a query, so all of the configuration settings are cached until then, and then "replayed" on the real connection. If the method doesn't do any database work (either because it never does, or because it exits early based on some condition before the db access code fires), then there's basically no database cost. But if it does, then things work as expected.
Don't rely on this DataSource proxying logic though - it's best to be explicit about which services are transactional and which aren't, and within each service which methods are transactional and which aren't. The best way to do this is by annotating methods as needed, or adding a single annotation at the class level if all methods use the same settings.
You can get more info in this talk I did about transactions in Grails.
First, if your performance concerns are due to the fact your services are transactional then you have reached nirvana. I say this because there are going to be plenty of other bottle necks in your application long before this is a major (or even minor) concern. So, don't fret about that.
Typically in Grails a transaction relates to the transactional state of a database connection or hibernate session. Though it could be anything managed by the JTA with the proper Spring configuration.
In simple terms, it's usually means (by default) a database transaction.
I building multitenancy into the unit of work for a set of services. I want to keep the tenancy question out of the way of day-to-day business domain work, and I do not want to touch every existing consumer in the system (I am retrofitting the multitenancy onto a system without any prior concept of a tenant).
Most messages in the system will be contexted by a tenant. However, there will be some infrastructure messages which will not be, particularly for the purpose of automating tenant creation. I need a way of determining whether to use a tenant-contexted unit of work, or a infrastructure unit of work uncontexted by a tenant because the way I interact with the database is different depending on whether I have tenant context. The unit of work is built in the process of spinning up the dependencies of the consumer.
As such I need a way of peeking at the message or its metadata before consuming it, and specifically, I need to be able to peek at it during the dependency building. I was intended to have a tag interface to mark tenant management messages out from normal business domain messages, but any form of identifying the difference could work. If I am in a unit of work resulting from an HTTP request, I can look at WebApi's HttpContext.Current and see the headers of the current request, etc. How do I do something analogous to this if I am in a unit of work resulting from messaging?
I see there is a way to intercept messages with BeforeConsumingMessage() but I need a way of correlating it to the current unit of work I am spinning up and I'm not seeing how that would work for me. Pseudocode for what I am trying to do:
if MessageContext.Message.GetType() = typeof<ITenantInfrastructureMessage>:
database = new Database(...)
else:
tenantId = MessageContext.Headers.TenantId;
database = new TenantDatabase(..., tenantId)
I am working in C#/.NET using MassTransit with RabbitMQ and Autofac with MassTransit's built-in support for both.
Your best option is to override at the IConsumerFactory<T> extension point, and extract the tenant from the message (either via a message header, or some message property) and register that in the container child lifetime scope so that subsequent resolutions from the actual consumer class (and it's dependencies) are properly matched to the tenant in the message.
In our systems, we have a TenantContext that is registered in a newly created LifetimeScope (we're using Autofac), after which we resolve the consume from the child scope, and the dependencies that use the tenant context get the proper value since it's registered as part of building the child container for the message scope.
It works extremely well, we even built up extension methods to make it easy for developers registering consumers to specify "tenant context providers" that go from a message type to the proper tenant id, which is used to build the TenantContext.
You can do similar things with activity factories in Courier routing slips (which are a specialization of a consumer).
I have an ASP.NET MVC application that uses NHibernate to persist data into a SQL Server database.
There are cases where I want to save an entry into a database (initially triggered by a call into an action method on a controller) but there's no need to block the caller.
Is it "safe" to try to implement a fire-and-forget mechanism into the database that will put the database call into a Task and then invoke it on the background so control can return immediately to the caller? (OR accomplish the same thing with BackgroundWorker or the "async/await" keywords) I need a solution where NHibernate will not get tripped up by ASP.NET trying to clean up its ISession, which is per-request. I'm using Autofac for lifetime management on the session. I assume that the database operation would have a slightly longer lifetime than the web request itself, and I'm not sure how smoothly that would work.
It is not safe to do this; I have a blog post on the subject. The problem is that when you have no requests in progress, it is possible that your entire AppDomain can be torn down. Also, consider what would happen if the database insert failed for some reason? If you return early, then there's no way to notify the client of an error.
A reliable solution must store the data in some kind of persistent place before returning success to the caller. This can be directly in the database, or in a queue of some kind (to be later processed by an independent worker).
I have JMS plugin installed in Grails 2.2.0. It works famously. However, I now want to extend the functionality to all my domain objects to persist them across a JMS ActiveMQ connection.
The database is on the remote system which has an ActiveMQ broker on it. I can talk to it fine. But, I'm unclear on how to implement the code to save(), search(), update(), etc from my controller using JMS rather than a local DB. I think I have save() figured out but not for sure. Then how do I hook into the dynamic finders and such to supply the needed service so the actual search criteria is sent over a queue and the response is sent back on another queue and populates the correct Grails results for display? My head hurts from reading so much and getting no where fast.
Thanks,
Marshall