I would like to propagate JTA state (= the transaction) between a transactional REST endpoint that emits a message to a reactive-messaging connector.
#Inject
#Channel("test")
Emitter<String> emitter;
#POST
#Transactional
public Response test() {
emitter.send("test");
}
and
#ApplicationScoped
#Connector("test")
public class TestConnector implements OutgoingConnectorFactory {
#Inject
TransactionManager tm;
#Override
public SubscriberBuilder<? extends Message<?>, Void> getSubscriberBuilder(Config config) {
return ReactiveStreams.<Message<?>>builder()
.flatMapCompletionStage(message -> {
tm.getTransaction(); // = null
return message.ack();
})
.ignore();
}
}
As I understand, context-propagation is responsible for making the transaction available (see io.smallrye.context.jta.context.propagation.JtaContextProvider#currentContext). The problem seems to be, that currentContext gets created on subscription, which happens when the injection point (Emitter<String> emitter) get its instance. Which is too early to properly capture the transaction.
What am I missing?
By the way, I am having the same problem when using #Incoming / #Outgoing instead of the emitter. I have decided to give you this example because it is easy to understand and reproduce.
At the moment, you need to pass the current Transaction in the message metadata. Thus, it will be propagated to your different downstream components (as well as the connector).
Note that, Transaction tends to be attached to the request scope, which means that in your connector, it may already be too late to use it. So, make sure your endpoint is asynchronous and only returns when the emitted message is acknowledged.
Context Propagation is not going to help in this case as the underlying streams are built at startup time (at build time in Quarkus) so, there are no capture contexts.
Related
I'm able to make Spring+Rabbit work with the non-functional way (prior to 2.0?), but I'm trying to use with the functional pattern as the previous one is deprecated.
I've been following this doc: https://docs.spring.io/spring-cloud-stream/docs/3.1.0/reference/html/spring-cloud-stream.html#_binding_and_binding_names
The queue (consumer) is not being created in Rabbit with the new method. I can see the connection being created but without any consumer.
I have the following in my application.properties:
spring.cloud.stream.function.bindings.approved-in-0=approved
spring.cloud.stream.bindings.approved.destination=myTopic.exchange
spring.cloud.stream.bindings.approved.group=myGroup.approved
spring.cloud.stream.bindings.approved.consumer.back-off-initial-interval=2000
spring.cloud.stream.rabbit.bindings.approved.consumer.queueNameGroupOnly=true
spring.cloud.stream.rabbit.bindings.approved.consumer.bindingRoutingKey=myRoutingKey
which is replacing:
spring.cloud.stream.bindings.approved.destination=myTopic.exchange
spring.cloud.stream.bindings.approved.group=myGroup.approved
spring.cloud.stream.bindings.approved.consumer.back-off-initial-interval=2000
spring.cloud.stream.rabbit.bindings.approved.consumer.queueNameGroupOnly=true
spring.cloud.stream.rabbit.bindings.approved.consumer.bindingRoutingKey=myRoutingKey
And the new class
#Slf4j
#Service
public class ApprovedReceiver {
#Bean
public Consumer<String> approved() {
// I also saw that it's recommended to not use Consumer, but use Function instead
// https://docs.spring.io/spring-cloud-stream/docs/3.1.0/reference/html/spring-cloud-stream.html#_consumer_reactive
return value -> log.info("value: {}", value);
}
}
which is replacing
// BindableApprovedChannel.class
#Configuration
public interface BindableApprovedChannel {
#Input("approved")
SubscribableChannel getApproved();
}
// ApprovedReceiver.class
#Service
#EnableBinding(BindableApprovedChannel.class)
public class ApprovedReceiver {
#StreamListener("approved")
public void handleMessage(String payload) {
log.info("value: {}", payload);
}
}
Thanks!
If you have multiple beans of type Function, Supplier or Consumer (which could be declared by third party libraries), the framework does not know which one to bind to.
Try setting the spring.cloud.function.definition property to approved.
https://docs.spring.io/spring-cloud-stream/docs/3.1.3/reference/html/spring-cloud-stream.html#spring_cloud_function
In the event you only have single bean of type java.util.function.[Supplier/Function/Consumer], you can skip the spring.cloud.function.definition property, since such functional bean will be auto-discovered. However, it is considered best practice to use such property to avoid any confusion. Some time this auto-discovery can get in the way, since single bean of type java.util.function.[Supplier/Function/Consumer] could be there for purposes other then handling messages, yet being single it is auto-discovered and auto-bound. For these rare scenarios you can disable auto-discovery by providing spring.cloud.stream.function.autodetect property with value set to false.
Gary's answer is correct. If adding the definition property alone doesn't resolve the issue I would recommend sharing what you're doing for your supplier.
This is also a very helpful general discussion for transitioning from imperative to functional with links to repos with more in depth examples: EnableBinding is deprecated in Spring Cloud Stream 3.x
I have a simple listener with 3 methods. and a repository with autowired on that. While saving an object from afterWrite it works nicely. but when saving item from onError methods no exception occurs, however it is not saving any data. Thankful for suggestions.
public class WriteListener implements ItemWriteListener{
public void beforeWrite(List items) {
System.out.println("Going to write following items: "+ items.toString());
}
public void onWriteError(Exception exception, List items) {
System.out.println("Error occurred when writing items!");
testRepository.save(items.get(0)); //not working
}
public void afterWrite(List items) {
testRepository.save(items.get(0)); //not nicely and save data
Based on the limited information provided, most likely the cause is the exception itself. The exception would have marked current transaction as dirty thus spring would have rolled it back.
If you still want to store data in your listener despite existing exception, use it in a separate transaction context. Simplest way for that would be to use #Async annotation on your listener and marking it Transactional explicitly to ensure it initiate a new transaction. Check out Spring Event which covers this topic in little bit more depth.
Note: read the end of the answer for the way I implemented #Nonika's suggestions
What's the "right way" to send a websocket event on data insert?
I'm using a Spring Boot server with SQL/JPA and non-stomp websockets. I need to use "plain" websockets as I'm using Java clients where (AFAIK) there's no stomp support.
When I make a change to the database I need to send the event to the client so I ended up with an implementation like this:
#Transactional
public void addEntity(...) {
performActualEntityAdding();
sendEntityAddedEvent(eventData);
}
#Transactional
public void sendEntityAddedEvent(String eventData) {
TransactionSynchronizationManager.registerSynchronization(new TransactionSynchronizationAdapter() {
#Override
public void afterCommit() {
sendEntityAddedEventAsync(eventData);
}
});
}
#Async
public void sendEntityAddedEventAsync(String eventData) {
// does the websocket session sending...
}
This works. If I would just call the sendEntityAddedEventAsync it would also work for real world scenarios but it fails on unit tests because the event would arrive before transaction commit. As such when the unit test invokes a list of the entities after the event it fails.
This feels like a hack that shouldn't be here. Is there a better way to ensure a commit?
I tried multiple alternative approaches and the problem is that they often worked for 10 runs of the unit tests yet failed every once in a while. That isn't acceptable.
I tried multiple approaches to solve this such as different transaction annotations and splitting the method to accommodate them. E.g read uncommitted, not supported (to force a commit) etc. Nothing worked for all cases and I couldn't find an authoritative answer for this (probably common) use case that wasn't about STOMP (which is pretty different).
Edit
One of my original attempts looked something like this:
// this shouldn't be in a transaction
public void addEntity(...) {
performActualEntityAdding();
sendEntityAddedEvent(eventData);
}
#Transactional
public void performActualEntityAdding(...) {
//....
}
#Async
public void sendEntityAddedEventAsync(String eventData) {
// does the websocket session sending...
}
The assumption here is that when sendEntityAddedEventAsync is invoked the data would already be in the database. It wasn't for a couple of additional milliseconds.
A few additional details:
Test environment is based on h2 (initially I mistakenly wrote hsql)
Project is generated by JHipster
Level 2 cache is used but disabled as NONE for these entities
Solution (based on #Nonika's answer):
The solution for me included something similar to this:
public class WebEvent extends ApplicationEvent {
private ServerEventDAO event;
public WebEvent(Object source, ServerEventDAO event) {
super(source);
this.event = event;
}
public ServerEventDAO getEvent() {
return event;
}
}
#Transactional
public void addEntity(...) {
performActualEntityAdding();
applicationEventPublisher.publishEvent(new WebEvent(this, evtDao));
}
#Async
#TransactionalEventListener
public void sendEntityAddedEventAsync(WebEvent eventData) {
// does the websocket session sending...
}
This effectively guarantees that the data is committed properly before sending the event and it runs asynchronously to boot. Very nice and simple.
Spring is using AdviceMode.PROXY for both #Async and #Transactional this is quote from the javadoc:
The default is AdviceMode.PROXY. Please note that proxy mode allows
for interception of calls through the proxy only. Local calls within
the same class cannot get intercepted that way; an Async annotation on
such a method within a local call will be ignored since Spring's
interceptor does not even kick in for such a runtime scenario. For a
more advanced mode of interception, consider switching this to
AdviceMode.ASPECTJ.
This rule is common for almost all spring annotations which requires proxy to operate.
Into your first example, you have a #Transactional annotation on both addEntity(..) and performActualEntityAdding(..). I suppose you call addEntity from another class so #Transactional works as expected. process in this scenario can be described in this flow
// -> N1 transaction starts
addEntity(){
performActualEntityAdding()//-> we are still in transaction N1
sendEntityAddedEvent() // -> call to this #Async is a class local call, so this advice is ignored. But if this was an async call this would not work either.
}
//N1 transaction commits;
That's why the test fails. it gets an event that there is a change into the db, but there is nothing because the transaction has not been committed yet.
Scenario 2.
When you don't have a #Transactional addEntity(..) then second transaction for performActualEntityAdding not starts as there is a local call too.
Options:
You can use some middleware class to call these methods to trigger
spring interceptors.
you can use Self injection with Spring
if you have Spring 5.0 there is handy #TransactionalEventListener(phase = TransactionPhase.AFTER_COMMIT)
I'm using jpa , spring data and entity listeners to audit my entities precisely on postUpdate , postPersist , PostRemove
This is a pseudo code of my entity listener class
public class EntityListener extends AuditingEntityListener {
#PostUpdate
public void postPersist(Object auditedEntity) {
writer.saveEntity(auditedEntity,"UPDATE");
}
This the pseudo code of the Writer class
public class Writer {
#Async
public void saveEntity(Object auditedEntity, String action) {
try {
//some code to prepare the history entity
historyDAO.save(entity);
} catch (Exception e) {
}
}
when an exception is thrown in Writer class , the auditedEntity is updated or inserted however the historyEntity where i store the audit action doesnt
The problem is i need to invoke the saveEntity method in another thread for performance issue (#Async) but in that case a new transaction is open instead of the previously one which opened
how can i solve the rollack issue for both transactions
so when an exception is throwen both historyEntity and auditedEntity not persisted
I understand that you want to rollback both the child and the parent transaction when an exception is thrown from within Writer.saveEntity.
The problem is that the thread with the original transaction would still need to wait for all these complicated operations to finish before it could mark the transaction as committed. You can't easily span a transaction across multiple threads, either.
The only thing you could probably do to speed things up is you could run the logic of generating the history entities in parallel, and then save them all just before the transaction commits.
One way of doing that that I can think of is using a Hibernate interceptor:
public class AuditInterceptor extends EmptyInterceptor {
private List<Callable<BaseEntity>> historyEntries;
private ExecutorService executor;
...
public void beforeTransactionCompletion(Transaction tx) {
List<Future<BaseEntity>> futures = executor.invokeAll(historyEntries);
if (executor.awaitTermination(/* some timeout here */)) {
futures.stream().map(Future::get).forEach(entity -> session.save(object));
} else {
/* rollback */
}
}
}
Your listener code then becomes:
#PostUpdate
public void postPersist(Object auditedEntity) {
interceptor.getHistoryEntries().add(new Callable<BaseEntity> {
/* history entry generation logic goes here */
});
}
(note that the above code is greatly simplified, you could use any other asynchronous execution API, the basic idea is that you need to block in AuditInterceptor.beforeTransactionCompletion, waiting for all the history entries to be generated)
However, I would strongly advise against using the above technique, as it is rather complicated and error prone.
If you look here: https://docs.jboss.org/hibernate/orm/5.1/userguide/html_single/chapters/events/Events.html, you'll find that Hibernate interceptors have more interesting methods that could help you gather auditing info, and that perhaps your implementation could make use of them, possibly avoiding the need for complicated logic altogether (Hibernate already does track changes to fields of individual entities, so you get that information for free).
Why reinvent the wheel, though? If you dig even deeper, you'll find the Hibernate Envers module (http://hibernate.org/orm/envers/, works for both JPA and pure Hibernate) which gives you business auditing out of the box. Envers already digs into the above mechanism, so hopefully the performance issue would go away.
Final note: have you measured how long history entry generation takes? I would guess that executing for loops and if statements might be cheaper than database access operations. If I were you, I wouldn't do any of the above unless I was absolutely sure that's where the performance bottleneck was.
UPD 1: Upon further research I think the following information may be useful:
I obtain datasource through JNDI lookup on WildFly 9.0.2, then 'wrap' it into in instance of HikariDataSource (e. g. return new HikariDataSource(jndiDSLookup(dsName))).
the transaction manager that ends up being used is JTATransactionManager.
I do not configure the transaction manager in any way.
ORIGINAL QUESTION:
I am experiencing an issue with JPA/Hibernate and (maybe) Spring-Boot where DB changes introduced in a transactional method of one class called from a transactional method of another class are committed even though the changes in the caller method are rolled back (as they should be).
Here are my transactional services
StuffService:
#Service
#Transactional(rollbackFor = IOException.class)
public class StuffService {
#Inject private BarService barService;
#Inject private StuffRepository stuffRepository;
public Stuff updateStuff(Stuff stuff) {
try {
if (null != barService.doBar(stuff)) {
stuff.setSomething(SOMETHING);
stuff.setSomethingElse(SOMETHING_ELSE);
return stuffRepository.save(stuff);
}
} catch (FirstCustomException e) {
logger.error("Blah", e);
throw new SecondCustomException(e.getMessage());
}
throw new SecondCustomException("Blah 2");
}
// other methods
}
and BarService:
#Service
#Transactional
public class BarService {
#Inject private EntityARepository entityARepository;
#Inject private EntityBRepository entityBRepository;
/*
* updates existing entity A and persists new entity B.
*/
public EntityA doBar(Stuff stuff) throws FirstCustomException {
EntityA a = entityARepository.findOne(/* some criteria */);
a.setSomething(SOMETHING);
EntityB b = new EntityB();
b.setSomething(SOMETHING);
b.setSomethingElse(SOMETHING_ELSE);
entityBRepository.save(b);
return entityARepository.save(a);
}
// other methods
}
EntityARepository and EntityBRepository are very similar Spring-Boot repositories defined like this:
public interface EntityARepository extends JpaRepository<EntityA, Long>{
EntityA findOne(/* some criteria */);
}
FirstCustomException extends Throwable
SecondCustomException extends RuntimeException
Stuff entity is versioned, and every once in a while it is concurrently updated by StuffService.updateStuff(). In that case changes to one of the stuff instances are rolled back, as expected, but everything that happens in the barService.doBar() ends up being committed.
This puzzles me quite a lot since transaction propagation on both methods should be REQUIRED (the default one) and both methods belong to different classes, hence #Transactional should apply for both.
I did see Transaction is not completely rolled back after server throws OptimisticLockException1
But it did not really answer my question.
Can anyone please give me an idea of what's going on?
Thank you.
This isn't a 'nested' transaction - these services are operating in completely independent transactions. If you want the rollback of one to affect the other, you need to have them take part in the same transaction rather than start its own.
Or if your issue is that there is a problem with the version of 'stuff' passed into the doBar method and you want it verified, you will need to do something with the stuff instance that would cause an optimistic lock check, and so result in an exception if it is stale. see EntityManager.lock