Hazelcast and Spring-managed Transactions: A Sample Integration

Implementing a Transactional Receiver

See the first post for a basic introduction on Hazelcast and Spring Boot integration.

Hazelcast Transactions

Hazelcast, differently from some other No-SQL, in-memory stores, has the ability to make access to distributed collection ACID (… at least when you are not in a split-brain condition, but this is fair enough).Hazelcast supports transactions in two ways:

1. Transactional distributed collections; Map, Multimap, Queue, Set and List.
2. Transactional distributed tasks: Automatically wraps the execution in a transaction.

Transaction and locking behaviour is reasonably similar to RDBMS we are al used to: While in a transaction, writing a distributed collection entry take a lock on it; reading doesn’t take a lock unless you do it explicitly with a “get for update” (e.g. TransactionalMap.getForUpdate(...)), remarkably similar to SQL SELECT … FOR UPDATE.Beyond locks, all updates are recorded in a “transaction log” and applied only on commit, giving a REPEATABLE_READ isolation by default.You may also find some similarities with the JPA EntityManager/Hibernate Session, as accessed entries are locally cached during the transaction. But beware similarities end here. For example, modifying properties of a distributed collection entry does NOT modify the actual distributed object, as you have always a copy (not a proxy) of the actual entry.

The Not-so-transparent Transactional Context and Transactional Collections

If you are accustomed to the Spring’s @Transactional and TransactionManager, Hazelcast transactions will appear remarkably non-transparent:Here is an example from the official guide:

‍

HazelcastInstance hz = Hazelcast.newHazelcastInstance();
TransactionContext txCxt = hz.newTransactionContext();
TransactionalMap<String,Employee> employees = context.getMap(“employees”);

employees.put(“1”,new Employee());

txCxt.commitTransaction();
‍

IMap employeesIMap = hz.getMap(“employees”);
System.out.println(employeesIMap.get(“1”));

As you may see, transactional collections are retrieved through a Transactional Context. Note that the interface of the transactional collections are different from the non-transactional versions (e.g. TransactionalMap vs IMap). Apparently, this is a API design decision: The developer has to know he’s working with a transactional collections, due to performance implications. This make sense, but makes practically impossible to transparently inject transactions à la Spring style.

A Transactional Consumer

So, get back to the sample chat application I introduced in the first post of this series.My specific (but not so uncommon) problem, is making message polling more robust. If something happens during the poll and transmission, I want the messages to remain in the queue.

I’m focusing on the transactional polling: The green dotted line on sending the messages to the recipient.First of all, I have to use a TransactionalQueue. Note that Queue, differently from other collections, are transactional on read too, as polling actually modify the content of the Queue. When transactional, polls are recorded in a “transaction log”, entries are actually removed only on commit.But a transactional Queue will not suffice. I have to wrap the whole poll request (including converting, serialising and sending messages to the client) in a try - catch handling the transaction externally. The way Spring does annotating the Controller method @Transactional.So I have to make the Hazelcast transaction Spring-managed.

No out-of-the-box Spring-managed Transaction

Although transactions may integrate with (heavyweight) XA, there is no support for Spring managed transactions. A feature request is long-pending.I was short of time for implementing it from scratch, So I googled for some open source project and I found the remarkable HazelcastMQ, by Mike Pilone: https://github.com/mpilone/hazelcastmq.His goal is much broader than mine. HazelcastMQ provides a high-level messaging layer above Hazelcast. But what interested me was the integration with Spring-managed transactions.So I extracted just a couple of classes from HazelcastMQ, adapting them to my limited scope.

• HazelcastTransactionManager: Implements Spring’s PlatformTransactionManager, binding the Hazelcast Transaction Context to the current thread. Runtime, it depends on the HazelcastInstance and has to be initialised as Spring Bean.
• HazelcastUtils: Utility class used by the application to retrieve the transactional objects (currently, only Queues are available).

Back to Code

The sample chat application, upgrade for supporting transactional polling, is in the transactions branch of the repository:https://github.com/nicusX/springboot-hazelcast-example/tree/transactionsFor comparison, both transactional and non-transactional versions are available. Which implementation to use is controlled by the presence of the “transactional-polling” Spring Bean Profile.Registering the HazelcastTransactionManager makes Spring managing the Hazelcast transaction. Annotating the controller method @Transactional, make it rollback (and returning the polled elements into the Queue) on any exception, including those caused by response mapping and serialisation.

@Transactional
@RequestMapping(value = "/recipients/{recipient}/poll", method = RequestMethod.GET)
public List receiveAll(@PathVariable("recipient") String recipient) {
  final List messages = chatService.receive(recipient);
  return messages.stream().map( ChatController::map ).collect(Collectors.toList());
}

Only for demonstrating the transactional implementation, the controller throws an exception on receiving a “poison pill”: A message with EPOCH timestamp:

if ( message.getTimestamp().equals(Instant.EPOCH)) {
  throw new RuntimeException("Simulated Exception");
}

Service

The implementation of the service method is the same both for transactional and non-transactional polling. Note I had to use BaseQueue as only common interface of both IQueue and TransactionalQueue.

@Override
public List receive(String recipient) {
  final BaseQueue recipientQueue = recipientQueueForPolling(recipient);
  final List messages = new ArrayList<>();
  while ( true ) {
      final ChatMessage message = recipientQueue.poll();
      if ( message == null ) break;
      messages.add(message);
  }
  return Collections.unmodifiableList(messages);
}

The non-transactional version simply retrieves the IQueue (as BaseQueue) from HazelcastInstance:

// Non transactional Queue
protected BaseQueue recipientQueueForPolling(String recipient) {
  return hazelcastInstance.getQueue(“recipient-” + recipient);
}

While the transactional version uses the custom HazelcastUtils, that returns a TransactionalQueue:

// Transactional Queue
protected BaseQueue recipientQueueForPolling(String recipient) {
  return HazelcastUtils.getTransactionalQueue( “recipient-” + recipient, hazelcastInstance, true);
}

Conclusion: A half-successful Integration

This transaction integration works fine for my limited goal of implementing transactional polling. However, is very limited and has a lot of space for improvement. It currently supports only a transactional queue, but extending  HazelcastUtils to support other collections should not be hard.I also haven’t tested the synchronisation with other Spring-managed transactions. For example JPA or JMS. I would possibly work on this subject for a future post.HazelcastMQ also provides a second integration pattern. It uses a factory, the snappily named  TransactionAwareHazelcastInstanceProxyFactory, to provide a wrapped HazelcastInstance returning transactional collections bound to the current Spring-managed transaction. To be honest, I wasn’t able to make it work, and I wasn’t super-happy with the solutions required.More specifically, the two HazelcastInstance beans (for the same actual instance) you have to handle: One for transactional and another for non-transactional collections. There is also an “elegant hack”, proxying transactional collections to make them implementing non-transactional interfaces.Stay-tuned to the OpenCredo blog for the next installation of my exploration into Spring and Hazelcast!This blog is written exclusively by the OpenCredo team. We do not accept external contributions.