Sessions & Transactions

As of version 3.6, MongoDB supports the concept of sessions. The use of sessions enables MongoDB’s Causal Consistency model, which guarantees running operations in an order that respects their causal relationships. Those are split into ServerSession instances and ClientSession instances. In this section, when we speak of a session, we refer to ClientSession.

Operations within a client session are not isolated from operations outside the session.

Both MongoOperations and ReactiveMongoOperations provide gateway methods for tying a ClientSession to the operations. MongoCollection and MongoDatabase use session proxy objects that implement MongoDB’s collection and database interfaces, so you need not add a session on each call. This means that a potential call to MongoCollection#find() is delegated to MongoCollection#find(ClientSession).

Methods such as (Reactive)MongoOperations#getCollection return native MongoDB Java Driver gateway objects (such as MongoCollection) that themselves offer dedicated methods for ClientSession. These methods are NOT session-proxied. You should provide the ClientSession where needed when interacting directly with a MongoCollection or MongoDatabase and not through one of the #execute callbacks on MongoOperations.

ClientSession support

The following example shows the usage of a session:

  • Imperative

  • Reactive

ClientSessionOptions sessionOptions = ClientSessionOptions.builder()
    .causallyConsistent(true)
    .build();

ClientSession session = client.startSession(sessionOptions); (1)

template.withSession(() -> session)
    .execute(action -> {

        Query query = query(where("name").is("Durzo Blint"));
        Person durzo = action.findOne(query, Person.class);  (2)

        Person azoth = new Person("Kylar Stern");
        azoth.setMaster(durzo);

        action.insert(azoth);                                (3)

        return azoth;
    });

session.close()                                              (4)
1 Obtain a new session from the server.
2 Use MongoOperation methods as before. The ClientSession gets applied automatically.
3 Make sure to close the ClientSession.
4 Close the session.
When dealing with DBRef instances, especially lazily loaded ones, it is essential to not close the ClientSession before all data is loaded. Otherwise, lazy fetch fails.
ClientSessionOptions sessionOptions = ClientSessionOptions.builder()
.causallyConsistent(true)
.build();

Publisher<ClientSession> session = client.startSession(sessionOptions); (1)

template.withSession(session)
.execute(action -> {

        Query query = query(where("name").is("Durzo Blint"));
        return action.findOne(query, Person.class)
            .flatMap(durzo -> {

                Person azoth = new Person("Kylar Stern");
                azoth.setMaster(durzo);

                return action.insert(azoth);                            (2)
            });
    }, ClientSession::close)                                            (3)
    .subscribe();                                                       (4)
1 Obtain a Publisher for new session retrieval.
2 Use ReactiveMongoOperation methods as before. The ClientSession is obtained and applied automatically.
3 Make sure to close the ClientSession.
4 Nothing happens until you subscribe. See the Project Reactor Reference Guide for details.

By using a Publisher that provides the actual session, you can defer session acquisition to the point of actual subscription. Still, you need to close the session when done, so as to not pollute the server with stale sessions. Use the doFinally hook on execute to call ClientSession#close() when you no longer need the session. If you prefer having more control over the session itself, you can obtain the ClientSession through the driver and provide it through a Supplier.

Reactive use of ClientSession is limited to Template API usage. There’s currently no session integration with reactive repositories.

MongoDB Transactions

As of version 4, MongoDB supports Transactions. Transactions are built on top of Sessions and, consequently, require an active ClientSession.

Unless you specify a MongoTransactionManager within your application context, transaction support is DISABLED. You can use setSessionSynchronization(ALWAYS) to participate in ongoing non-native MongoDB transactions.

To get full programmatic control over transactions, you may want to use the session callback on MongoOperations.

The following example shows programmatic transaction control:

Programmatic transactions
  • Imperative

  • Reactive

ClientSession session = client.startSession(options);                   (1)

template.withSession(session)
    .execute(action -> {

        session.startTransaction();                                     (2)

        try {

            Step step = // ...;
            action.insert(step);

            process(step);

            action.update(Step.class).apply(Update.set("state", // ...

            session.commitTransaction();                                (3)

        } catch (RuntimeException e) {
            session.abortTransaction();                                 (4)
        }
    }, ClientSession::close)                                            (5)
1 Obtain a new ClientSession.
2 Start the transaction.
3 If everything works out as expected, commit the changes.
4 Something broke, so roll back everything.
5 Do not forget to close the session when done.

The preceding example lets you have full control over transactional behavior while using the session scoped MongoOperations instance within the callback to ensure the session is passed on to every server call. To avoid some of the overhead that comes with this approach, you can use a TransactionTemplate to take away some of the noise of manual transaction flow.

Mono<DeleteResult> result = Mono
    .from(client.startSession())                                                             (1)

    .flatMap(session -> {
        session.startTransaction();                                                          (2)

        return Mono.from(collection.deleteMany(session, ...))                                (3)

            .onErrorResume(e -> Mono.from(session.abortTransaction()).then(Mono.error(e)))   (4)

            .flatMap(val -> Mono.from(session.commitTransaction()).then(Mono.just(val)))     (5)

            .doFinally(signal -> session.close());                                           (6)
      });
1 First we obviously need to initiate the session.
2 Once we have the ClientSession at hand, start the transaction.
3 Operate within the transaction by passing on the ClientSession to the operation.
4 If the operations completes exceptionally, we need to stop the transaction and preserve the error.
5 Or of course, commit the changes in case of success. Still preserving the operations result.
6 Lastly, we need to make sure to close the session.

The culprit of the above operation is in keeping the main flows DeleteResult instead of the transaction outcome published via either commitTransaction() or abortTransaction(), which leads to a rather complicated setup.

Unless you specify a ReactiveMongoTransactionManager within your application context, transaction support is DISABLED. You can use setSessionSynchronization(ALWAYS) to participate in ongoing non-native MongoDB transactions.

Transactions with TransactionTemplate / TransactionalOperator

Spring Data MongoDB transactions support both TransactionTemplate and TransactionalOperator.

Transactions with TransactionTemplate / TransactionalOperator
  • Imperative

  • Reactive

template.setSessionSynchronization(ALWAYS);                                     (1)

// ...

TransactionTemplate txTemplate = new TransactionTemplate(anyTxManager);         (2)

txTemplate.execute(new TransactionCallbackWithoutResult() {

    @Override
    protected void doInTransactionWithoutResult(TransactionStatus status) {     (3)

        Step step = // ...;
        template.insert(step);

        process(step);

        template.update(Step.class).apply(Update.set("state", // ...
    };
});
1 Enable transaction synchronization during Template API configuration.
2 Create the TransactionTemplate using the provided PlatformTransactionManager.
3 Within the callback the ClientSession and transaction are already registered.
Changing state of MongoTemplate during runtime (as you might think would be possible in item 1 of the preceding listing) can cause threading and visibility issues.
template.setSessionSynchronization(ALWAYS);                                          (1)

// ...

TransactionalOperator rxtx = TransactionalOperator.create(anyTxManager,
                                   new DefaultTransactionDefinition());              (2)


Step step = // ...;
template.insert(step);

Mono<Void> process(step)
    .then(template.update(Step.class).apply(Update.set("state", …))
    .as(rxtx::transactional)                                                         (3)
    .then();
1 Enable transaction synchronization for Transactional participation.
2 Create the TransactionalOperator using the provided ReactiveTransactionManager.
3 TransactionalOperator.transactional(…) provides transaction management for all upstream operations.

Transactions with MongoTransactionManager & ReactiveMongoTransactionManager

MongoTransactionManager / ReactiveMongoTransactionManager is the gateway to the well known Spring transaction support. It lets applications use the managed transaction features of Spring. The MongoTransactionManager binds a ClientSession to the thread whereas the ReactiveMongoTransactionManager is using the ReactorContext for this. MongoTemplate detects the session and operates on these resources which are associated with the transaction accordingly. MongoTemplate can also participate in other, ongoing transactions. The following example shows how to create and use transactions with a MongoTransactionManager:

Transactions with MongoTransactionManager / ReactiveMongoTransactionManager
  • Imperative

  • Reactive

@Configuration
static class Config extends AbstractMongoClientConfiguration {

    @Bean
    MongoTransactionManager transactionManager(MongoDatabaseFactory dbFactory) {  (1)
        return new MongoTransactionManager(dbFactory);
    }

    // ...
}

@Component
public class StateService {

    @Transactional
    void someBusinessFunction(Step step) {                                        (2)

        template.insert(step);

        process(step);

        template.update(Step.class).apply(Update.set("state", // ...
    };
});
1 Register MongoTransactionManager in the application context.
2 Mark methods as transactional.
@Transactional(readOnly = true) advises MongoTransactionManager to also start a transaction that adds the ClientSession to outgoing requests.
@Configuration
public class Config extends AbstractReactiveMongoConfiguration {

    @Bean
    ReactiveMongoTransactionManager transactionManager(ReactiveMongoDatabaseFactory factory) {  (1)
        return new ReactiveMongoTransactionManager(factory);
    }

    // ...
}

@Service
public class StateService {

    @Transactional
    Mono<UpdateResult> someBusinessFunction(Step step) {                                  (2)

        return template.insert(step)
            .then(process(step))
            .then(template.update(Step.class).apply(Update.set("state", …));
    };
});
1 Register ReactiveMongoTransactionManager in the application context.
2 Mark methods as transactional.
@Transactional(readOnly = true) advises ReactiveMongoTransactionManager to also start a transaction that adds the ClientSession to outgoing requests.

Special behavior inside transactions

Inside transactions, MongoDB server has a slightly different behavior.

Connection Settings

The MongoDB drivers offer a dedicated replica set name configuration option turing the driver into auto detection mode. This option helps identifying the primary replica set nodes and command routing during a transaction.

Make sure to add replicaSet to the MongoDB URI. Please refer to connection string options for further details.

Collection Operations

MongoDB does not support collection operations, such as collection creation, within a transaction. This also affects the on the fly collection creation that happens on first usage. Therefore make sure to have all required structures in place.

Transient Errors

MongoDB can add special labels to errors raised during transactional operations. Those may indicate transient failures that might vanish by merely retrying the operation. We highly recommend Spring Retry for those purposes. Nevertheless one may override MongoTransactionManager#doCommit(MongoTransactionObject) to implement a Retry Commit Operation behavior as outlined in the MongoDB reference manual.

Count

MongoDB count operates upon collection statistics which may not reflect the actual situation within a transaction. The server responds with error 50851 when issuing a count command inside of a multi-document transaction. Once MongoTemplate detects an active transaction, all exposed count() methods are converted and delegated to the aggregation framework using $match and $count operators, preserving Query settings, such as collation.

Restrictions apply when using geo commands inside of the aggregation count helper. The following operators cannot be used and must be replaced with a different operator:

  • $where$expr

  • $near$geoWithin with $center

  • $nearSphere$geoWithin with $centerSphere

Queries using Criteria.near(…) and Criteria.nearSphere(…) must be rewritten to Criteria.within(…) respective Criteria.withinSphere(…). Same applies for the near query keyword in repository query methods that must be changed to within. See also MongoDB JIRA ticket DRIVERS-518 for further reference.

The following snippet shows count usage inside the session-bound closure:

session.startTransaction();

template.withSession(session)
    .execute(action -> {
        action.count(query(where("state").is("active")), Step.class)
        ...

The snippet above materializes in the following command:

db.collection.aggregate(
   [
      { $match: { state: "active" } },
      { $count: "totalEntityCount" }
   ]
)

instead of:

db.collection.find( { state: "active" } ).count()