Spring is the most popular Java framework. It has lots of out-of-box solutions for web, security, caching, and data access. Spring Data especially makes the life of a developer much easier. We don’t have to worry about database connections and transaction management. The framework does the job. But the fact that it hides some important details from us may lead to hard-tracking bugs and issues. So, let’s deep dive into @Transactional annotation.
Default Rollback Behaviour
Assume that we have a simple service method that creates 3 users during one transaction. If something goes wrong, it throws java.lang.Exception.
@Service
public class PersonService {
@Autowired
private PersonRepository personRepository;
@Transactional
public void addPeople(String name) throws Exception {
personRepository.saveAndFlush(new Person("Jack", "Brown"));
personRepository.saveAndFlush(new Person("Julia", "Green"));
if (name == null) {
throw new Exception("name cannot be null");
}
personRepository.saveAndFlush(new Person(name, "Purple"));
}
}
And here is a simple unit test.
@SpringBootTest
@AutoConfigureTestDatabase
class PersonServiceTest {
@Autowired
private PersonService personService;
@Autowired
private PersonRepository personRepository;
@BeforeEach
void beforeEach() {
personRepository.deleteAll();
}
@Test
void shouldRollbackTransactionIfNameIsNull() {
assertThrows(Exception.class, () -> personService.addPeople(null));
assertEquals(0, personRepository.count());
}
}
Do you think the test will pass or not? Logic tells us that Spring should roll back the transaction due to an exception. So, personRepository.count() ought to return 0, right? Well, not exactly.
expected: <0> but was: <2>
Expected :0
Actual :2
That requires some explanations. By default, Spring rolls back transaction only if an unchecked exception occurs. The checked ones are treated like restorable. In our case, Spring performs commit instead of rollback. That’s why personRepository.count() returns 2.
The easiest way to fix it is to replace a checked exception with an unchecked one (e.g., NullPointerException). Or else we can use the annotation’s attribute rollbackFor.
For example, both of these cases are perfectly valid.
@Service
public class PersonService {
@Autowired
private PersonRepository personRepository;
@Transactional(rollbackFor = Exception.class)
public void addPeopleWithCheckedException(String name) throws Exception {
addPeople(name, Exception::new);
}
@Transactional
public void addPeopleWithNullPointerException(String name) {
addPeople(name, NullPointerException::new);
}
private <T extends Exception> void addPeople(String name, Supplier<? extends T> exceptionSupplier) throws T {
personRepository.saveAndFlush(new Person("Jack", "Brown"));
personRepository.saveAndFlush(new Person("Julia", "Green"));
if (name == null) {
throw exceptionSupplier.get();
}
personRepository.saveAndFlush(new Person(name, "Purple"));
}
}
@SpringBootTest
@AutoConfigureTestDatabase
class PersonServiceTest {
@Autowired
private PersonService personService;
@Autowired
private PersonRepository personRepository;
@BeforeEach
void beforeEach() {
personRepository.deleteAll();
}
@Test
void testThrowsExceptionAndRollback() {
assertThrows(Exception.class, () -> personService.addPeopleWithCheckedException(null));
assertEquals(0, personRepository.count());
}
@Test
void testThrowsNullPointerExceptionAndRollback() {
assertThrows(NullPointerException.class, () -> personService.addPeopleWithNullPointerException(null));
assertEquals(0, personRepository.count());
}
}
Rollback on Exception Suppressing
Not all exceptions have to be propagated. Sometimes it is acceptable to catch it and log information about it.
Suppose that we have another transactional service that checks whether the person can be created with the given name. If it is not, it throws IllegalArgumentException.
@Service
public class PersonValidateService {
@Autowired
private PersonRepository personRepository;
@Transactional
public void validateName(String name) {
if (name == null || name.isBlank() || personRepository.existsByFirstName(name)) {
throw new IllegalArgumentException("name is forbidden");
}
}
}
Let’s add validation to our PersonService.
@Service
@Slf4j
public class PersonService {
@Autowired
private PersonRepository personRepository;
@Autowired
private PersonValidateService personValidateService;
@Transactional
public void addPeople(String name) {
personRepository.saveAndFlush(new Person("Jack", "Brown"));
personRepository.saveAndFlush(new Person("Julia", "Green"));
String resultName = name;
try {
personValidateService.validateName(name);
}
catch (IllegalArgumentException e) {
log.error("name is not allowed. Using default one");
resultName = "DefaultName";
}
personRepository.saveAndFlush(new Person(resultName, "Purple"));
}
}
If validation does not pass, we create a new person with the default name.
Ok, now we need to test it.
@SpringBootTest
@AutoConfigureTestDatabase
class PersonServiceTest {
@Autowired
private PersonService personService;
@Autowired
private PersonRepository personRepository;
@BeforeEach
void beforeEach() {
personRepository.deleteAll();
}
@Test
void shouldCreatePersonWithDefaultName() {
assertDoesNotThrow(() -> personService.addPeople(null));
Optional<Person> defaultPerson = personRepository.findByFirstName("DefaultName");
assertTrue(defaultPerson.isPresent());
}
}
But the result is rather unexpected.
Unexpected exception thrown: org.springframework.transaction.UnexpectedRollbackException: Transaction silently rolled back because it has been marked as rollback-only
That’s weird. The exception has been suppressed. Why did Spring roll back the transaction? Firstly, we need to understand the @Transactional management approach.
Internally Spring uses the aspect-oriented programming pattern. Skipping the complex details, the idea behind it is to wrap an object with the proxy that performs the required operations (in our case, transaction management). So, when we inject the service that has any @Transactional method, actually Spring puts the proxy.
Here is the workflow for the defined addPeople method.
The default @Transactional propagation is REQUIRED. It means that the new transaction is created if it’s missing. And if it’s present already, the current one is supported. So, the whole request is being executed within a single transaction.
Anyway, there is a caveat. If the RuntimeException throws out of the transactional proxy, Spring marks the current transaction as rollback only. That’s exactly what happened in our case. PersonValidateService.validateName throws IllegalArgumentException. Transactional proxy tracks it and sets on the rollback flag. Later executions during the transaction have no effect because they ought to be rolled back in the end.
What’s the solution? There are several ones. For example, we can add noRollbackFor attribute to PersonValidateService.
@Service
public class PersonValidateService {
@Autowired
private PersonRepository personRepository;
@Transactional(noRollbackFor = IllegalArgumentException.class)
public void validateName(String name) {
if (name == null || name.isBlank() || personRepository.existsByFirstName(name)) {
throw new IllegalArgumentException("name is forbidden");
}
}
}
Another approach is to change the transaction propagation to REQUIRES_NEW. In this case, PersonValidateService.validateName will be executed in a separate transaction. So, the parent one will not be rollbacked.
@Service
public class PersonValidateService {
@Autowired
private PersonRepository personRepository;
@Transactional(propagation = Propagation.REQUIRES_NEW)
public void validateName(String name) {
if (name == null || name.isBlank() || personRepository.existsByFirstName(name)) {
throw new IllegalArgumentException("name is forbidden");
}
}
}
Possible Kotlin Issues
Kotlin has many common things with Java. But exception management is not the case.
Kotlin eliminated the idea of checked and unchecked exceptions. Basically, any exception in the language is unchecked because we don’t need to specify throws SomeException in the method declaration. The pros and cons of this decision should be a topic for another story. But now I want to show you the problems it may bring with Spring Data usage.
Let’s rewrite the very first example of the article with java.lang.Exception in Kotlin.
@Service
class PersonService(
@Autowired
private val personRepository: PersonRepository
) {
@Transactional
fun addPeople(name: String?) {
personRepository.saveAndFlush(Person("Jack", "Brown"))
personRepository.saveAndFlush(Person("Julia", "Green"))
if (name == null) {
throw Exception("name cannot be null")
}
personRepository.saveAndFlush(Person(name, "Purple"))
}
}
@SpringBootTest
@AutoConfigureTestDatabase
internal class PersonServiceTest {
@Autowired
lateinit var personRepository: PersonRepository
@Autowired
lateinit var personService: PersonService
@BeforeEach
fun beforeEach() {
personRepository.deleteAll()
}
@Test
fun `should rollback transaction if name is null`() {
assertThrows(Exception::class.java) { personService.addPeople(null) }
assertEquals(0, personRepository.count())
}
}
The test fails just like in Java.
expected: <0> but was: <2>
Expected :0
Actual :2
There are no surprises. Spring manages transactions in the same way in either Java or Kotlin. But in Java, we cannot execute a method that throws java.lang.Exception without taking care of it. Kotlin allows it. That may bring unexpected bugs, so, you should pay extra attention to such cases.
Conclusion
That’s all I wanted to tell you about Spring @Transactional annotation. If you have any questions or suggestions, please leave your comments down below. Thanks for reading!
Top comments (3)
Hi Semyon,
Thanks a lot for the article, it’s very demonstrative and useful.
One question: will removal of @Transactional annotation from the method validateName also solve considered problem?
@dimosn
Hi Dimos,
Thank you! In this case, the removal of
@Transactionalwill definitely help to resolve the issue. Because the transaction is started anyway onaddPeoplemethod call.But you should take into account that the
validateNamemethod might be complex and consists of several database calls. Also, there is a chance that someone might call the method from non-transactional context. This would lead to multiple transactions during one business operation which can cause the inconsistent final result.Therefore, knowing the rollback behavior patterns is useful anyways :)
Yes, sure, Semyon, here the content of the
validateNamemethod is simplified for clarity, I can imagine :) Thanks!