Prior to this commit, a Testcontainer that was managed as a bean
would not have been started in time if it was accessed before
the bean factory's configuration had been frozen. A common way
for this to occur is when using JPA. The entity manager factory
bean is LoadTimeWeaverAware which causes it to be created before
configuration is frozen. Creating this bean requires the DataSource
which in turn requires the JdbcConnectionDetails and its JDBC URL.
Getting the JDBC URL From the connection details requires the
container hosting the SQL database to have been started.
This commit updates ContainerConnectionDetails, the super-class for
all Testcontainer-based ConnectionDetails implementations, to publish
an event when the Container is retrieved from the details. When this
event is published, TestcontainersLifecycleBeanPostProcessor
initializes all containers that are defined as beans.
Closes gh-40585
Update `TestcontainersLifecycleBeanPostProcessor` so that containers
are now initialized either on the first `postProcessAfterInitialization`
call with a frozen configuration or just before a test container
property is supplied.
Prior to this commit, it was assumed that the first post-process call
after the configuration was frozen was suitably early to initialize
the containers. This turns out to not be no always the case.
Specifically, in the `finishBeanFactoryInitialization` method of
`AbstractApplicationContext` we see that `LoadTimeWeaverAware` beans
are obtained before the configuration is frozen. One such bean is
`DefaultPersistenceUnitManager` which is likely to need datasource
properties that will require a started container.
To fix the problem, the `TestcontainersPropertySource` now publishes
a `BeforeTestcontainersPropertySuppliedEvent` to the ApplicationContext
just before any value is supplied. By listening for this event, we can
ensure that containers are initialized and started before any dynamic
property is read.
Fixes gh-38913
Update `TestcontainersLifecycleBeanPostProcessor` so that containers
can actually be started in parallel.
Prior to this commit, `initializeStartables` would collect beans
and in the process trigger the `postProcessAfterInitialization` method
on each bean. This would see that `startablesInitialized` was `true`
and call `startableBean.start` directly. The result of this was that
beans were actually started sequentially and when the `start` method
was finally called it had nothing to do.
The updated code uses an enum rather than a boolean so that the
`postProcessAfterInitialization` method no longer attempts to start
beans unless `initializeStartables` has finished.
Fixes gh-38831