We get more control over the handling and in particular the registration
of the endpoint this way. It was practically impossible to disable the
AgentServlet bean when in a parent context of the management server
because of lifecyce issues - you don't know that the user wants a
separate management server until too late.
This approach also makes it possible to test with spring-test MVC
support.
You can contribute additional HttpMessageConverters
by simply adding beans of that type in a Spring Boot
context. If a bean you add is of a type that would have been included
by default anyway (like MappingJackson2HttpMessageConverter for JSON
conversions) then it will replace the default value. A convenience
bean is provided of type MessageConverters (always available if you
use the default MVC configuration) which has some useful methods to
access the default and user-enhanced message converters (useful, for
example if you want to manually inject them into a custom
RestTemplate).
There are also some convenient configuration shortcuts for Jackson2.
The smallest change that might work is to just add beans of type
Module to your context. They will be registered with the default
ObjectMapper and then injected into the default message
converter. In addition, if your context contains any beans of type
ObjectMapper then all of the Module beans will be registered with
all of the mappers.
Previously it was @ConditionalOnMissingBean(CommandLineRunner.class)
which caued obvious problems when user wanted to add an unrelated
CLR.
Extended feature set so that a JobRegistry can also be used (with
spring.batch.job.name) and the whole idea can be switched off with
spring.boot.job.enabled.
ServerProperties formerly had an @OnMissingBeanCondition
that didn't restrict the hierarchy. It also asserts that
the current context (not including parents) contains such
a bean. This led to an inevitable failure when there was
an existing instance in the parent context.
Fixed by a) searching only the current context, b) not
adding a ServerProperties bean if the context is not a
web app.
Previously all EventListeners were eagerly instantiated
but that can cause problems because it happens quite early
in the lifecycle. Better to be explicit about the
supported types.
This leverages existing capabilities of teh JDK and the OS
to grab a port at random and not have it stolen by another
process. It's very hard to avoid that race condition in
pure Java code, so why bother?
User can set port<0 to disable autoStart of connectors (e.g.
to start a web application context but not have it listen on
any port). In that case the actual socket port will be set to
0 (and therefore if it ever starts up the local port will
be random).
The AutoConfigurationReportLoggingInitializer wasn't working in
non-GenericApplicationContext becasue teh BeanFatcory wasn't available
for registering its listener during initialization. Instead of
relying on that rather fragile state I decided to give any
ApplicationContextInitializer that was itself an ApplicationListener
an explicit callback with a ContextRefreshedEvent, and move that
interface up a level in the logging initializer. Works much better.
Using containsBean() involves looking in the parent bean factory
if there is one, and that would mean that the same report woykd be used
for multiple contexts, which wouldn't make sense.
