When the RefreshEndpoint makes a new Environment it creates a new
ApplicationContext with (hopefully) empty sources. If the user has
set spring.main.sources it won't be empty, so we should take care
to mask it off in the process.
Fixes gh-79
Repots DOWN if there are errors in RefreshScope or
ConfigurationProperties rebinding, N.B. RefreshScope instantiates
its beans lazily after a refresh by default, so you won't see an
error until the bean is used.
Fixes gh-85
The fact that PropertySourceBootstrapConfiguration was re-using the
"bootstrap" property source name from the listener was a problem. The
listener has to be able to manage its own temporary property source name.
Fixes gh-83
The bootstrap.properties need to stick with the default properties
during the process of initializing the application.properties
otherwise the ordering ends up wrong because of the addLast()
semantics in an Environment merge. To do this is a bit ugly with
the additional constraint that the default properties has to
remain a MapPropertySource (so that other processors can append
to it if needed). So we created a custom extension of
MapPropertySource that also carries all the bootstrap properties
during the phase where the application.properties are being
processed, but unpacks them as soon as possible afterwards
in an ApplicationContextInitializer, preserving the order, but
making the bootstrap.properties available effectively for the
whole of the startup and initializaion phase of the main context.
The biggest problem addressed here is one where an
EnvironmentPostProcessor (reasonably) adds entries to the
defaultProperties in the bootstrap context, but then that
property source is not merged with the parent, or is merged
too late (because it only happens when the application context
parent is set). The result would be that things that were activated
during bootstrap would be not be activated in the main context, or
would be activated too late (early enough for beans to bind to but
not for other listeners and post processors to get access to the
additional properties).
See https://github.com/spring-cloud/spring-cloud-sleuth/issues/126
We only need one special name for a property source. In the end it
gets used twice, but this simplifies the logic a bit and ensures that
there is never a PropertySource with that name and things like
spring.application.name=bootstrap in it.
This ensures that any fetaures that require a bean to be instantiated
(for instance @Scheduled) behave the same for @RefreshScope and normal
singletons.
Fixes gh-74
Otherwise multiple instances of a scoped target can be created
if several threads need one at the same time. This isn't always
going to be a problem, but it breaks the singleton semantics, and
might lead to surprising results and/or performance issues if bean
creation is expensive.
It's kind of a corner case, but not a very small corner: if a /refresh
leads to a change in a key that was overridden from its default value (e.g.
n a profile-specific config file) then the key is not identified as changed.
The change is still applied.
This commit fixes the key computation and adds a couple of tests.
Fixes gh-73
The problem is that we do want bootstrap properties to override
default properties in most cases, but if they are decrypted those
values have to be added with the highest possible precedence.
Fixes gh-54
EnvironmentChangeEvent should come before the RefreshScopeRefreshedEvent
so that all the @ConfigurationProperties are fully bound when the refresh
happens.
The main driver for this was a config server use case, where the server
wants to configure its own properties in the remote repo. It was working
up to a point, but the beans in bootstrap context were not refreshed
after the Environment was built so you had to manually refresh them to
get them to point to (e.g.) config repos configured remotely.
Fixes https://github.com/spring-cloud/spring-cloud-config/issues/139
The incoming (existing) environment needs to be used to initalize the
context that is created to grab the new environment, but we shouldn't
re-use the existing environment, rather just copy the property sources
and profiles.
The effect of not doing this is not often noticeable, but you sometimes
get a race condition between a context and its parent closing in different
threads without it.
