Prior to this commit, configuring a reactive Elasticsearch client would
auto-configure an Actuator Health check using a synchronous client, with
the default configuration properties (so tarting localhost:9200).
This would lead to false reports of unhealthy Elasticsearch clusters
when using reactive clients.
This commit reproduces the logic for MongoDB repositories: if a reactive
variant is available, it is selected for the health check
infrastructure.
See gh-21042
Previously, if TomcatMetricsBinder destroy() was called before it had
received an ApplicationStartedEvent an NPE would be thrown due to
TomcatMetrics being null. This NPE was then caught and logged at
warning level by the disposable bean adapter.
This prevents the NPE by checking that the TomcatMetrics instance is
null before calling close() on it.
See gh-22141
Previously, the thread dump endpoint's response could exceed
WebClient's in-memory buffer limt when there were a large number of
threads or the threads had large stacks.
This commit disables WebClient's in-memory buffer size limit so that
the test passing is not dependent on the number of active threads and
their stack sizes.
Closes gh-22101
Prior to Spring Data Redis version 2.2.8, the contents of the
Properties object returned from the
ReactiveRedisConnection.ServerCommands.info API were the same
for clustered and non-clustered Redis configurations, containing a set
of key/value pairs. This allowed ReactiveRedisHealthIndicator to get
a version property using a well-known key. Starting with Spring Data
Redis 2.2.8, the info property keys contain a host:port prefix in a
clustered Redis configuration. This prevented
ReactiveRedisHealthIndicator from getting the version property as
before and resulted in the health always being reported as DOWN.
This commit adjusts ReactiveRedisHealthIndicator to detect the
clustered configuration from Spring Data Redis and find the version
property for one of the reported cluster nodes.
Fixes gh-22061
Previously, Spring Boot's modules published Gradle Module Metadata
(GMM) the declared a platform dependency on spring-boot-dependencies.
This provided versions for each module's own dependencies but also had
they unwanted side-effect of pulling in spring-boot-dependencies
constraints which would influence the version of other dependencies
declared in the same configuration. This was undesirable as users
should be able to opt in to this level of dependency management, either
by using the dependency management plugin or by using Gradle's built-in
support via a platform dependency on spring-boot-dependencies.
This commit reworks how Spring Boot's build uses
spring-boot-dependencies and spring-boot-parent to provide its own
dependency management. Configurations that aren't seen by consumers are
configured to extend a dependencyManagement configuration that has an
enforced platform dependency on spring-boot-parent. This enforces
spring-boot-parent's version constraints on Spring Boot's build without
making them visible to consumers. To ensure that the versions that
Spring Boot has been built against are visible to consumers, the
Maven publication that produces pom files and GMM for the published
modules is configured to use the resolved versions from the module's
runtime classpath.
Fixes gh-21911
Update `EndpointDiscoverer` so that `@Endpoint` and `@EndpointExtension`
beans are created as late as possible.
Prior to this commit, endpoint beans and extension beans would be
created during the discovery phase which could cause early bean
initialization. The problem was especially nasty when using an embedded
servlet container since `ServletEndpointRegistrar` is loaded as the
container is initialized. This would trigger discovery and load all
endpoint beans, including the health endpoint, and all health indicator
beans.
Fixes gh-20714
Prior to this commit, there was a property server.error.include-details
that allowed configuration of the message and errors attributes in a
server error response.
This commit separates the control of the message and errors attributes
into two separate properties named server.error.include-message and
server.error.include-binding-errors. When the message attribute is
excluded from a servlet response, the value is changed from a
hard-coded text value to an empty value.
Fixes gh-20505
This commit updates HazelcastHealthIndicator and
HazelcastCacheMeterBinderProvider so that they work with
Hazelcast 4 while retaining compatibility with Hazelcast 3. Reflection
is used when necessary.
This commit also adds a smoke test that validates those features are
working when Hazelcast 4 is on the classpath.
Closes gh-21169
Prior to this commit, there was a cycle between `StatusAggregator` and
`SimpleStatusAggregator`, which caused a static initialization bug -
depending on which class (the implementation or its interface) was
loaded first.
This commit turns the static field of the `StatusAggregator` interface
into a static method to avoid this problem.
Fixes gh-21211
Prior to this commit, default error responses included the message
from a handled exception. When the exception was a BindException, the
error responses could also include an errors attribute containing the
details of the binding failure. These details could leak information
about the application.
This commit removes the exception message and binding errors detail
from error responses by default, and introduces a
`server.error.include-details` property that can be used to cause
these details to be included in the response.
Fixes gh-20505
Update the `HealthEndpointGroups` customization support to use a
post-processor rather than a mutable registry. Although this approach
is slightly less flexible, it removes a lot of complexity from the
`HealthEndpointGroups` code. Specifically, it allows us to drop the
`HealthEndpointGroupsRegistry` interface entirely.
The probe health groups are now added via the post-processor if they
aren't already defined. Unlike the previous implementation, users are
no longer able to customize status aggregation and http status code
mapping rules _unless_ they also re-define the health indicators that
are members of the group.
See gh-20962
Relocate probe auto-configuration from the `kubernetes` package to
`availability` since probes could also be used on other platforms.
The classes have also been renamed to named to `AvailabilityProbes...`
See gh-20962
Rename `LivenessProbeHealthIndicator` to `LivenessStateHealthIndicator`
and `ReadinessProbeHealthIndicator` to `ReadinessStateHealthIndicator`.
Also introduce a general purpose `AvailabilityStateHealthIndicator`
class.
See gh-20962
Create a general purpose `AvailabilityState` interface and refactor
the existing `LivenessState` and `ReadinessState` to use it. A single
`AvailabilityChangeEvent` is now used to carry all availability state
updates.
This commit also renames `ApplicationAvailabilityProvider` to
`ApplicationAvailabilityBean` and extracts an `ApplicationAvailability`
interface that other beans can inject. The helps to hide the event
listener method, which is really internal.
Finally the state enums have been renamed as follows:
- `LivenessState.LIVE` -> `LivenessState.CORRECT`
- `ReadinessState.READY` -> `ReadinessState.ACCEPTING_TRAFFIC`
- `ReadinessState.UNREADY` -> `ReadinessState.REFUSING_TRAFFIC`
See gh-20962
With its initial fix in gh-18444, the `WebClient` instrumentation would
record all CANCEL signals, including:
* when a `timeout` expires and the response has not been received
* when the client partially consumes the response body
Since the second use case is arguable intentional, this commit restricts
the instrumentation and thus avoids recording two events for a single
request in that case.
Closes gh-18444
Prior to this commit, cancelled client requests (for example as a result
of a `timeout()` reactor operator would not be recorded by Micrometer.
This commit instruments the cancelled signal for outgoing client
requests and assigns a status `CLIENT_ERROR`.
The cancellation can be intentional (triggering a timeout and falling
back on a faster alternative) or considered as an error. The intent
cannot be derived from the signal itself so we're considering it as a
client error.
Closes gh-18444
The system keyspace has a replication factor of 1 and is local to each
node; it is therefore recommended to query system.local with a
consistency level of ONE or LOCAL_ONE.
Stronger consistency levels may result in an Unavailable error, but this
does not mean that the node is down.
See gh-20709
Prior to this commit, `LivenessState` and `ReadinessState` were
immutable classes. This was done in order to have additional behavior
and information in those classes.
Because the current implementation doesn't need this, this commit turns
those classes into simple enums.
Additional state and information can be added to the
`*StateChangedEvent` classes.
See gh-19593
Andy Wilkinson
Brian Clozel
Aleksander Lech
im47cn
Scott Frederick
dreis2211
Stephane Nicoll
lltx
Phillip Webb
Jon Schneider
Phillip Webb
Brian Clozel
Johnny Lim
Andreas Born
Alexandre Dutra