Support an alternative fat jar format that is more amenable to Docker
image layers.
The new format arranges files in the following structure:
BOOT-INF/
layers/
<layer-name #1>
/classes
/lib
<layer-name #2>
/classes
/lib
The `BOOT-INF/layers.idx` file provides the names of the layers and the
order in which they should be added (starting with the least changed).
The `JarLauncher` class can load layered jars in both fat and exploded
forms.
Closes gh-19767
Co-authored-by: Phillip Webb <pwebb@pivotal.io>
Update the `Repackager` class so that an additional `classpath.idx` file
is written into the jar that provides the original order of the
classpath. The `JarLauncher` class now uses this file when running as
an exploded archive to ensure that the classpath order is the same as
when running from the far jar.
Closes gh-9128
Co-authored-by: Phillip Webb <pwebb@pivotal.io>
This paves the way for publishing Gradle module metadata once the
problem caused by snapshot versions and our two-step publication
process has been addressed.
See gh-19609
This reverts commit b34a311d02 as,
having disabled the publishing of Gradle's module metadata (4f75ab5),
the changes are no longer needed.
See gh-19609
Previously, enforcedPlatform dependencies were using to pull in the
constraints defined in spring-boot-dependencies and
spring-boot-parent and applied them strictly so that the constrained
version had to be used. This worked as intended in Spring Boot's own
build but incorrectly enforced those same strict version requirements
on external consumers of Spring Boot's modules.
This commit reworks how Spring Boot defines its internal dependency
management so that platform dependencies are exposed to external
consumers while enforced platform dependencies are using internally.
See gh-19609
Previously, when RunProcess handled a SIGINT it would immediately
attempt to destroy the process that it had run. This created a race
condition between the SIGINT being handled by the child process
and RunProcess destroying the child. The exact behavior of destroy
is implementation dependent and it may result in forcible termination
of the process where shutdown hooks are not called. This is what
happens on Windows. The exit code in such a case is 1 which prevents
anything from waiting for the process to complete from detecting
that it ended as a result of a SIGINT, leaving it with no choice but
to report an error. This is what happens with mvn spring-boot:run
with a forked process on Windows and results in the build failing.
This commit updates RunProcess to allow the child process to handle
the SIGINT itself, waiting for up to five seconds for that to happen
before the process is then destroyed. Given this time, the child
process exits with 130 which RunMojo already handles correctly as
indicating that the process died due to SIGINT and the build completes
with success as a result.
Fixes gh-18936
Previously, Maven's default behaviour was relied up which resulted
in the artifact ID being appended to each URL as it was inherited.
This behaviour can only be disabled in Maven 3.6 and later, a version
that we cannot use due to an incompatibility with the Flatten Plugin.
This commit works around Maven's default behaviour by defining
properties for the SCM URL, connection, and developer connection and
then explicitly defining the settings in each pom using these
properties. The explicit definition of the properties in each pom
prevents them being inherited from the parent, thereby disabling the
unwanted appending of the artifact ID to the URL.
Fixes gh-18328
Apply checkstyle rule to ensure that private and package private
classes do not have unnecessary public methods. Test classes have
also been unified as much as possible to use default scoped
inner-classes.
Closes gh-7316
Since the move to JUnit 5, a number of tests were failing on Windows.
The majority were failing due to open file handles preventing the
clean up of the tests' temporary directory. This commit addresses
these failures by updating the tests to close JarFiles, InputStreams,
OutputStreams etc.
A change has also been made to CachingOperationInvokerTests to make
a flakey test more robust. Due to System.currentTimeMillis() being
less precise on Windows than it is on *nix platforms, the test could
fail as it would not sleep for long enough for the TTL period to have
expired.
With JDK 8 being the baseline and JDK 7 not being supported anymore we
can get rid of the workaround for a JDK 7 bug in
ProcessBuilder.inheritIO on Windows machines.
Closes gh-12337
Madhura Bhave
Phillip Webb
Andy Wilkinson
Stephane Nicoll
dreis2211
Wagner Macedo
Spring Operator
Robert Thornton
igor-suhorukov
Andreas Gebhardt
Oliver B. Fischer
Johnny Lim
Dmytro Nosan