014-code commented on PR #30046: URL: https://github.com/apache/flink/pull/30046#issuecomment-7037369615
Additional reproduction, design, and validation details: ### Reproduction and root cause I compiled the current checkout's four classloader source files and exercised them with a real temporary JAR containing a resource. For each loading order, the reproducer opens the resource with `getResourceAsStream`, reads it, closes the stream, and then closes the classloader. It checks the cached `JarFile` through its public API and separately attempts to delete the JAR on Windows. On Java 17, with each loading order repeated three times: | Configuration | JAR closed after loader close | Windows deletion | | --- | --- | --- | | Parent-first, no safety-net wrapper | 3/3 | 3/3 succeeded | | Child-first, no safety-net wrapper | 3/3 | 3/3 succeeded | | Parent-first, safety-net wrapper | 0/3 | 0/3 succeeded | | Child-first, safety-net wrapper | 0/3 | 0/3 succeeded | The wrapper delegates `getResource` to the inner loader but inherits `URLClassLoader.getResourceAsStream`. That inherited method records the JAR it opens on the wrapper. The wrapper's `close()` only closes the inner loader, leaving the wrapper-owned cached JAR open even after the caller closes the resource stream. ### Why delegate the stream lookup The fix forwards `getResourceAsStream` through `ensureInner()`, consistent with the wrapper's existing resource and class-loading methods. The inner loader now owns the opened resource and releases it when the wrapper closes it. This also retains the existing closed-loader guard and parent-first/child-first resource lookup behavior. Calling `super.close()` on the wrapper is another possible direction, but delegation avoids creating a second resource owner in a class intended to forward operations to the inner loader. The patch does not change `close()` or add a separate cleanup mechanism. ### Regression test and red-green check The parameterized test uses a private copy of the existing test JAR for each loading order, reads and closes a real class resource, retrieves the cached JAR with `JarURLConnection`, closes the loader, and asserts that `JarFile.size()` rejects access to the closed JAR. Try-with-resources releases the JAR even when the assertion fails. The assertion does not rely on Windows deletion semantics, sleeps, reflective access to JDK internals, or mocked classloaders. Before applying the production change, both cases failed because `JarFile.size()` remained callable. After the fix, both passed. The initial red-green run used JUnit Platform Console with the current classloader and test-helper sources compiled directly; auxiliary Flink dependencies were published 2.2.1 artifacts. Separately, the Maven validation built the current reactor modules and compiled all production and test sources in `flink-core`. Its selected test classes passed 13/13 tests with no failures, errors, or skips. Checkstyle and Spotless also passed. ### Remaining validation This stays a draft because full `clean verify` has not passed. The initial Maven run was stopped while Apache Rat spent several minutes in license-matching regex processing, confirmed by a thread dump. A subsequent run compiled successfully but default Surefire forks printed Windows "Access is denied" without starting the tests. The successful targeted Maven run therefore used `-Drat.skip=true` and `-Dflink.forkCountUnitTest=0`, as documented in the PR body. The full module test suite, default forked execution, full license scan, and Linux/macOS execution remain unverified. Same-path JAR replacement was not tested separately. These limits are not included in the passing-test claim. -- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. To unsubscribe, e-mail: [email protected] For queries about this service, please contact Infrastructure at: [email protected]
