zentol commented on a change in pull request #11303: [FLINK-16245] Decoupling
user classloader from context classloader.
URL: https://github.com/apache/flink/pull/11303#discussion_r393526519
##########
File path:
flink-runtime/src/main/java/org/apache/flink/runtime/execution/librarycache/FlinkUserCodeClassLoaders.java
##########
@@ -82,4 +89,61 @@ public static ResolveOrder fromString(String resolveOrder) {
super(urls, parent);
}
}
+
+ /**
+ * Ensures that holding a reference on the context class loader
outliving the scope of user code does not prevent
+ * the user classloader to be garbage collected (FLINK-16245).
+ *
+ * <p>This classloader delegates to the actual user classloader. Upon
{@link #close()}, the delegate is nulled
+ * and can be garbage collected. Additional class resolution will be
resolved solely through the bootstrap
+ * classloader and most likely result in ClassNotFound exceptions.
+ */
+ private static class SafetyNetWrapperClassLoader extends URLClassLoader
+ implements Closeable {
+ private static final Logger LOG =
LoggerFactory.getLogger(SafetyNetWrapperClassLoader.class);
+
+ private FlinkUserCodeClassLoader inner;
+
+ SafetyNetWrapperClassLoader(FlinkUserCodeClassLoader inner) {
+ super(new URL[0], null);
+ this.inner = inner;
+ }
+
+ @Override
+ public void close() {
+ if (inner != null) {
+ try {
+ inner.close();
+ } catch (IOException e) {
+ LOG.warn("Could not close user
classloader", e);
+ }
+ }
+ inner = null;
+ }
+
+ @Override
+ protected Class<?> loadClass(String name, boolean resolve)
throws ClassNotFoundException {
+ if (inner == null) {
+ return super.loadClass(name, resolve);
Review comment:
> any job with leaked context classloader [...] fail eventually
Given that the CL is only closed when all tasks on the TE have stopped, how
could it affect a _job_? Leftover threads, sure, but what happens to those
shouldn't affect job execution, and failing as early as possible might actually
be a benefit.
> [throwing a CNFE] may trigger some unwanted fallback logic in user
libraries
This can happen in any case; for all we now the library could be catching
throwables.
I'd fail with an IllegalStateException; I think when users see a CNFE they
first consider it as a class-loading issue in Flink (regardless of the error
message).
----------------------------------------------------------------
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.
For queries about this service, please contact Infrastructure at:
[email protected]
With regards,
Apache Git Services
