Add a plugin that adds a handler to crash the JVM #109930
Conversation
thecoop
added
>test
|
Pinging @elastic/es-core-infra (Team:Core/Infra) |
...nal-modules/jvm-crash/src/javaRestTest/java/org/elasticsearch/test/jvm_crash/JvmCrashIT.java
Outdated
Show resolved
Hide resolved
There was a problem hiding this comment.
One question: do you know if the JVM is treating its internal code differently from the "user" code (i.e. the bytecode that gets interpreted/JITed?)
This plugin may be useful anyway to test how we respond to a JVM crash, but the usefulness can be limited if it ends up that native uncaught exceptions are treated differently based on where/when they originated
...nal-modules/jvm-crash/src/javaRestTest/java/org/elasticsearch/test/jvm_crash/JvmCrashIT.java
Outdated
Show resolved
Hide resolved
|
Other bugs may cause |
...modules/jvm-crash/src/javaRestTest/java/org/elasticsearch/test/jvm_crash/TeePrintStream.java
Show resolved
Hide resolved
...nal-modules/jvm-crash/src/javaRestTest/java/org/elasticsearch/test/jvm_crash/JvmCrashIT.java
Show resolved
Hide resolved
| private static void redirectStdout() { | ||
| if (originalOut == null) { | ||
| originalOut = System.out; | ||
| stdOutput = new ByteArrayOutputStream(); |
There was a problem hiding this comment.
Why can't we read the output from disk? The test nodes have an es.out file that should contain the crash info?
There was a problem hiding this comment.
The JVM crash is only output to standard out (JVM knows nothing about log4j), the test-cluster.log file doesn't have it
There was a problem hiding this comment.
es.out is not log4j, it's the stdout of the ES process.
There was a problem hiding this comment.
oof, I am mixing up testclusters and the new test framework. For things like gradlew run we have es.out, but it looks like the new cluster test framework starts the ES processes with inherited IO.
@mark-vieira is there a strong reason why the stdout/stderr is not captured via a file?
There was a problem hiding this comment.
We redirect stdout from the ES proccess to stdout in the test process so the crash should be captured in the test logs but perhaps not in a way that can be asserted on. If we need a general way to redirect logs to some other consumer we can certainly add an API for this use case. I think we've used the getNodeLog() API for these use cases in the past but that obviously doesn't work for JVM level stuff that doesn't go through log4j.
There was a problem hiding this comment.
And also it's very rare to have something that goes to stdout/stderr that does not go through log4j
|
@elasticmachine update branch |
A rest handler that purposefully causes a hard JVM crash for testing purposes