Java Debugging

Java Debugging

Find the root cause of a Java defect with a reproducible, hypothesis-driven loop, and attach a debugger safely when logs are not enough.

Safety rules

• Bind JDWP to loopback only (127.0.0.1). Never bind *:5005 on a reachable network; use an SSH tunnel or kubectl port-forward for remote hosts.

• Use suspend=n for running applications. suspend=y is acceptable only for short local test runs bound to loopback (waiting for the IDE to attach). Never leave a debug agent enabled on production; JDWP gives full code execution to whoever connects.

• Do not attach a debugger to production without explicit user instruction. Prefer reproducing locally or use read-only diagnostics (see the java-diagnostics skill).

• JAVA_TOOL_OPTIONS is picked up by every JVM started in that environment (build tools, forked test JVMs, child processes), causing Address already in use clashes. Prefer scoping the flag to the one process you want to debug.

Method

Reproduce deterministically, minimize, read the evidence (stack traces), test one falsifiable hypothesis with the least intrusive tool (log line or logpoint before breakpoints), then add a regression test and remove debug flags and temporary logging.

Version-dependent behavior (verified)

• Helpful NullPointerException messages (JEP 358) are off by default in JDK 14 (-XX:+ShowCodeDetailsInExceptionMessages) and on by default from JDK 15 (JEP 358 documents the message format). Source: https://openjdk.org/jeps/358.

• JDWP bare address=5005 binds all interfaces before JDK 9 and loopback from JDK 9 (see below).

• Gradle --debug-jvm passes a bare address, so on JDK 8 it listens on all interfaces.

Attach a debugger (JDWP)

Start the JVM:

java -agentlib:jdwp=transport=dt_socket,server=y,suspend=n,address=127.0.0.1:5005 -jar app.jar

Always write the explicit 127.0.0.1:5005 on every JDK version: a bare address=5005 listens on all interfaces before JDK 9 and only on loopback from JDK 9. On JDK 8 and older, the host:port form may not be honored, so also restrict access with a firewall or tunnel.

Attach with jdb:

jdb -attach 127.0.0.1:5005
# stop in com.example.OrderService.place | stop at com.example.OrderService:42
# cont | next | step | locals | print order | where | catch java.lang.NullPointerException

VS Code (Extension Pack for Java)

.vscode/launch.json:

{
  "version": "0.2.0",
  "configurations": [
    {
      "type": "java",
      "name": "Attach to JVM",
      "request": "attach",
      "hostName": "127.0.0.1",
      "port": 5005
    }
  ]
}

IntelliJ IDEA

Run | Edit Configurations | + | Remote JVM Debug; host 127.0.0.1, port 5005, debugger mode Attach. The dialog shows the matching JVM argument string.

Debug tests and frameworks

Surefire/Failsafe debug mode and --debug-jvm suspend until a debugger attaches, which is fine for a local run. Surefire and Failsafe accept a custom agent string instead of the bare flag, e.g. -Dmaven.surefire.debug="-agentlib:jdwp=transport=dt_socket,server=y,suspend=y,address=127.0.0.1:5005". Verified listeners on 127.0.0.1 (via lsof -nP -iTCP:5005 -sTCP:LISTEN) for Gradle 9.7 --debug-jvm, Maven Surefire 3.5 -Dmaven.surefire.debug and Quarkus 3.40 dev mode on JDK 26; Gradle 7.6.4, 8.0.2 (JDK 17), 8.5 (JDK 21) and 8.14.3 also listened on 127.0.0.1. Gradle passes a bare address=5005 (no host), so the bind address comes from the JDK: loopback on JDK 9+, but all interfaces (0.0.0.0:5005, checked in an eclipse-temurin:8-jdk container with Gradle 7.6.4 and 8.14.3) on JDK 8. Older Gradle/plugin versions were not tested; run the same lsof check and use an explicit loopback agent on shared networks or JDK 8.

Explicit Gradle configuration (Gradle 7.6+; debugOptions.host was added in 7.6; Kotlin DSL):

tasks.withType<Test> {
    debugOptions { enabled.set(true); host.set("127.0.0.1"); port.set(5005); server.set(true); suspend.set(true) }
}

// BootRun extends JavaExec, not Test; configure it separately when the Boot plugin is applied.
tasks.named<JavaExec>("bootRun") {
    debugOptions { enabled.set(true); host.set("127.0.0.1"); port.set(5005); server.set(true); suspend.set(false) }
}

Then run gradle test --tests '*OrderServiceTest' or gradle bootRun, without --debug-jvm (which enables suspension). Configure only the task you are debugging; do not run both listeners on port 5005 at once. For Gradle before 7.6, use jvmArgs("-agentlib:jdwp=transport=dt_socket,server=y,suspend=n,address=127.0.0.1:5005") on the target task instead (suspend=y for local tests), without also enabling debugOptions or --debug-jvm. On JDK 8 and older, retain the firewall/tunnel precaution above.

Containers

Publish the port to loopback only and let the JVM listen inside the container:

docker run --rm -p 127.0.0.1:5005:5005 \
  -e JAVA_TOOL_OPTIONS="-agentlib:jdwp=transport=dt_socket,server=y,suspend=n,address=*:5005" my-app

Inside the container the agent must bind * for the port mapping to work, so JDWP (full code execution) is reachable by every other container on the same Docker network; the 127.0.0.1: mapping only blocks access from other hosts. Use this only on a developer machine where no untrusted containers share the network: run it on a dedicated user-defined network (not the shared default bridge) containing only this container, or keep the agent on 127.0.0.1 inside the container and reach it with docker exec/kubectl port-forward style forwarding instead of publishing it. For Kubernetes use kubectl port-forward pod/<name> 5005:5005 against a dev/staging pod only.

Common pitfalls

• Hot code replace only swaps method bodies; adding/removing methods or fields or changing signatures needs a restart (or Quarkus dev mode / Spring DevTools).

• Missing debug info: code compiled without -g shows unavailable variables (Maven and Gradle include it by default). In lambdas, step into the lambda body.

• Heisenbugs/races: pausing changes timing. Prefer logpoints, thread dumps, or JFR; suspend only the current thread, not all threads.

• Async and reactive code: stack traces show scheduler frames, not the caller. Use IDE async stack traces, correlation IDs, or Reactor Hooks.onOperatorDebug() in dev only.

• Virtual threads (JDK 21+): there can be thousands; use jcmd <pid> Thread.dump_to_file -format=json file.json for an overview.

• Timeouts: long pauses can trip HTTP, DB, or heartbeat timeouts; raise them locally.

• toString() side effects: debugger evaluation calls toString(), which may trigger lazy loading or mutate state.

Optional code-navigation aid

If the Copilot CLI lsp tool is configured (see the lsp-setup skill), go-to-definition, find-references, and call hierarchy help trace the execution path. It is optional; grep and reading code work as well.

References

• Reading stack traces

• For runtime symptoms (CPU, memory, deadlocks, slow startup) use the java-diagnostics skill.

Official tool docs (prefer them over this skill for the JDK in use; change 21 to your version): jcmd https://docs.oracle.com/en/java/javase/21/docs/specs/man/jcmd.html, JDK troubleshooting guide https://docs.oracle.com/en/java/javase/21/troubleshoot/.