[Bug] Concurrent access to a shared PythonInterpreter can crash the JVM in async cross-language calls

[wenjin272]

Search before asking

• I searched in the issues and found nothing similar.

Description

The embedded pemja.core.PythonInterpreter owned by an operator is shared by cross-language action and resource paths. When multiple threads concurrently enter the same interpreter through Pemja APIs such as invoke, the JVM can terminate with a native SIGSEGV instead of returning a managed Java or Python exception.

This is broader than the Mem0 path where the problem was first observed. Any asynchronous cross-language path that can concurrently access the same interpreter is potentially affected. The participating threads may originate from:

• Flink's Java async executor;
• the mailbox or lifecycle thread;
• Python-created workers that call into Java and then re-enter the same interpreter through a Java adapter; or
• another JVM thread carrying a cross-language callback.

This does not mean that every asynchronous operation is unsafe. A Python worker that only runs Python code, a callback that only executes Java code without touching the interpreter, sequential cross-language calls, and calls using different interpreter instances are outside the demonstrated failure condition.

The failure was first observed in the cross-language Mem0 long-term-memory E2E test in this CI job:

https://github.com/apache/flink-agents/actions/runs/33615851952/job/100201365522

The original failing path was:

Java durable async worker
  -> Java MemorySet.add
  -> Java Mem0LongTermMemory wrapper
  -> shared Pemja PythonInterpreter.invoke
  -> Python Mem0LongTermMemory.add
  -> Mem0 chat / embedding / vector-store processing

A sanitized excerpt from a local reproduction is:

# A fatal error has been detected by the Java Runtime Environment:
# SIGSEGV (0xb)
# JRE version: Java(TM) SE Runtime Environment (21.0.8+12)
# Java VM: Java HotSpot(TM) 64-Bit Server VM (bsd-aarch64)

Current thread: JavaThread "flink-agents-java-async-pool-5-thread-2" [_thread_in_native]

C  [Python]  _PyFrame_Clear
C  [Python]  _PyEval_EvalFrameDefault
C  [pemja_core.cpython-311-darwin.so]  JcpPyObject_Call
C  [pemja_core.cpython-311-darwin.so]  Java_pemja_core_PythonInterpreter_invoke
j  pemja.core.PythonInterpreter.invoke(...)
j  org.apache.flink.agents.runtime.python.utils.PythonResourceAdapterImpl.callMethod(...)

Root-cause isolation

The failure was isolated from Mem0 with a ChatModel cross-language E2E test:

1. Configure AgentExecutionOptions.NUM_ASYNC_THREADS to 2.
2. Send two events with different keys so that two slow chat requests overlap.
3. Use a Java agent calling a Python ChatModel setup through the shared interpreter.

Both connection variants are affected:

• Java -> Python setup -> Java connection; and
• Java -> Python setup -> Python connection.

In the Python-connection variant, the JVM crashed before the Python connection wrapper's chat() entry log was emitted. The fatal stack was already inside PythonResourceAdapterImpl.callMethod -> PythonInterpreter.invoke, which excludes the connection implementation, Mem0's internal executor, and Elasticsearch as required causes.

Expected behavior: asynchronous cross-language execution completes safely, or an unsupported concurrent access pattern fails with a managed exception without terminating the JVM.

Actual behavior: concurrent access to the shared interpreter can terminate the entire JVM in CPython evaluation through Pemja.

How to reproduce

Isolated ChatModel reproduction

1. Build Flink Agents and install its Python dependencies.
2. Start Ollama with a chat model such as qwen3:1.7b.
3. Configure the cross-language ChatModel test with two async threads and two different input keys.
4. Use a Python ChatModel setup backed by either a Java or Python connection.
5. Make the two requests slow enough to overlap and run ChatModelCrossLanguageTest.

When the requests overlap, the Surefire JVM may abort before both requests complete and produce an hs_err_pid*.log.

Original Mem0 reproduction

1. Start Ollama with the nomic-embed-text model.
2. Start Elasticsearch and export ES_HOST=http://localhost:9200.
3. Export a valid ACTION_API_KEY and, if needed, ACTION_BASE_URL for the OpenAI-compatible chat model.
4. Run:

mvn -pl e2e-test/flink-agents-end-to-end-tests-resource-cross-language \
  -Dtest=Mem0LongTermMemoryTest test

The test sends interleaved facts for two keys. Mem0LongTermMemoryAgent.addItems() invokes memorySet.add() using durableExecuteAsync(). One of the overlapping additions may abort the Maven Surefire JVM and produce an hs_err_pid*.log.

Version and environment

• Flink Agents: 0.4-SNAPSHOT
• Flink: 2.3.0
• Pemja: 0.5.7
• CI: ubuntu-latest, Temurin JDK 21, Python 3.12
• Local reproduction: macOS aarch64, Oracle JDK 21.0.8, Python 3.11

Are you willing to submit a PR?

• I'm willing to submit a PR!