Version 2.1.0 segfaults when subscribing to non-existent topic

[ffissore]

Description

Since confluent-kafka 2.1.0, subscribing to a non-existent topic causes python to segfault.

How to reproduce

Using the files in this gist, run docker compose, then run script test.py

1. at first, with no modifications
2. then, with the admin client part commented out: in short, don't create the topic

The first run will be successful. The second run will log Segmentation fault (core dumped)

With confluent-kafka 2.0.2, poll returns a message with value b'Subscribed topic not available: test-cf67de5e-e79e-48f7-9300-763fb5b8bc05: Broker: Unknown topic or partition'

Checklist

Please provide the following information:

• confluent-kafka-python and librdkafka version (confluent_kafka.version() and confluent_kafka.libversion()): ('2.1.0', 33619968) ('2.1.0', 33620223)
• Apache Kafka broker version: 2.1.1 (confluent 5.1.2)
• Client configuration: {...}: in the gist
• Operating system: ubuntu 20.04
• Provide client logs (with 'debug': '..' as necessary)
• Provide broker log excerpts
• Critical issue

[ffissore]

Using gdb I got this far in debugging (not much actually)

Thread 1 "python" received signal SIGSEGV, Segmentation fault.
0x00007ffff62f7115 in rd_kafka_message_leader_epoch () from /home/federico/.pyenv/versions/3.11.2-debug/envs/test-env/lib/python3.11/site-packages/confluent_kafka/../confluent_kafka.libs/librdkafka-e24e6ccd.so.1

[emasab]

Hello Federico, thanks. I'm aware of the issue. Have done a fix in the .NET client but I'm going to do it for librdkafka too.

[Richetto]

Same issue here, upgrading to 2.1.0 from 1.6.0. Any update on the fix?
Thank you

[emasab]

Yes, we've merged the fix here and are planning a maintenance release soon.

[lpsinger]

This is still broken in confluent-kafka-python 2.1.1.

[emasab]

@lpsinger are you sure? I've just tried with a non existing topic and it gives

Traceback (most recent call last):
  File "consumer.py", line 98, in <module>
    raise KafkaException(msg.error())
cimpl.KafkaException: KafkaError{code=UNKNOWN_TOPIC_OR_PART,val=3,str="Subscribed topic not available: this_topic_doesnt_exist: Broker: Unknown topic or partition"}

[lpsinger]

We have a lightweight client library that wraps confluent-kaka-python and adds the configuration presets for our Kafka cluster: https://github.com/nasa-gcn/gcn-kafka-python

It's segfaulting with unknown topics. So it might be something with confluent-kafka-python 2.1.1 plus unknown topics plus OpenID Connect.

[lpsinger]

To test, go to https://gcn.nasa.gov, click "Start streaming GCN Notices", and follow the instructions.

[emasab]

@lpsinger Thanks for that, I've installed and reproduced. The error happens in a different place than the initial one for a fix we did to the consume batch in 2.1.0. It happens with the Consumer.consume method but not with Consumer.poll, with non-existent topics.

[ffissore]

I confirm that our code is working with version 2.1.1: our tests are green
As I'm not sure if you prefer another issue to track the bug with Consumer.consume, I'll leave closing this issue to you.

[milibopp]

When I run the test suite on version 2.1.1, one of the tests seems to segfault, see gist. The same happens, when I upgrade both librdkafka and this package to 2.2.0.

It seems like the same issue.

[emasab]

@milibopp It doesn't seem the same issue, in that test test_oauth_cb_principal_sasl_extensions it's not subscribing to any topics. I couldn't reproduce it by running that test, could your provide some hint to reproduce it?

[ksajan]

I am facing the similar issue in the current package version v2.2.0 in python 3.10.

[Vikash08Mishra]

I still see similar issue even in version >= 2.0.2 when using SSL, more details in ticket: #1690