[Bug]: apache beam python SDK hangs and crashes with segmentation fault errors with orjson 3.9.4

[dankuchler]

What happened?

A bug introduced in orjson dependency (ijl/orjson#415) might cause Beam Python pipelines to crash with a segmentation fault or get stuck. Beam uses orjson in BigQuery IO, users of this IO might be affected.

Mitigation

Until Beam 2.51.0 is released, consider any of the following workarounds:

• Use apache-beam==2.49.0 or below. To avoid running into another known issue, consider apache-beam==2.46.0 .

• Install orjson==3.9.1 or below in the runtime environment. For example, you can use a --requirements_file pipeline option with a file that includes:

  orjson==3.9.1
  

  We recommend the version orjson==3.9.1 since it was previously tested with Beam 2.49.0 SDK.

  For more information, see: https://beam.apache.org/documentation/sdks/python-pipeline-dependencies/

• Install an [updated version of orjson dependency] (https://pypi.org/project/orjson/#history) once 3.9.4 has a threading issue ijl/orjson#415 is fixed.

Original report

In our latest deployment of our apache beam pipeline our dependency for orjson (dependency of the python apache beam SDK) was upgraded from 3.9.2 to 3.9.4.

The apache beam SDK has a dependency on orjson < 4.0 here:

https://github.com/apache/beam/blob/master/sdks/python/setup.py#L233

With this upgrade of orjson from 3.9.2 to 3.9.4 we are periodically seeing our apache beam SDK hang or the workers crash with segmentation fault errors that we believe is related to this issue in the orjson project:

ijl/orjson#415

When reverting from orjson 3.9.4 to 3.9.2 it seems that the issues are resolved.

The python apache beam SDK may want to limit orjson to 3.9.2 or below until orjson issue 415 is resolved.

Issue Priority

Priority: 2

Issue Components

• Component: Python SDK

[tvalentyn]

thanks for reporting. did you see some stacktraces around the segmentation fault by chance?

[tvalentyn]

Only Beam 2.50.0 has orjson==3.9.4 in Beam containers.

[dankuchler]

The segmentation fault messages that we are seeing in the system log seems to indicate that it is coming from orjson:

python[1925]: segfault at 7ff9f1ff4000 ip 00007ffa3c72df53 sp 00007ffa391bd000 error 6 in orjson.cpython-310-x86_64-linux-gnu.so[7ffa3c716000+2f000]
python[82009]: segfault at 70 ip 00007f91e601bb09 sp 00007f9189bfa080 error 4 in orjson.cpython-310-x86_64-linux-gnu.so[7f91e600f000+2f000]
python[1915]: segfault at 50 ip 00007f0431827b09 sp 00007f0407ffc1c0 error 4 in orjson.cpython-310-x86_64-linux-gnu.so[7f043181b000+2f000]

[dankuchler]

In the mitigation section - I believe 3.9.2 is fine - we have been running on orjson 3.9.2 and the issue in the orjson project mentioned people were mitigating by reverting to version 3.9.2