[
https://issues.apache.org/jira/browse/SPARK-61490?page=com.atlassian.jira.plugin.system.issuetabpanels:all-tabpanel
]
Hyukjin Kwon reassigned SPARK-61490:
------------------------------------
Assignee: Hyukjin Kwon
> Release the instance dict of the Python client's PyO3 classes when they are
> freed, which PyO3 0.28 keeps
> --------------------------------------------------------------------------------------------------------
>
> Key: SPARK-61490
> URL: https://issues.apache.org/jira/browse/SPARK-61490
> Project: Spark
> Issue Type: Bug
> Components: Connect
> Affects Versions: connect-rust-4.3.0
> Reporter: Hyukjin Kwon
> Assignee: Hyukjin Kwon
> Priority: Major
> Labels: pull-request-available
> Fix For: connect-rust-4.3.0
>
>
> PyO3 0.28 never releases the __dict__ of a #[pyclass(dict)] instance:
> PyClassObjectContents::dealloc calls clear_dict, which is PyDict_Clear, and
> does not decref the dict (PyO3 0.29 adds release_dict). Every instance of the
> Python client's 28 dict classes (Column, DataFrame, SparkSession, DataType
> and its Rust subclasses, GroupedData, readers and writers, Observation, ...)
> whose __dict__ was created (an attribute set, vars(), dir(), monkeypatching,
> a __reduce__ that reads it) keeps an empty dict, 64 bytes, after it is freed.
> And since PyDict_Clear clears outside CPython's trashcan, freeing a long
> chain of instances linked through their attributes (100k Columns) overflows
> the stack and crashes the process.
> PyO3 cannot be upgraded here: arrow-pyarrow 58.3.0 pins pyo3 0.28.
> Fix: wrap the tp_dealloc of each dict class, and of the Rust subclasses of
> DataType, so that the instance dict is taken out of its slot before PyO3's
> dealloc and released after it, only where the type object holds the
> tp_dealloc that PyType_GetSlot reports. Note: on CPython 3.11 and 3.12,
> object.__new__ also creates a dict for every instance of a type with a dict
> offset, which PyO3 0.28 overwrites; that creation leak needs PyO3 0.29.
--
This message was sent by Atlassian Jira
(v10.3.26#10030026)
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]