Talks - Vinícius Gubiani Ferreira: PEP 683: Immortal Objects - A new approach for memory managing

PEP 683 introduces “immortal objects” - objects that live throughout the entire program execution with an unchanging reference count

Immortal objects bypass Python’s garbage collector, reference counting, and Global Interpreter Lock (GIL), improving memory management and performance

Common immortal objects include None, True, False, empty strings, and small integers (-5 to 256)

Instagram created PEP 683 to address:

• CPU cache invalidation issues
• Data race conditions
• Copy-on-write (COW) memory problems
• Memory consumption in pre-forking web servers

Immortal objects are implemented by setting a very high reference count value in the object’s C structure

The change resulted in decreased memory consumption and increased shared memory efficiency as request numbers grew

Currently, creating custom immortal objects requires C-level modifications - no Python-level API exists yet

PEP 683 works alongside PEP 684 (making GIL optional) to improve Python’s parallel processing capabilities

Developers need to be careful with immortal objects as they effectively create memory leaks that persist until program termination

The implementation faced challenges with:

• Platform compatibility across different OS/architectures
• Backward compatibility
• Performance optimization
• Accidental immortalization of objects