State of Python 3.13 performance: Free-threading

Original Article ↗ Hacker News Discussion ↗

Overall sentiment

  • Many welcome CPython performance work and free-threading as overdue, given Python’s ubiquity.
  • Others question whether chasing speed in CPython is worth the ecosystem churn, given Python’s semantic limits and reliance on C/accelerated libraries.

Python performance limits

  • One camp argues Python can “be fast enough” (e.g., via PyPy, JITs, optimized C extensions), and that for many workloads this erases the need to rewrite prototypes.
  • Skeptics say even PyPy/JS VMs remain far slower than low-level languages and often sacrifice ecosystem compatibility.
  • Several note Python’s highly dynamic semantics (mutable classes, arbitrary-precision ints, etc.) make “C-like” performance for general code unrealistic; fast paths will always rely on restricted subsets or offloaded native code.

Free-threading / GIL removal

  • Removing the GIL is seen as crucial for CPUs with many cores and for workloads where Python “glue” becomes the parallelism bottleneck (data loaders, preprocessing pipelines, outer loops around NumPy/PyTorch).
  • Others argue most existing Python is single-threaded and will see little benefit; extra complexity and thread-safety bugs may outweigh gains.
  • There is debate over how much code actually “relied” on the GIL for implicit safety versus being accidentally safe.

Backwards compatibility and churn

  • Multiple comments compare to the 2→3 transition and note long-lived ABI instability and standard-library removals across 3.x.
  • Some feel Python 3.x breaks “too much” for a language with so much FFI and legacy code; others stress most removals followed long deprecation periods.
  • Threading and ABI changes are seen as another wave of breakage, but some argue that since compatibility is already fragile, performance wins might justify it.

Dependency management & packaging

  • Recurrent pain point: getting old projects running due to unpinned or incompatible dependencies; conda/npm/CRAN analogies raised.
  • Advocates recommend pinning (or locking) full dependency sets; critics note this complicates libraries and long-term maintenance.
  • Tools like Poetry and uv are cited as better defaults than raw pip, but pip’s behavior is still the de facto experience for many.

Implementation details

  • Core enabler for free-threading is thread-safe, “biased” reference counting plus many fine-grained locks.
  • Cyclic GC still stops all threads in free-threaded mode; precise performance trade-offs and crash bugs in early 3.13/3.14t builds are acknowledged.