remove PYNULL_CACHE (#767)

* remove PYNULL_CACHE

* restore PYNULL_CACHE but with a lock now

* back to removing PYNULL_CACHE but now skip the finalizer

* revert to some default settings in benchmark ci

* docstring clarification

---------

Co-authored-by: Christopher Rowley <github.com/cjdoris>

Author: cjdoris

.github/workflows/benchmark_pr.yml

@@ -8,9 +8,5 @@ jobs:
     steps:
       - uses: MilesCranmer/AirspeedVelocity.jl@action-v1
         with:
-          julia-version: "1"
           tune: "true"
-          # Post to "summary" tab of workflow run:
-          job-summary: "true"
-          # Run benchmark using PR's version of the script:
           bench-on: ${{ github.event.pull_request.head.sha }}

CHANGELOG.md

@@ -1,9 +1,10 @@
 # Changelog
 
 ## Unreleased
-
 * Add configuration via Preferences in addition to environment variables (e.g. `exe`
   rather than `JULIA_PYTHONCALL_EXE`.)
+* Internals: removed the cache of unused Python objects. This makes `pydel!` faster and
+  removes a race condition in free-threaded python.
 
 ## 0.9.32 (2026-05-14)
 * Added `juliacall.TypeValue.__numpy_dtype__` attribute to allow converting Julia types

src/Core/Py.jl

@@ -37,8 +37,6 @@ decref(x::Py) = Base.GC.@preserve x (decref(getptr(x)); x)
 
 Base.unsafe_convert(::Type{C.PyPtr}, x::Py) = getptr(x)
 
-const PYNULL_CACHE = Py[]
-
 """
     pynew([ptr])
 
@@ -50,12 +48,7 @@ points at, i.e. the new `Py` object owns a reference.
 Note that NULL Python objects are not safe in the sense that most API functions will probably
 crash your Julia session if you pass a NULL argument.
 """
-pynew() =
-    if isempty(PYNULL_CACHE)
-        Py(Val(:new), C.PyNULL)
-    else
-        pop!(PYNULL_CACHE)
-    end
+pynew() = Py(Val(:new), C.PyNULL)
 
 const PyNULL = pynew()
 
@@ -86,21 +79,17 @@ DANGER! Use this function ONLY IF the Julia object `x` could have been garbage-c
 anyway, i.e. was about to become unreachable. This means you MUST KNOW that no other part of
 the program has the Julia object `x`.
 
-This decrements the reference count, sets the pointer to NULL and appends `x` to a cache
-of unused objects (`PYNULL_CACHE`).
+This decrements the reference count and sets the pointer to NULL.
 
-This is an optimization to avoid excessive allocation and deallocation in Julia, which can
-be a significant source of slow-down in code which uses a lot of Python objects. It allows
-`pynew()` to pop an item from `PYNULL_CACHE` instead of allocating one, and avoids calling
-the relatively slow finalizer on `x`.
+Use this to eagerly free a Python object, rather than waiting for Julia's GC to finalize
+it at some indeterminate point in the future.
 """
 function pydel!(x::Py)
     ptr = getptr(x)
     if ptr != C.PyNULL
         C.Py_DecRef(ptr)
         setptr!(x, C.PyNULL)
     end
-    push!(PYNULL_CACHE, x)
     return
 end