Repository navigation
profiling.sampling: --blocking sample-read error rate ~33% on AMD EPYC 9V45 (Windows), ~0.5% on other CPUs #158552
Description
Activity
cc @pablogsal but I'm not sure we can do something on our end. It could be a mix of Windows and that kind of CPU.
There may be a race in our Windows blocking path. We call
NtSuspendProcessand immediately start reading memory, without checking that the target threads have finished suspending.Could you run the diagnostic script below (save it as
diagnose.py) with the same interpreter and target on a failing EPYC 9V45 runner, and ideally one unaffected runner?python diagnose.py --target target.py python diagnose.py --target target.py --no-cachePlease paste the full output, CPU model, and exact CPython build or commit. The script records the exception chains and compares normal suspension with a delay and with
GetThreadContext, which should synchronize suspension.In particular, we need to establish whether the failing blocking runs produce the same
Failed to parse initial frame in chainerror you observed locally without blocking.Your original target is fine for this check; no need to reduce it yet.
diagnose.py
"""Compare sampling failures on a single-threaded target, using Python 3.15. Run: python diagnose.py [--target target.py] [--seconds 3] [--no-cache] Windows x64 also tests GetThreadContext after pause_threads(). """ import argparse from collections import Counter import contextlib import ctypes import os import platform import runpy import subprocess import sys import threading import time def leaf(x): return x + 1 def calls(): x = 0 while True: x = leaf(leaf(leaf(x))) def worker(target): print(threading.get_native_id(), flush=True) sys.stdin.readline() with open(os.devnull, "w") as output, contextlib.redirect_stdout(output): if target: sys.argv = [target] sys.path.insert(0, os.path.dirname(os.path.abspath(target))) runpy.run_path(target, run_name="__main__") else: calls() @contextlib.contextmanager def windows_context(tid): # Only the main thread of the explicitly single-threaded target is checked. # CONTEXT layout: https://learn.microsoft.com/windows/win32/api/winnt/ns-winnt-context from ctypes import wintypes kernel32 = ctypes.WinDLL("kernel32", use_last_error=True) kernel32.OpenThread.argtypes = (wintypes.DWORD, wintypes.BOOL, wintypes.DWORD) kernel32.OpenThread.restype = wintypes.HANDLE kernel32.GetThreadContext.argtypes = (wintypes.HANDLE, ctypes.c_void_p) kernel32.GetThreadContext.restype = wintypes.BOOL kernel32.CloseHandle.argtypes = (wintypes.HANDLE,) kernel32.CloseHandle.restype = wintypes.BOOL handle = kernel32.OpenThread(0x0008, False, tid) # THREAD_GET_CONTEXT if not handle: raise ctypes.WinError(ctypes.get_last_error()) try: # AMD64 CONTEXT is 1232 bytes, aligned to 16 bytes. ContextFlags follows # six 64-bit home slots. Only control registers are requested. storage = ctypes.create_string_buffer(1232 + 15) address = (ctypes.addressof(storage) + 15) & ~15 ctypes.c_uint32.from_address(address + 48).value = 0x00100001 def synchronize(): if not kernel32.GetThreadContext(handle, address): raise ctypes.WinError(ctypes.get_last_error()) yield synchronize finally: kernel32.CloseHandle(handle) def error_chain(exc): parts = [] while exc is not None: parts.append(f"{type(exc).__name__}: {exc}") exc = exc.__cause__ return " <- ".join(parts) def measure(args, mode): from _remote_debugging import RemoteUnwinder command = [sys.executable, os.path.abspath(__file__), "--worker"] if args.target: command += ["--target", args.target] errors = Counter() retries = Counter() count = 0 with subprocess.Popen(command, stdin=subprocess.PIPE, stdout=subprocess.PIPE, text=True) as child: try: tid = int(child.stdout.readline()) unwinder = RemoteUnwinder(child.pid, native=True, opcodes=True, gc=True, cache_frames=not args.no_cache, debug=True) manager = windows_context(tid) if mode == "context" else contextlib.nullcontext() with manager as synchronize: child.stdin.write("go\n") child.stdin.flush() time.sleep(0.1) start = deadline = time.perf_counter() while time.perf_counter() - start < args.seconds and child.poll() is None: paused = False try: if mode != "live": unwinder.pause_threads() paused = True if mode == "delay": time.sleep(0.001) elif mode == "context": synchronize() try: unwinder.get_stack_trace() except (RuntimeError, OSError, UnicodeError, MemoryError) as exc: errors[error_chain(exc)] += 1 # Check a few failures again without resuming the target. if paused and sum(retries.values()) < 10: time.sleep(0.001) try: unwinder.get_stack_trace() except (RuntimeError, OSError, UnicodeError, MemoryError) as again: retries[error_chain(again)] += 1 else: retries["retry succeeded while still paused"] += 1 finally: if paused: unwinder.resume_threads() count += 1 deadline += 0.001 remaining = deadline - time.perf_counter() if remaining > 0: time.sleep(remaining) elapsed = time.perf_counter() - start finally: if child.poll() is None: child.terminate() child.wait() failed = sum(errors.values()) rate = 100 * failed / count if count else 0 print(f"{mode}: {failed}/{count} failed ({rate:.2f}%), {elapsed:.3f}s", flush=True) for message, frequency in errors.most_common(5): print(f" {frequency}: {message}", flush=True) for message, frequency in retries.most_common(5): print(f" same-pause retry {frequency}: {message}", flush=True) def main(): parser = argparse.ArgumentParser(description=__doc__) parser.add_argument("--target", help="Optional single-threaded Python script") parser.add_argument("--seconds", type=float, default=3) parser.add_argument("--no-cache", action="store_true") parser.add_argument("--worker", action="store_true", help=argparse.SUPPRESS) args = parser.parse_args() if args.worker: worker(args.target) return print(sys.version) print(platform.platform(), platform.processor(), f"cache={not args.no_cache}", flush=True) modes = ["live", "pause", "delay"] if sys.platform == "win32" and platform.machine().lower() in ("amd64", "x86_64") and ctypes.sizeof(ctypes.c_void_p) == 8: modes.append("context") for mode in modes: measure(args, mode) if __name__ == "__main__": main()
Thanks, @pablogsal. I ran
diagnose.pyon 12 GitHub-hosted Windows runners (imagewindows-2025-vs2026), three of which landed on an AMD EPYC 9V45. Build on every runner:3.15.0rc1 (tags/v3.15.0rc1:37e98da, Aug 4 2026, 11:37:18) [MSC v.1951 64 bit (AMD64)].Your race hypothesis matches the data. On the EPYC 9V45 runners,
pausefails about as often aslive; a 1 ms delay mostly fixes it, andGetThreadContextbrings it to 0. On every other CPU,pauseis already 0.Failure rates with the built-in single-threaded loop (3 s each):
Runner CPU live pause delay context live (no-cache) pause (no-cache) delay (no-cache) context (no-cache) 0 Intel(R) Xeon(R) 6973P-C 16.03% 0.00% 0.00% 0.00% 11.57% 0.00% 0.00% 0.00% 1 AMD EPYC 7763 64-Core Processor 12.83% 0.00% 0.00% 0.00% 12.90% 0.00% 0.00% 0.00% 2 AMD EPYC 7763 64-Core Processor 13.93% 0.00% 0.00% 0.00% 11.97% 0.00% 0.00% 0.00% 3 AMD EPYC 9V74 80-Core Processor 12.40% 0.00% 0.00% 0.00% 12.00% 0.00% 0.00% 0.00% 4 AMD EPYC 9V74 80-Core Processor 12.53% 0.00% 0.00% 0.00% 12.47% 0.00% 0.00% 0.00% 5 AMD EPYC 9V74 80-Core Processor 13.07% 0.00% 0.00% 0.00% 13.53% 0.00% 0.00% 0.00% 6 AMD EPYC 9V74 80-Core Processor 12.90% 0.00% 0.00% 0.00% 12.80% 0.00% 0.00% 0.00% 7 AMD EPYC 9V45 96-Core Processor 11.10% 10.93% 0.56% 0.00% 11.27% 11.30% 0.44% 0.00% 8 AMD EPYC 9V74 80-Core Processor 12.27% 0.00% 0.00% 0.00% 15.53% 0.00% 0.00% 0.00% 9 AMD EPYC 7763 64-Core Processor 12.67% 0.00% 0.00% 0.00% 12.43% 0.00% 0.00% 0.00% 10 AMD EPYC 9V45 96-Core Processor 10.90% 10.27% 0.56% 0.00% 11.17% 10.77% 0.26% 0.00% 11 AMD EPYC 9V45 96-Core Processor 14.03% 16.60% 0.32% 0.00% 14.47% 13.80% 0.94% 0.00% The failing blocking (
pause) runs on the 9V45 produce the same chain we saw locally without blocking:RuntimeError: Failed to unwind stack for thread <- RuntimeError: Failed to collect frames <- RuntimeError: Failed to parse initial frame in chainand the same-pause retries after 1 ms reported
retry succeeded while still paused.One caveat: our original target (
tests/test_expression_arbiter.py) finishes in under 0.2 s underdiagnose.py, so for that target only thelivemode has meaningful samples (33-48 % on all CPUs); on one 9V45 runner the script then hitProcessLookupErrorfrompause_threads()because the target had already exited. The built-in loop results above are the reliable comparison.Full output: AMD EPYC 9V45 (runner 10)
CPU: AMD EPYC 9V45 96-Core Processor BUILD: 3.15.0rc1 (tags/v3.15.0rc1:37e98da, Aug 4 2026, 11:37:18) [MSC v.1951 64 bit (AMD64)] ('CPython', 'tags/v3.15.0rc1', '37e98da') === builtin target 3.15.0rc1 (tags/v3.15.0rc1:37e98da, Aug 4 2026, 11:37:18) [MSC v.1951 64 bit (AMD64)] Windows-2025Server-10.0.26100-SP0 AMD64 Family 26 Model 2 Stepping 1, AuthenticAMD cache=True live: 327/2999 failed (10.90%), 3.001s 327: RuntimeError: Failed to unwind stack for thread <- RuntimeError: Failed to collect frames <- RuntimeError: Failed to parse initial frame in chain pause: 308/3000 failed (10.27%), 3.001s 308: RuntimeError: Failed to unwind stack for thread <- RuntimeError: Failed to collect frames <- RuntimeError: Failed to parse initial frame in chain same-pause retry 10: retry succeeded while still paused delay: 9/1614 failed (0.56%), 3.001s 9: RuntimeError: Failed to unwind stack for thread <- RuntimeError: Failed to collect frames <- RuntimeError: Failed to parse initial frame in chain same-pause retry 9: retry succeeded while still paused context: 0/1741 failed (0.00%), 3.001s === target tests/test_expression_arbiter.py 3.15.0rc1 (tags/v3.15.0rc1:37e98da, Aug 4 2026, 11:37:18) [MSC v.1951 64 bit (AMD64)] Windows-2025Server-10.0.26100-SP0 AMD64 Family 26 Model 2 Stepping 1, AuthenticAMD cache=True live: 23/53 failed (43.40%), 0.053s 16: RuntimeError: Failed to unwind stack for thread <- RuntimeError: Failed to collect frames <- RuntimeError: Failed to parse initial frame in chain 2: RuntimeError: Failed to read interpreter state buffer <- ProcessLookupError: [Errno 3] No such process 1: RuntimeError: Failed to unwind stack for thread <- RuntimeError: Failed to collect frames <- RuntimeError: Failed to parse frame object in chain <- RuntimeError: Unhandled frame owner type in async frame <- RuntimeError: Unhandled frame owner 112. 1: RuntimeError: Failed to unwind stack for thread <- RuntimeError: Failed to collect frames <- RuntimeError: Failed to parse frame object in chain <- RuntimeError: Unhandled frame owner type in async frame <- RuntimeError: Unhandled frame owner 104. 1: RuntimeError: Failed to unwind stack for thread <- RuntimeError: Failed to collect frames <- RuntimeError: Failed to parse frame object in chain <- RuntimeError: Unhandled frame owner type in async frame <- RuntimeError: Unhandled frame owner 103. Traceback (most recent call last): File "D:\a\MoHan-PC-Desktop-Assistant\MoHan-PC-Desktop-Assistant\cpython-158552\diagnose.py", line 168, in <module> main() ~~~~^^ File "D:\a\MoHan-PC-Desktop-Assistant\MoHan-PC-Desktop-Assistant\cpython-158552\diagnose.py", line 164, in main measure(args, mode) ~~~~~~~^^^^^^^^^^^^ File "D:\a\MoHan-PC-Desktop-Assistant\MoHan-PC-Desktop-Assistant\cpython-158552\diagnose.py", line 107, in measure unwinder.pause_threads() ~~~~~~~~~~~~~~~~~~~~~~^^ ProcessLookupError: Process 6892 has terminated === builtin target --no-cache 3.15.0rc1 (tags/v3.15.0rc1:37e98da, Aug 4 2026, 11:37:18) [MSC v.1951 64 bit (AMD64)] Windows-2025Server-10.0.26100-SP0 AMD64 Family 26 Model 2 Stepping 1, AuthenticAMD cache=False live: 335/3000 failed (11.17%), 3.000s 335: RuntimeError: Failed to unwind stack for thread <- RuntimeError: Failed to process frame chain <- RuntimeError: Failed to parse initial frame in chain pause: 323/2999 failed (10.77%), 3.000s 323: RuntimeError: Failed to unwind stack for thread <- RuntimeError: Failed to process frame chain <- RuntimeError: Failed to parse initial frame in chain same-pause retry 9: retry succeeded while still paused same-pause retry 1: RuntimeError: Failed to unwind stack for thread <- RuntimeError: Failed to process frame chain <- RuntimeError: Failed to parse initial frame in chain delay: 4/1559 failed (0.26%), 3.002s 4: RuntimeError: Failed to unwind stack for thread <- RuntimeError: Failed to process frame chain <- RuntimeError: Failed to parse initial frame in chain same-pause retry 4: retry succeeded while still paused context: 0/1713 failed (0.00%), 3.002s === target tests/test_expression_arbiter.py --no-cache 3.15.0rc1 (tags/v3.15.0rc1:37e98da, Aug 4 2026, 11:37:18) [MSC v.1951 64 bit (AMD64)] Windows-2025Server-10.0.26100-SP0 AMD64 Family 26 Model 2 Stepping 1, AuthenticAMD cache=False live: 22/50 failed (44.00%), 0.051s 15: RuntimeError: Failed to unwind stack for thread <- RuntimeError: Failed to process frame chain <- RuntimeError: Failed to parse initial frame in chain 2: RuntimeError: Failed to unwind stack for thread <- RuntimeError: Failed to process frame chain <- RuntimeError: Failed to parse frame object in chain <- RuntimeError: Failed to read function name from code object <- RuntimeError: Failed to read PyUnicodeObject <- OSError: ReadProcessMemory failed for PID 2132 at address 0x1 (size 64, partial read 0 bytes): Windows error 299 <- OSError: [WinError 299] Only part of a ReadProcessMemory or WriteProcessMemory request was completed 1: RuntimeError: Failed to unwind stack for thread <- RuntimeError: Failed to process frame chain <- RuntimeError: Failed to parse frame object in chain <- RuntimeError: Unhandled frame owner type in async frame <- RuntimeError: Unhandled frame owner 38. 1: RuntimeError: Failed to unwind stack for thread <- RuntimeError: Failed to process frame chain <- RuntimeError: Failed to parse frame object in chain <- RuntimeError: Unhandled frame owner type in async frame <- RuntimeError: Unhandled frame owner 104. 1: RuntimeError: Failed to unwind stack for thread <- RuntimeError: Failed to process frame chain <- RuntimeError: Failed to parse frame object in chain <- RuntimeError: Failed to read function name from code object <- RuntimeError: Failed to read PyUnicodeObject <- OSError: ReadProcessMemory failed for PID 2132 at address 0xc0000000 (size 64, partial read 0 bytes): Windows error 998 <- OSError: [WinError 998] Invalid access to memory location Traceback (most recent call last): File "D:\a\MoHan-PC-Desktop-Assistant\MoHan-PC-Desktop-Assistant\cpython-158552\diagnose.py", line 168, in <module> main() ~~~~^^ File "D:\a\MoHan-PC-Desktop-Assistant\MoHan-PC-Desktop-Assistant\cpython-158552\diagnose.py", line 164, in main measure(args, mode) ~~~~~~~^^^^^^^^^^^^ File "D:\a\MoHan-PC-Desktop-Assistant\MoHan-PC-Desktop-Assistant\cpython-158552\diagnose.py", line 107, in measure unwinder.pause_threads() ~~~~~~~~~~~~~~~~~~~~~~^^ ProcessLookupError: Process 4276 has terminatedFull output: AMD EPYC 7763 (runner 1, unaffected)
CPU: AMD EPYC 7763 64-Core Processor BUILD: 3.15.0rc1 (tags/v3.15.0rc1:37e98da, Aug 4 2026, 11:37:18) [MSC v.1951 64 bit (AMD64)] ('CPython', 'tags/v3.15.0rc1', '37e98da') === builtin target 3.15.0rc1 (tags/v3.15.0rc1:37e98da, Aug 4 2026, 11:37:18) [MSC v.1951 64 bit (AMD64)] Windows-2025Server-10.0.26100-SP0 AMD64 Family 25 Model 1 Stepping 1, AuthenticAMD cache=True live: 385/3000 failed (12.83%), 3.000s 385: RuntimeError: Failed to unwind stack for thread <- RuntimeError: Failed to collect frames <- RuntimeError: Failed to parse initial frame in chain pause: 0/3000 failed (0.00%), 3.001s delay: 0/1887 failed (0.00%), 3.001s context: 0/3000 failed (0.00%), 3.000s === target tests/test_expression_arbiter.py 3.15.0rc1 (tags/v3.15.0rc1:37e98da, Aug 4 2026, 11:37:18) [MSC v.1951 64 bit (AMD64)] Windows-2025Server-10.0.26100-SP0 AMD64 Family 25 Model 1 Stepping 1, AuthenticAMD cache=True live: 64/190 failed (33.68%), 0.190s 46: RuntimeError: Failed to unwind stack for thread <- RuntimeError: Failed to collect frames <- RuntimeError: Failed to parse initial frame in chain 2: RuntimeError: Failed to unwind stack for thread <- RuntimeError: Failed to collect frames <- RuntimeError: Failed to parse frame object in chain <- RuntimeError: Unhandled frame owner type in async frame <- RuntimeError: Unhandled frame owner -87. 2: RuntimeError: Failed to unwind stack for thread <- RuntimeError: Failed to collect frames <- RuntimeError: Failed to parse frame object in chain <- RuntimeError: Failed to read function name from code object <- RuntimeError: Failed to read PyUnicodeObject <- OSError: ReadProcessMemory failed for PID 892 at address 0xc0000000 (size 64, partial read 0 bytes): Windows error 998 <- OSError: [WinError 998] Invalid access to memory location 2: RuntimeError: Failed to unwind stack for thread <- RuntimeError: Failed to collect frames <- RuntimeError: Failed to parse frame object in chain <- RuntimeError: Failed to read function name from code object <- RuntimeError: Failed to read PyUnicodeObject <- OSError: ReadProcessMemory failed for PID 892 at address 0x26262626 (size 64, partial read 0 bytes): Windows error 998 <- OSError: [WinError 998] Invalid access to memory location 2: RuntimeError: Failed to unwind stack for thread <- RuntimeError: Failed to collect frames <- RuntimeError: Failed to parse frame object in chain <- RuntimeError: Unhandled frame owner type in async frame <- RuntimeError: Unhandled frame owner -98. Traceback (most recent call last): File "D:\a\MoHan-PC-Desktop-Assistant\MoHan-PC-Desktop-Assistant\cpython-158552\diagnose.py", line 168, in <module> main() ~~~~^^ File "D:\a\MoHan-PC-Desktop-Assistant\MoHan-PC-Desktop-Assistant\cpython-158552\diagnose.py", line 164, in main measure(args, mode) ~~~~~~~^^^^^^^^^^^^ File "D:\a\MoHan-PC-Desktop-Assistant\MoHan-PC-Desktop-Assistant\cpython-158552\diagnose.py", line 107, in measure unwinder.pause_threads() ~~~~~~~~~~~~~~~~~~~~~~^^ ProcessLookupError: Process 7808 has terminated === builtin target --no-cache 3.15.0rc1 (tags/v3.15.0rc1:37e98da, Aug 4 2026, 11:37:18) [MSC v.1951 64 bit (AMD64)] Windows-2025Server-10.0.26100-SP0 AMD64 Family 25 Model 1 Stepping 1, AuthenticAMD cache=False live: 387/3000 failed (12.90%), 3.000s 387: RuntimeError: Failed to unwind stack for thread <- RuntimeError: Failed to process frame chain <- RuntimeError: Failed to parse initial frame in chain pause: 0/3000 failed (0.00%), 3.001s delay: 0/1872 failed (0.00%), 3.001s context: 0/3000 failed (0.00%), 3.001s === target tests/test_expression_arbiter.py --no-cache 3.15.0rc1 (tags/v3.15.0rc1:37e98da, Aug 4 2026, 11:37:18) [MSC v.1951 64 bit (AMD64)] Windows-2025Server-10.0.26100-SP0 AMD64 Family 25 Model 1 Stepping 1, AuthenticAMD cache=False live: 56/155 failed (36.13%), 0.155s 34: RuntimeError: Failed to unwind stack for thread <- RuntimeError: Failed to process frame chain <- RuntimeError: Failed to parse initial frame in chain 5: RuntimeError: Failed to unwind stack for thread <- RuntimeError: Failed to process frame chain <- RuntimeError: Failed to parse frame object in chain <- RuntimeError: Unhandled frame owner type in async frame <- RuntimeError: Unhandled frame owner -87. 3: RuntimeError: Failed to unwind stack for thread <- RuntimeError: Failed to process frame chain <- RuntimeError: Failed to parse frame object in chain <- RuntimeError: Unhandled frame owner type in async frame <- RuntimeError: Unhandled frame owner 10. 2: RuntimeError: Failed to unwind stack for thread <- RuntimeError: Failed to process frame chain <- RuntimeError: Failed to parse frame object in chain <- RuntimeError: Failed to read function name from code object <- RuntimeError: Failed to read PyUnicodeObject <- OSError: ReadProcessMemory failed for PID 3708 at address 0x202626 (size 64, partial read 0 bytes): Windows error 299 <- OSError: [WinError 299] Only part of a ReadProcessMemory or WriteProcessMemory request was completed 2: RuntimeError: Failed to unwind stack for thread <- RuntimeError: Failed to process frame chain <- RuntimeError: Failed to parse frame object in chain <- RuntimeError: Failed to read function name from code object <- RuntimeError: Failed to read PyUnicodeObject <- OSError: ReadProcessMemory failed for PID 3708 at address 0xc0000000 (size 64, partial read 0 bytes): Windows error 998 <- OSError: [WinError 998] Invalid access to memory location Traceback (most recent call last): File "D:\a\MoHan-PC-Desktop-Assistant\MoHan-PC-Desktop-Assistant\cpython-158552\diagnose.py", line 168, in <module> main() ~~~~^^ File "D:\a\MoHan-PC-Desktop-Assistant\MoHan-PC-Desktop-Assistant\cpython-158552\diagnose.py", line 164, in main measure(args, mode) ~~~~~~~^^^^^^^^^^^^ File "D:\a\MoHan-PC-Desktop-Assistant\MoHan-PC-Desktop-Assistant\cpython-158552\diagnose.py", line 107, in measure unwinder.pause_threads() ~~~~~~~~~~~~~~~~~~~~~~^^ ProcessLookupError: Process 6224 has terminated- added a commit that references this issue
on Oct 5, 2026
Bug report
Bug description
On Windows,
python -m profiling.sampling run --blocking ...reports a sample-readerror rate of roughly 31–35% when the host CPU is an AMD EPYC 9V45, while the
same command on other CPUs reports about 0.5%. The high rate matches what we see
without
--blockingon other machines (about 36–39%), so it looks as if theblocking (suspend-then-read) path is not effective on that CPU.
Command (wall mode, 1 kHz, 40 s duration cap; the target finishes on its own):
target.pyis a CPU-bound, single-threaded workload (many short Python calls)that runs for about 4 s on the slower CPUs and about 2 s on the EPYC 9V45.
Observed on GitHub-hosted Windows runners (image
windows-2025-vs2026,version
20260922.246.2), same commit, same command, six consecutive runs:Two other targets profiled in the same jobs without
--blocking(but with--all-threads) stay at about 2 % on every CPU, including the EPYC 9V45.Locally (AMD Ryzen 5 5600X, Windows 11), 10 runs each of the same target:
--blocking0.36–0.51 %; without--blocking36.38–38.79 %. Without blocking,the dominant read error was
Failed to parse initial frame in chain.We do not have access to an EPYC 9V45 machine to bisect further, and we have not
yet reduced this to a standalone script; happy to run diagnostics on GitHub
runners if that helps.
CPython versions tested on:
3.15 (3.15.0rc1)
Operating systems tested on:
Windows
Linked PRs