conditional-futures

Make multi-threaded concurrency backward- and forward-compatible for the free-threaded future of Python.

Multi-Threading In and Out of Free-Threading

The following is a table of performance results for the execution of a function across each row of a NumPy array (code), with (no GIL) python3.14t and (GIL enabled) python3.14, and with and without ThreadPoolExecutor. Performance improves with python3.14t but degrades with python3.14.

ConditionalThreadPoolExecutor provides a single interface to get the best result in either context.

Introduction

The new free-threaded version of Python (with the GIL disabled) offers extraordinary performance improvements in multi-threading CPU-bound processes. Upgrading your code to take advantage of this performance, however, is problematic. The same multi-threaded code, if run with the GIL enabled, can actually perform significantly worse than single-threaded execution. Even when using a free-threaded interpreter, importing an incompatible C-extension will automatically re-enable the GIL.

For code that will run across many interpreters with or without the GIL, we need interfaces that perform multi-threaded processing only when the GIL is disabled.

The conditional-futures package provides ConditionalThreadPoolExecutor, a drop-in replacement for ThreadPoolExecutor that adapts based on the runtime state of the GIL.

When running under free-threaded Python with the GIL disabled ConditionalThreadPoolExecutor behaves like a normal thread pool. When running under a GIL-enabled build, it falls back on single-threaded execution, potentially avoiding a significant degradation in performance. The same implementation offers optimal performance in all contexts.

Note that, even with the GIL enabled, multi-threading can perform well for I/O-bound processes. ConditionalThreadPoolExecutor is appropriate only for CPU-bound processes that perform worse with the GIL.

Example

The performance of function application on the rows of a 2D NumPy array can be used to show both the benefits of free-threaded Python and the need for ConditionalThreadPoolExecutor.

First, using the free-threaded build of Python 3.14, we can create an array and apply a function to each row of that array. The ipython %time utility is used to measure duration.

$ python3.14t
>>> array = np.arange(100_000_000).reshape(100_000, 1_000)
>>> func = lambda row: (row[row % 2 == 0]**2).sum()
>>> %time _ = np.fromiter((func(row) for row in array), dtype=float, count=array.shape[0])
CPU times: user 580 ms, sys: 662 μs, total: 580 ms
Wall time: 581 ms

Using ConditionalThreadPoolExecutor with this GIL-disabled build of Python we can take advantage of multi-threaded performance on a CPU-bound process: the same routine is almost twice as fast:

>>> with ConditionalThreadPoolExecutor(max_workers=4) as ex:
...     %time _ = np.fromiter(ex.map(func, array), dtype=float, count=array.shape[0])
...
CPU times: user 1.31 s, sys: 98 ms, total: 1.41 s
Wall time: 352 ms

Now, if using the standard Python 3.14 interpreter (with the GIL enabled), ThreadPoolExecutor degrades performance: the same operation takes six times as long!

$ python3.14
>>> array = np.arange(100_000_000).reshape(100_000, 1_000)
>>> func = lambda row: (row[row % 2 == 0] ** 2).sum()
>>> with ThreadPoolExecutor(max_workers=4) as ex:
...     %time _ = np.fromiter(ex.map(func, array), dtype=float, count=array.shape[0])
...
CPU times: user 1.9 s, sys: 2.21 s, total: 4.12 s
Wall time: 2.33 s

Using ConditionalThreadPoolExecutor, one implementation performs optimally in both contexts. Running the same code with python3.14, ConditionalThreadPoolExecutor does not perform as well as with python3.14t, but provides the best option available: single-threaded performance.

>>> with ConditionalThreadPoolExecutor(max_workers=4) as ex:
...     %time _ = np.fromiter(ex.map(func, array), dtype=float, count=array.shape[0])
...
CPU times: user 532 ms, sys: 773 μs, total: 533 ms
Wall time: 533 ms

Installation

pip install conditional-futures

What is New in conditional-futures

1.0.2

Extended documentation.

1.0.1

Extended documentation.

1.0.0

Initial release.