This could be, in part, due to the recent (just prior to 3.13.0b1) change (#118539) toconfigure option--with-system-libmpdec to now default toyes rather thanno, meaning that, by default, an external ("system") version of libmpdec needs to be provided for cpython builds rather than automatically using the now-deprecated vendored libmpdec long included in the cpython source. If so, the failing buildbot configurations need to be changed to install an externallibmpdec package or to specify--with-system-libmpdec=no (for 3.13). Presumably, many downstream builders of Python 3.13 will inadvertently run into this, too.@zware,@erlend-aasland,@Yhg1s

Thanks for the info, Ned! I tuned all that out while it was happening.

On the one hand, I'm happy to do what it takes so that Python doesn'tneedlibmpdec for Python to run.decimal is a standard library, but - although CPython intended never to need to use it anymore - we still ship the pure Python workalike_pydecimal.py.

On the other ...libmpdec is a major piece of work, and is enormously faster than the Python workalike for high-precision applications. Case in point: on the failing buildbots that left behind enough clues to nail it,test_int'stest_whitebox_dec_str_to_int_inner_failsafe() timed out after some 20 minutes. On my box, it takes well under 5 seconds.

It's not just C v Python speed at work.limpmdec also implements sophisticated (& very involved) numeric algorithms with far better asymptotics than_pydecimal.py, or CPython'sC code, attempt. So a system that "suddenly" stops supplying the C version will sorely annoy some serious users. It's so much slower it will look to them like new bugs introduced infinite loops.

Yes, I think part of the issue is the name of the configure option:--with-system-libmpdec tends to imply that libmpdec is a standard system library provided by default by most operating system platforms when that's not typically the case.

@Yhg1s, I think we should consider doing something for 3.13 to make this less of a potential footgun for our downstream builders and users, if nothing else, making it less easy to overlook the fact that_decimal is accidentally no longer being built due to the lack of alibmpdec. I'm not sure what the best way to do that is. A first step might be to create or re-open an issue and mark asrelease blocker?

The--with-system-libmpdec issue should be fixed by#119196.

I mentioned this to Ned yesterday in person, but yeah, I think the temporary solution is to fall back to the bundled version if the system-installed version isn't available. We may want to warn that the bundled version is going away if that's still the plan, and we can warn more aggressively going forward (even going so far as to reflect that decimal is using the bundled libmpdec in the module somewhere, so we can have tests start failing without breaking the build), but that's for later I think.

@Yhg1s, yep that's what was merged in#119196; we print a warning (IIRC, autoconf outputs warnings to stderr) that a fallback is used, and that the bundled copy is deprecated and scheduled for removal.