Re: Packages using more than their assigned MAKE_JOBS

* On 2026-08-10 at 12:23 BST, Greg Troxel wrote:

>Jonathan Perkin <[email protected]> writes:
>
>> The issue is py-scikit-build-core which is the build system for this.
>> It delegates to cmake and ninja, and ninja ultimately uses all cores
>> by default (and when building 4 versions of py-duckdb simultaneously
>> in bulk builds, 4x all cores on the host, plus all the other builds,
>> which is enough to kill jenkins.jar for me a few times...)
>
>I must be missing something.
>
>It seems obvious that ninja should run 1 thread only, unless it has been
>given an argument that means -jN (however ninja spells it).   And, that
>if cmake is invoked, it use 1 thread and ask all child make-type
>programs to use only 1 thread, unless cmake was given an explicit
>argument that means -jN.
>
>Is ninja deciding to use hw.ncpu or something as a default, but only
>when run under cmake under scikit-build, because of some environment
>var, or ?  Ninja called from cmake, normally, seems to get this right.

ninja is actually worse than that, it's nproc+2 by default:

   https://github.com/ninja-build/ninja/blob/v1.13.2/src/ninja.cc#L253-L264

scikit-build invokes cmake directly:

   https://github.com/scikit-build/scikit-build-core/blob/v1.0.3/src/scikit_build_core/cmake.py#L345-L350

with ninja as the default generator:

   https://github.com/scikit-build/scikit-build-core/blob/v1.0.3/src/scikit_build_core/builder/generator.py#L90-L105

So yes, anywhere we call ninja or cmake ourselves directly in pkgsrc, we 
ensure the correct flags are passed, but here we have build tools 
calling cmake themselves without flags, and so things go sideways.

In anticipation of a potential thread: I don't think we should patch 
ninja.  We might disagree with the defaults, but users will expect it to 
behave like upstream versions.  "Why is pkgsrc ninja so much slower than 
other OS?", etc.

-- 
Jonathan Perkin                    pkgsrc.smartos.org
Open Source Complete Cloud   www.tritondatacenter.com