Summary
GROUP=Everything fails on current main when QA runs after the Enzyme and Mooncake groups in the same Julia process. The functional groups themselves pass, and GROUP=QA passes in a fresh process, so the regular per-group CI matrix does not expose this.
On ac2b103c99ea5b1ef6f7c3af6de53fc36ad496d6 with Julia 1.12.6, SciMLTesting 2.4.0, and ExplicitImports 1.15.0:
- Core: 51/51 plus BigFloat 7/7 pass
- Enzyme: 68/68 pass
- Mooncake: 13/13 pass
- nopre/JET: 7/7 plus BigFloat 7/7 pass
- QA: 18 pass, 3 errors
The three QA errors are:
no_implicit_imports: FunctionWrappersWrappersEnzymeExt implicitly imports Enzyme, EnzymeCore, BatchDuplicated, Const, Duplicated, Forward, and ForwardWithPrimal.all_qualified_accesses_are_public: FunctionWrappersWrappersMooncakeExt accesses non-public Core.Typeof and Mooncake's build_rrule, primal, rrule!!, and tangent_type.all_explicit_imports_are_public: the Mooncake extension imports non-public @is_primitive, CoDual, MinimalCtx, NoRData, NoTangent, fdata, and zero_tangent.
Switching group environments in-process also prints extension-loading errors because packages loaded by an earlier group are absent from the newly activated environment. This appears to be related state leakage, though the three ExplicitImports findings above are genuine and should not be hidden with catch-all allowlists.
Reproduction
git checkout ac2b103c99ea5b1ef6f7c3af6de53fc36ad496d6
GROUP=Everything julia +1.12 --startup-file=no --project=. -e 'using Pkg; Pkg.test()'
This exits 1 after the QA summary:
Test Summary: | Pass Error Total
QA | 18 3 21
In contrast, a fresh-process QA run succeeds:
GROUP=QA julia +1.12 --startup-file=no --project=. -e 'using Pkg; Pkg.test()'
Test Summary: | Pass Total
QA | 21 21
Regression boundary
The first repository commit where this exact GROUP=Everything command exposes the failure is f8b7e1d6a009709b33e6370810264963cce1dd19 (Update QA envs for SciMLTesting v2.1): it visits Core, Enzyme, Mooncake, nopre, then fails QA with the same three errors.
Its immediate parent 6ccf3d9b0669d242da50b746dfc2a0a9f60c3725 exits 0 with SciMLTesting 1.8.0, but under those older semantics GROUP=Everything only runs Core. Thus f8b7e1d exposed pre-existing extension issues rather than introducing the extension code itself.
Possible resolution
- Make the Enzyme extension's imports explicit.
- Replace Mooncake internal API usage with documented public API, or promote and document the required API in Mooncake before depending on it.
- Consider whether SciMLTesting's
Everything execution should isolate per-environment groups so loaded extensions and unavailable extension dependencies cannot leak across environment switches.
Summary
GROUP=Everythingfails on currentmainwhen QA runs after the Enzyme and Mooncake groups in the same Julia process. The functional groups themselves pass, andGROUP=QApasses in a fresh process, so the regular per-group CI matrix does not expose this.On
ac2b103c99ea5b1ef6f7c3af6de53fc36ad496d6with Julia 1.12.6, SciMLTesting 2.4.0, and ExplicitImports 1.15.0:The three QA errors are:
no_implicit_imports:FunctionWrappersWrappersEnzymeExtimplicitly importsEnzyme,EnzymeCore,BatchDuplicated,Const,Duplicated,Forward, andForwardWithPrimal.all_qualified_accesses_are_public:FunctionWrappersWrappersMooncakeExtaccesses non-publicCore.Typeofand Mooncake'sbuild_rrule,primal,rrule!!, andtangent_type.all_explicit_imports_are_public: the Mooncake extension imports non-public@is_primitive,CoDual,MinimalCtx,NoRData,NoTangent,fdata, andzero_tangent.Switching group environments in-process also prints extension-loading errors because packages loaded by an earlier group are absent from the newly activated environment. This appears to be related state leakage, though the three ExplicitImports findings above are genuine and should not be hidden with catch-all allowlists.
Reproduction
git checkout ac2b103c99ea5b1ef6f7c3af6de53fc36ad496d6 GROUP=Everything julia +1.12 --startup-file=no --project=. -e 'using Pkg; Pkg.test()'This exits 1 after the QA summary:
In contrast, a fresh-process QA run succeeds:
GROUP=QA julia +1.12 --startup-file=no --project=. -e 'using Pkg; Pkg.test()'Regression boundary
The first repository commit where this exact
GROUP=Everythingcommand exposes the failure isf8b7e1d6a009709b33e6370810264963cce1dd19(Update QA envs for SciMLTesting v2.1): it visits Core, Enzyme, Mooncake, nopre, then fails QA with the same three errors.Its immediate parent
6ccf3d9b0669d242da50b746dfc2a0a9f60c3725exits 0 with SciMLTesting 1.8.0, but under those older semanticsGROUP=Everythingonly runs Core. Thusf8b7e1dexposed pre-existing extension issues rather than introducing the extension code itself.Possible resolution
Everythingexecution should isolate per-environment groups so loaded extensions and unavailable extension dependencies cannot leak across environment switches.