conda-smithy: BUG: conda-smithy v3.31 mismatches cuda_compiler & image, breaks zip_keys
I’m getting the following rerender in https://github.com/conda-forge/arrow-cpp-feedstock/pull/1325, which mismatches the cuda_compiler with the image, and then predictably fails.
--- a/.azure-pipelines/azure-pipelines-linux.yml
+++ b/.azure-pipelines/azure-pipelines-linux.yml
@@ -11,7 +11,7 @@ jobs:
linux_64_cuda_compiler_version11.2: # CUDA
CONFIG: linux_64_cuda_compiler_version11.2
UPLOAD_PACKAGES: 'True'
- DOCKER_IMAGE: quay.io/condaforge/linux-anvil-cuda:11.2 # CUDA
+ DOCKER_IMAGE: quay.io/condaforge/linux-anvil-cos7-x86_64 # !! no CUDA !!
linux_64_cuda_compiler_versionNone:
CONFIG: linux_64_cuda_compiler_versionNone
UPLOAD_PACKAGES: 'True'
--- a/.ci_support/linux_64_cuda_compiler_version11.2.yaml
+++ b/.ci_support/linux_64_cuda_compiler_version11.2.yaml # CUDA
@@ -7,15 +7,15 @@ bzip2:
cuda_compiler:
-- nvcc # CUDA
+- None # !! no CUDA !!
cuda_compiler_version:
- '11.2' # CUDA
cuda_compiler_version_min:
@@ -23,23 +23,23 @@ cuda_compiler_version_min:
docker_image:
-- quay.io/condaforge/linux-anvil-cuda:11.2
+- quay.io/condaforge/linux-anvil-cos7-x86_64
gflags:
--- a/.ci_support/linux_64_cuda_compiler_versionNone.yaml
+++ b/.ci_support/linux_64_cuda_compiler_versionNone.yaml # no CUDA
@@ -9,13 +9,13 @@ c_compiler:
cuda_compiler:
-- None # no CUDA
+- nvcc # !! CUDA !!
cuda_compiler_version:
- None # no CUDA
cuda_compiler_version_min:
In fact, for 4/4 builds on the arrow feedstock, the rerender randomly mismatches at least two jobs, seemingly by messing up the zip_keys. Not sure what’s happening here or whether #1815 had a bug or uncovered another problem.
About this issue
- Original URL
- State: closed
- Created 4 months ago
- Comments: 15 (15 by maintainers)
I’ve profiled this coarsely and quickly worked on some changes at the end of last month and but only got to put in https://github.com/conda/conda-build/issues/5224 today. (
conda-build=24.3release processes is starting now, so that would only be available in>=24.5.) With those changes, it is still very far from how quick it should be, but at least it should hopefully avoid >10 minute re-rendering times. Re-renderingarrow-cpp-feedstockwent down to about 3 minutes andctng-compilers-feedstockto about 4 minutes for me – still very bad, IMO; but I had no idea how atrociously slow these things are currently. (Please do file bug reports for such things!)FWIW, I tried your changes from gh-1851 on https://github.com/conda-forge/python-feedstock/pull/656/commits/73873d8de6b643f04ab08356d2da703dbcea552e (which failed with
3.30.4but works with3.31.0; i.e., the reason for gh-1815) and it still works fine for that case, too 👍.The compilers runs ~30 minutes on GHA so a factor of 4 would be a big improvement no matter what. Thank you!
PYTHONHASHSEED=4 will cause the error for those trying this out.
Awesome that you reduced this, thanks for tackling this so quickly!
Yeah… Local rerenders take up to 25min for me now - not fun. 😑