Description
In #110, we added checks on package size for wheels across RAPIDS.
In the 1.5 years that's been in place, I've seen it be valuable for catching unintended growth in package size and enforcing the social norm that we generally want RAPIDS packages to get smaller over time.
We prioritized wheels CI because we were focused on pypi.org's size restrictions.
Proposal: we should add similar checks in CI for RAPIDS conda packages.
The packaging for conda packages is often quite different (e.g. different choices around static vs. dynamic linking, vendoring) and RAPIDS has some
Benefits of this work
Will further help us to detect unexpected growth in package sizes.
The packaging for conda packages is often quite different from wheels (e.g. different choices around static vs. dynamic linking, vendoring) and RAPIDS has some packages that are only built as conda packages and not wheels (example: libraft-headers-only).
Acceptance Criteria
- conda CI enforce package size limits across RAPIDS
Approach
pydistcheck (https://github.com/jameslamb/pydistcheck) has support for conda packages (both .tar.bz2 and .conda), and could be used to do this enforcement similar to the way these are implemented in wheels CI, e.g. https://github.com/rapidsai/cudf/blob/b314d96c9473f70af9de53b7858efdd50b42be34/ci/validate_wheel.sh
That would be my preferred approach, but I'm of course biased because pydistcheck is something I wrote.
Other options include:
- checking sizes with Linux tools like
du or stat
- checking sizes custom code, e.g. in Python
Notes
Inspired by this conversation: rapidsai/shared-workflows#585 (review)
Description
In #110, we added checks on package size for wheels across RAPIDS.
In the 1.5 years that's been in place, I've seen it be valuable for catching unintended growth in package size and enforcing the social norm that we generally want RAPIDS packages to get smaller over time.
We prioritized wheels CI because we were focused on
pypi.org's size restrictions.Proposal: we should add similar checks in CI for RAPIDS
condapackages.The packaging for conda packages is often quite different (e.g. different choices around static vs. dynamic linking, vendoring) and RAPIDS has some
Benefits of this work
Will further help us to detect unexpected growth in package sizes.
The packaging for conda packages is often quite different from wheels (e.g. different choices around static vs. dynamic linking, vendoring) and RAPIDS has some packages that are only built as conda packages and not wheels (example:
libraft-headers-only).Acceptance Criteria
Approach
pydistcheck(https://github.com/jameslamb/pydistcheck) has support for conda packages (both.tar.bz2and.conda), and could be used to do this enforcement similar to the way these are implemented in wheels CI, e.g. https://github.com/rapidsai/cudf/blob/b314d96c9473f70af9de53b7858efdd50b42be34/ci/validate_wheel.shThat would be my preferred approach, but I'm of course biased because
pydistcheckis something I wrote.Other options include:
duorstatNotes
Inspired by this conversation: rapidsai/shared-workflows#585 (review)