fix linux PGO wheel build #1557
Conversation
CodSpeed Performance ReportMerging #1557 will not alter performanceComparing Summary
|
| - name: prepare profiling directory | ||
| shell: bash | ||
| # making this ahead of the compile ensures that the local user can write to this | ||
| # directory; the maturin action (on linux) runs in docker so would create as root | ||
| run: mkdir -p ${{ github.workspace }}/profdata |
There was a problem hiding this comment.
This turned out to be the critical change; as per the comment without this the linux jobs (which are built inside manylinux docker containers) were failing to write out their profiling information.
There was a problem hiding this comment.
is there a way to build with "require profiling information" so build fails if it can't be found?
There was a problem hiding this comment.
The annoying thing is that the build already emits profiling information for the proc macros, so there will always be some. I'm also not sure of a way to enforce.
At least with the way CI is now set up, if the PGO breaks we will see codspeed regressions.
| - name: prepare profiling directory | ||
| shell: bash | ||
| # making this ahead of the compile ensures that the local user can write to this | ||
| # directory; the maturin action (on linux) runs in docker so would create as root | ||
| run: mkdir -p ${{ github.workspace }}/profdata |
There was a problem hiding this comment.
is there a way to build with "require profiling information" so build fails if it can't be found?
Change Summary
Changes CI to use the same PGO-optimized wheel build steps for both codspeed and release. I have a concern that the PGO step may not be working as intended; running the wheel through codspeed will be a good sanity check.
Related issue number
N/A
Checklist
pydantic-core(except for expected changes)