This commit has two major goals:
- fix the caching of the QPY files for both the `main` and `stable/*`
branches
- increase the number of compatibility tests between the different
symengine versions that might be involved in the generation and
loading of the QPY files.
Achieving both of these goals also means that it is sensible to move the
job to GitHub Actions at the same time, since it will put more pressure
on the Azure machine concurrency we use.
Caching
-------
The previous QPY tests attempted to cache the generated files for each
historical version of Qiskit, but this was unreliable. The cache never
seemed to hit on backport branches, which was a huge slowdown in the
critical path to getting releases out. The cache restore keys were also
a bit lax, meaning that we might accidentally have invalidated files in
the cache by changing what we wanted to test, but the restore keys
wouldn't have changed.
The cache files would fail to restore as a side-effect of ed79d42
(Qiskitgh-11526); QPY was moved to be on the tail end of the lint run, rather
than in a test run. This meant that it was no longer run as part of the
push event when updating `main` or one of the `stable/*` branches. In
Azure (and GitHub Actions), the "cache" action accesses a _scoped_
cache, not a universal one for the repository [^1][^2]. Approximately,
base branches each have their own scope, and PR events open a new scope
that is a child of the target branch, the default branch, and the source
branch, if appropriate. A cache task can read from any of its parent
scopes, but write events go to the most local scope. This means that we
haven't been writing to long-standing caches for some time now. PRs
would typically miss the cache on the first attempt, hit their
cache for updates, then miss again once entering the merge queue.
The fix for this is to run the QPY job on branch-update events as well.
The post-job cache action will then write out to a reachable cache for
all following events.
Cross-symengine tests
---------------------
We previously were just running a single test with differing versions of
symengine between the loading and generation of the QPY files. This
refactors the QPY `run_tests.sh` script to run a full pairwise matrix of
compatibility tests, to increase the coverage.
[^1]: https://docs.github.com/en/actions/writing-workflows/choosing-what-your-workflow-does/caching-dependencies-to-speed-up-workflows#restrictions-for-accessing-a-cache
[^2]: https://learn.microsoft.com/en-us/azure/devops/pipelines/release/caching?view=azure-devops#cache-isolation-and-security