You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
[FEATURE] Migrate experimental features to PGM_Handle instead of PGM_Options #934
This feature instead sets the experimental feature to be part of the handle instead of the options.
Breaking changes to C API (at a major version bump)
To the PGM_Handle, add the experimental_features member field
Update the PGM_set_experimental_features function to remove the PGM_Options* opt from it and move it to the handle header + implemenation.
In the implementation of PGM_set_experimental_features, set the experimental_features member field of the PGM_Handle
Update the C++ API to be able to handle this. Currently, it uses a separate handle for the options, which may result in issues.
Update the tests to be able to handle this. Currently, it uses the PGM_Options for this.
Update the PGM_set_experimental_features function to remove the PGM_Options* opt from it.
NOTE: changing the function signature is not only an API but also an ABI breaking change. This is known and accepted after considering the following reasons:
The PGM_set_experimental_features already is marked both [DANGER MODE] and documented Features marked as 'experimental' as well as the behavior of experimental functionality itself may change over time.
There is a major version bump. It is listed as breaking and documented as such in the release notes and the announcement.
Since we are once again encountering the necessity for this (see also #1334), should we revisit the decision to not plan this?
Some points for the rationale:
it's experimental... it may change... there's no backwards compatibility issues as long as PGM_set_experimental_features(handle, options, disabled) results in the same behavior. It's a non-issue.
we're looking to do a major version bump. Now is the time.
the Python library now uses a thread-local handle, rather than one handle for all threads. Parallel testing is no longer an issue. We just need to disable experimental features after a single thread test is done. Some pattern like with experimental_features_enabled(): ... could do the trick
the whole C++ library is officially experimental (although we know of a user that uses this) so we can change the behavior if we need to. The main issue here is how to set experimental features on an object. A potential resolution could be to optionally provide a C++ constructor overload that takes an external handle instead of one that is created internally in the object.
Activity
mgovers commentedon Jun 11, 2026
#772 (comment)