Repository navigation
Generate centralized Boost.Serialization registrations for Python wrappers #211
Description
Context
borglab/gtsam#2853 fixes Python factor serialization that could fail with RuntimeError: unregistered class or depend on whether gtsam or gtsam_unstable was imported first. The PR shares Boost export-key declarations across binding translation units while retaining each extension's export implementations. The shared header currently lists 19 noise-model and robust-estimator types manually, which must remain synchronized with linear.i and the unstable export list.
In Fan's approving review, he asked to keep an open bug and revisit whether wrap should generate a centralized serialization library instead of maintaining this binding-side workaround.
Follow-up
Investigate generating the required Boost.Serialization registrations from wrap in a centralized, consistently included place, so each binding translation unit sees the export keys before serialization is instantiated without duplicate key specializations. If a centralized library is the right design, migrate the manual declarations in gtsam#2853 to generated output and remove the synchronization burden.
Preserve existing archive identifiers and stable/unstable Python serialization behavior, including stable-only loading and both extension import orders. Keep Python-specific registration choices out of public GTSAM headers unless compatibility with existing C++ archive registrations is addressed. Add or retain regression coverage for those cases.
This is follow-up design work; gtsam#2853 is the immediate bug fix.
Activity