constraints. This avoids traversing the archive everytime object data is needed and gives an overall consistent ~2x speedup here with files containing between 136 and 500 Alembic objects. Also this somewhat nicely de- duplicates code between data creation (upon import) and data streaming (modifiers and constraints). The only worying part is what happens when a CacheFile is deleted and/or has its path changed. For now, we traverse the whole scene and for each object using the CacheFile we free the pointer and NULL-ify it (see BKE_cachefile_clean), but at some point this should be re-considered and make use of the dependency graph.