I was able to reproduce the issue with artificially terminating the process during an import and have now implemented a simple journaling facility for the metadata and thumbnail caches, so that from the next release on, such corrupted cache entries will be detected and recomputed. Performance-wise this also seems to be okay.