Currently there is two-level caching of meta-objects (tables and indexes) in the handler. Only the global cache has reference counting. This causes handler instances sometimes to get stale meta-object references to altered tables (or rebuilt indexes) when being passed between threads. If the alter table is done in the same MySQL server then using global cache only would allow for detecting this without needing retries (for alter tables in other servers retries will still be needed). To implement this the local cache in the ndb objects need to be bypassed for the handler. (It cannot be removed since it is still useful for applications using NDB API directly).