We expect that wal_sync() will flush all prepared transactions on disk, but this is not the case for transactions waiting for journal size threshold. In particular in vinyl we flush prepared transactions before starting index build. We do not have a rollback trigger for these transactions thus we make sure if they will be rolled back we don't need to undo their effect. So if there is a prepared transaction waiting for journal size threshold and it will be rolled back it effect will stay. So secondary index will be inconsistent.