In bare-metal the default is to cache a saved copy of the query dll, and serialize it to the slaves (where it also cached).
In k8s the default is to load the query dll directly from the 'queries' plane both on the Thor manager and the workers.
If the queries plane is slow storage (e.g. blob storage) this can cause excessive latency (probably especially on large clusters, where all workers are trying to read simultaneously).
Add an option "saveQueryDlls" which copies and serializes the query dll in similar way to bare-metal detail (default off)