As of now we migrated back to role-less cluster the decision that will strike back when primary will become a bottleneck that has to process all that at the same time:

• apply finalizing processing
• serve primitives for data in dbroot1
• handle syscat queries processing

For a cluster with a shared storage the good but tedious solution is to separate metadata and syscat data from generic data in dbroot1 and mount only metadata and syscat.
The moderate solution is to write less into dbroot1 after we have at last 3-nodes cluster.