We currently define a by_ns index on the page_revisions table, which lists all revisions in a namespace alphabetically. On large wikis, this would map hundreds of millions of title / revision pairs into a single Cassandra partition.
We currently don't use this index at all, so I think we should consider simply dropping it. Use cases like alphabetically listing titles in a specific namespace can be tackled later in connection with T106455. We might want to consider some form of range index partitioning in the storage backends before we do so.