For optimal performance, data and queries should be distributed evenly across nodes and shards. If some shards are getting more traffic they could become a bottleneck for Scylla.
There could be multiple explanations for these performance issues: data-model, non-prepared statement, or driver.