Per a comment in #17644, we need an acceptance test for scatter. One option is to bundle the split+scatter test that uncovered #17644 into an acceptance test:

Yes, scatter needs better testing. The test would essentially be kv --concurrency 512 --splits 1000000 --max-ops 1. Question is how to determine whether scatter worked well. I suppose we could look at the number of replicas vs ranges and the balance across the cluster.