When running a uniform write workload, on a uniformly sized keyspace, ranges will split in half at approximately the same time.
Normally, when a few ranges split this isn't expected to have any impact as https://github.com/cockroachdb/cockroach/issues/22348 removes the latch contention latency. However, scanning the range data simultaneously for many splits can cause CPU exhaustion, which does result in higher tail latency.
This issue is to smear, or space out size-based range splits so that they do not all occur at the same time for uniform workloads.
When running a uniform write workload, on a uniformly sized keyspace, ranges will split in half at approximately the same time.
Normally, when a few ranges split this isn't expected to have any impact as https://github.com/cockroachdb/cockroach/issues/22348 removes the latch contention latency. However, scanning the range data simultaneously for many splits can cause CPU exhaustion, which does result in higher tail latency.
This issue is to smear, or space out size-based range splits so that they do not all occur at the same time for uniform workloads.
Jira issue: CRDB-39670
Epic CRDB-39952