3 racks with more than one node per rack cannot be deployed from scratch

[pchmieli]

Bug Report

What did you do?
Create cassandracluster with one dc, 3 racks and 2 nodesPerRacks:

  topology:
    dc:
    - name: dc1
      nodesPerRacks: 2
      rack:
      - labels:
          dynatrace.tier: cassandra-v1
          topology.kubernetes.io/zone: us-east4-b
        name: rack1
      - labels:
          dynatrace.tier: cassandra-v1
          topology.kubernetes.io/zone: us-east4-a
        name: rack2
      - labels:
          dynatrace.tier: cassandra-v1
          topology.kubernetes.io/zone: us-east4-c
        name: rack3
      resources: {}

What did you expect to see?
A 3 statefulsets with 2 pods each, all ready and running.

What did you see instead? Under which circumstances?
Roll-out stuck on 4th node (rack1 2/2, rack2 1/2, rack3 not created yet). Cassandra fails with:

Token allocation failed: the number of racks 2 in datacenter dc1 is lower than its replication factor 3.
Fatal configuration error; unable to start server.  See log for stacktrace.
ERROR [main] 2025-09-22 16:32:13,409 CassandraDaemon.java:900 - Fatal configuration error
org.apache.cassandra.exceptions.ConfigurationException: Token allocation failed: the number of racks 2 in datacenter dc1 is lower than its replication factor 3.

It's connected with the config field introduced from what I know in cassandra 4.1:

allocate_tokens_for_local_replication_factor: 3

Workarounds

1. Removing allocate_tokens_for_local_replication_factor field from config helps to create >3 cluster from scratch but it's no go in my case.
2. Creating 3 nodes cluster first and just once it's ready scaling it out to 6 or more nodes.

Environment

• casskop version:
  2.3.0

• Kubernetes version information:
  Azure: v1.32.6
  GCP: v1.32.9-gke.1072000

• Kubernetes cluster kind:
  Azure / GCP

• Cassandra version:
  4.1.10

Possible Solution
Whenever deploying a cluster with nodesPerRacks > 1 start from deploying "first layer" first = ensure that each statefulset has exactly 1 pod and only then proceed to next step (which is scaling statefulset to the desired size).

[pchmieli]

I've got a solution which is manually tested already, will soon come back with some PR. Would be cool if @cscetbon or @AKamyshnikova review the issue itself.