Geo-replication with and without Global Zookeeper

[shasank112001]

I have been looking at enabling geo-replication for my Pulsar setup.

I have gone down the path of setting it up with the Global Zookeeper and also without it, and I have a few questions associated with these setups.

As the role of global zookeeper is not very well documented, and the geo-replication without global zookeeper is not even mentioned in the documentation, its hard to understand what the global zookeeper actually brings to the table.

I have made the following observations in my current testing:

• Tenant and Namespace can be replicated via the GZK but Topics cannot (unless the namespace level replication is used)
• If Replication is required at topic level, with or without GZK, user's have to follow the following steps:
  1. Create Tenant and Namespace in both clusters (via GZK or manually)
  2. Create Topic in cluster-a
  3. Create Topic in cluster-b (required only if auto-topic-creation is disabled)
  4. Set Replication on Topic to cluster-a,cluster-b
  5. Set Replication on Topic to cluster-a,cluster-b (required to set up bi-directional replication)
• Replicated Subscriptions when created via the client in GZK setup leads to the backlog being replicated, however, if subscription is created and replication is enabled via ADMIN API, the subscription is not created on both clusters.

Therefore, I have the following questions:

1. Is the 3rd observation listed above a bug or was there a particular reason for this behaviour?
2. Why are topic policies and topic replication not enabled via GZK, and if I plan to only use Topic level Replication, does it make sense to have GZKs?
3. In a replicated seup, what configuration is under the responsibility of GZK? (I noticed that partitioned-topic metadata is persisted on the GZK instances, but I have not been able to figure out why thats the case and what is it used for)
4. Is there a way to enable geo-replication (on topic or namespace level) without having auto-topic-creation=true?
5. If I go down the route of replication without GZK, is it possible to later add the GZK and have all the replicated tenants and namespaces be synced?