Documentation: Suggest adding documentation for supported bundle split algorithms

[frankjkelly]

Is your enhancement request related to a problem? Please describe.
Add documentation for how to choose between load balancing algorithms

Describe the solution you'd like

• supportedNamespaceBundleSplitAlgorithms supports two algorithms for namespace bundle split range_equally_divide,topic_count_equally_divide
• Suggest adding some documentation around that to here https://pulsar.apache.org/docs/en/administration-load-balance/ including (1) documentation how each algorithm works and how each algorithm can be adjusted (any parameters to control operation) (2) documentation how and when to choose between the default range_equally_divide and topic_count_equally_divide

CC: @jennifer88huang

[Anonymitaet]

@frankjkelly thanks for raising this up. Would you like to add docs accordingly? Then we can help review, thanks

[frankjkelly]

@Anonymitaet I would love to but I really do not have the expertise to understand how these mechanisms work at a detailed level and to provide some guidance on how to choose between them.

[Anonymitaet]

Currently, we have documented the usage in broker.conf as below, but do not provide more explanations in the user doc.

# Supported algorithms name for namespace bundle split.
# "range_equally_divide" divides the bundle into two parts with the same hash range size.
# "topic_count_equally_divide" divides the bundle into two parts with the same topics count.
supportedNamespaceBundleSplitAlgorithms=range_equally_divide,topic_count_equally_divide

[Anonymitaet]

I have a few questions:

• How does each algorithm work with other split bundle configurations?
  

• Is each algorithm limited by other split bundle configurations?

[frankjkelly]

Thanks @Anonymitaet exactly - it gets pretty interesting to understand how all those parameters will interact.
I'm particularly interested in how each algorithm behaves when we add brokers (we're hoping to implement Kubernetes autoscaling)- how quickly would those brokers be discovered and how quickly some load would be redirected to those new brokers.

In my limited experience with range_equally_divide it's not until we hit the loadBalancerBrokerOverloadedThresholdPercentage and the broker is overloaded that some namespace bundles will be split and offloaded but at that point usually the load on the broker is pretty high so that the offload seems to be "expensive" or at least take a long time and then our Pulsar Clients start timing out (we have them set with a 3 second timeout due to the near-real-time nature of our use-cases).

For us we'd rather have a way that was a faster balancing (but perhaps less precisely load balanced) than one that was more accurately load balanced but takes longer to shed load. Again we're pretty new to this but would appreciate to hear perspectives or pointers to code. Thanks!

[congbobo184]

@Anonymitaet each algorithm is affected by these configurations the same. range_equally_divide use hash range, topic_count_equally_divide is evenly distributed.