Which is better, ModularLoadManager or ExtensibleLoadManager?

[Denovo1998]

Which is more stable in the production environment, ModularLoadManager compared with ExtensibleLoadManager?

Has anyone in this community or in the company compared it?

[lhotari]

I'd say ModularLoadManager. There have been some issues in ExtensibleLoadManager which have been fixed recently (for example #25379 and #25132 (released in 4.0.9/4.1.3)). The fix #25379 is available in 4.2.0 release and will be released also in 4.0.10. The 4.1.x branch is out of support which means that there won't be any future releases.
We are getting closer to the goal of making ExtensibleLoadManager stable.

[lhotari]

But assuming both are in normal operating conditions, will the load balancing effect of ExtensibleLoadManager be better than ModularLoadManager?

ExtensibleLoadManager (PIP-192) supports PIP-307 which helps with more advanced features. Some comments at apache/pulsar-helm-chart#668 (comment).

[Denovo1998]

@lhotari
I have always had a question that I couldn't figure out. Why do we have to split the hot bundle and then unload it (splitting is fine, but unloading feels strange).

If a bundle contains a hot topic, doesn't that mean this topic is very important for the business? Any operation that causes it to become unstable will make it even more unstable.

In this case, why not unload other bundles with low traffic owned by the broker? These bundles with not so high traffic may mostly contain "unimportant topics", and connection jitter doesn't matter. Directly "bind" the "high traffic important" bundle to the current broker without any operation.

[lhotari]

If a bundle contains a hot topic, doesn't that mean this topic is very important for the business? Any operation that causes it to become unstable will make it even more unstable.

In this case, why not unload other bundles with low traffic owned by the broker? These bundles with not so high traffic may mostly contain "unimportant topics", and connection jitter doesn't matter. Directly "bind" the "high traffic important" bundle to the current broker without any operation.

I've been thinking about a similar flaw in how load balancing is performed and about a similar solution. I think it would make more sense than the current strategies.

[lhotari]

One of the problems with rolling restarts is that there isn't persistent information about the statistics for each bundle. In the StreamNative Cloud "Graceful Cluster Rollout", this is addressed by assigning all bundles on a specific broker instance to a replacement broker instance.

[Denovo1998]

It feels like the missing capability is not just a better balancing algorithm, but persistent per-bundle statistics / placement memory and a warmer handoff mechanism.

I have some rudimentary ideas, and I feel that I may need PIP.

There should be data support for the modification of loadmanager. I haven’t been engaged in pulsar-related job yet, and I don‘t have a production environment at hand. But I have an r730 server. I think I will do some scenario tests in my own environment according to the limit load measured by the machine itself before.