Document missing property - druid.announcer.skipSegmentAnnouncementOnZk

[capistrant]

Description

This property appears to have been omitted from the docs. I don't see why we would want to avoid exposing it. If operators are going to use http server view, why waste the space in zookeeper announcing segments?

I'm not sure on formatting here. I put it in it's own table since it renders the configs in the above table irrelevant. I can change it however we see fit. The important thing I think, is simply making it available in some way so an operator can decide how to properly consume their zookeeper resources depending on the server view they use.

---

Key changed/added classes in this PR

N/A

---

This PR has:

• added documentation for new or modified features or behaviors.

[maytasm]

@capistrant @abhishekagarwal87 Since HTTP-based segment loading is enabled by default, is there any reason the default of druid.announcer.skipSegmentAnnouncementOnZk is still false? Why do we still need to announce segments on zk?

[abhishekagarwal87]

It might have been so because one could switch back to zk based segment announcements without waiting for all the segments to be announced on zk first. Otherwise, if one wants to switch back to http, they would have to first set this flag, wait for segment announcements to be available on zookeeper and then switch back the announcement mechanism to zk.

Given that http based mechanism has been the de-facto one for a while and is very stable, we could indeed consider switching the default to true.

[maytasm]

It might have been so because one could switch back to zk based segment announcements without waiting for all the segments to be announced on zk first. Otherwise, if one wants to switch back to http, they would have to first set this flag, wait for segment announcements to be available on zookeeper and then switch back the announcement mechanism to zk.

Given that http based mechanism has been the de-facto one for a while and is very stable, we could indeed consider switching the default to true.

I have seen some cases where huge number of watches in ZooKeeper is causing performance issue. While there are improvements like #17482, I wonder if just switching this to true will be an easy win. I feel that most people are already using http segment discovery (it has been default since v25) but doesn't know about the existence of this flag. Hence, maybe switching the default of this flag to true will be beneficial to the community. I can raise a PR to switch the default to true unless there is any concern.
We can also deprecate and remove zk segment discovery first before changing the default of this flag to make sure that all Druid clusters will be using http segment discovery. WDYT?

[gianm]

I have seen some cases where huge number of watches in ZooKeeper is causing performance issue. While there are improvements like #17482, I wonder if just switching this to true will be an easy win. I feel that most people are already using http segment discovery (it has been default since v25) but doesn't know about the existence of this flag. Hence, maybe switching the default of this flag to true will be beneficial to the community. I can raise a PR to switch the default to true unless there is any concern.
We can also deprecate and remove zk segment discovery first before changing the default of this flag to make sure that all Druid clusters will be using http segment discovery. WDYT?

IMO, setting the default to true would be good. I don't think there are big backwards compatibility concerns anymore, since http segment discovery has been supported for a while. We would just need a release note.

[maytasm]

I have seen some cases where huge number of watches in ZooKeeper is causing performance issue. While there are improvements like #17482, I wonder if just switching this to true will be an easy win. I feel that most people are already using http segment discovery (it has been default since v25) but doesn't know about the existence of this flag. Hence, maybe switching the default of this flag to true will be beneficial to the community. I can raise a PR to switch the default to true unless there is any concern.
We can also deprecate and remove zk segment discovery first before changing the default of this flag to make sure that all Druid clusters will be using http segment discovery. WDYT?

IMO, setting the default to true would be good. I don't think there are big backwards compatibility concerns anymore, since http segment discovery has been supported for a while. We would just need a release note.

Raised a PR here: #18445
Thanks