Hey Colin, The KIP makes sense overall. Nice to clarify the contract between clients and the controllers. The use of `DirectToKRaftControllerQuorum` will help prevent misconfiguration. In fact, I wonder if we can return a fatal error instead of NOT_CONTROLLER so that the client would immediately fail. For example, could we use INVALID_REQUEST or something like that? Either that or we need to make sure clients treat NOT_CONTROLLER as a fatal error. Without that, it would probably get picked up with default retry logic and the user might not see the problem.
I also wonder if we can relax the requirements on the Metadata request so that we can use it to list topics and partition state (e.g. URPs). It would be useful to query the controllers as the metadata source of truth when we suspect that the broker metadata may have diverged. Thanks, Jason On Thu, Apr 20, 2023 at 5:53 PM Colin McCabe <cmcc...@apache.org> wrote: > On Wed, Apr 19, 2023, at 20:56, Philip Nee wrote: > > Hey Colin, > > > > I still need to finish reading and understanding the KIP, but I have a > > couple of comments despite being ignorant of most of the KRaft stuff. > > (Sorry!) > > > > Firstly, does it make sense to create an extension of the current > > AdminClient only to handle these specific KRaft use cases? It seems > > cumbersome to have two sets of bootstrap configurations to make the > > AdminClient generic enough to handle these specific cases, instead, maybe > > it is more obvious (to me) to just extend the AdminClient. What I'm > > thinking is KraftAdminClient which continuously uses *bootstrap.servers*, > > but make this class only serves the Kraft controllers APIs. > > Hi Philip, > > Thanks for taking a look. > > We would not want to create a new Admin client class in order to > communicate directly with the controllers. The RPCs accepted by the > controllers have the same format as the those accepted by the brokers. > There is no difference in what is sent over the wire or the data structures > that are used in the client. > > I know you mentioned you haven't had time to read all the KRaft stuff (and > there is a lot, I understand). But this is one area that would probably be > clarified if you were able to read at least KIP-500 and KIP-590. RPCs sent > to the broker are forwarded to the controller (mostly) without modification. > > > > > Secondly, if we want to continue with the design, I'm not yet sure why we > > can't continue using the *bootstrap.servers*? I assume when the client > gets > > the metadata, it should know who it is talking to. I'm just reconsidering > > your alternative again. > > > > A bad idea: Why don't we continue using *bootstrap.servers* but have a > > separated config like *kraft.controller* = true/false. I feel like most > > users might not know what is a controller and causes some mistakes down > the > > road. > > > > Well, you called it a bad idea, and I can't help but agree! :) > > I think "the user might not know what a controller is" is a good sign that > they shouldn't be interacting with the controller. Like many AdminClient > APIs, this one is intended for use by administrators only. Most users > indeed should not need to know what a controller is or interact with it > directly. If they do interact it should be very clear that they are doing > so. The --controller-bootstrap flag makes it very clear, as does the > separate configuration. > > Let me give an example of the kind of problems that arise if you want to > reuse bootstrap.servers for this purpose. > > If the user grasb a 3.4 Kafka AdminClient and set bootstrap.servers to a > set of controller servers, and set direct.to.controller to true, the > unknown (in 3.4) configuration will be ignored, and a normal metadata > request will be sent without the direct to controller flag. In that case it > will give back an error. Confusing, right? > > Using controller.servers clarifies this situation because the 3.4 client > will not support that config, and will complain about the lack of > bootstrap.servers. > > Actually, the situation could get even more confusing than what I > described since some older preproduction versions of the KRaft controller > did implement the METADATA RPC. So if you send them a METADATA request > without any special information, it's not clear what you'll get. Indeed, > it would be dependent on the client version and the controller version. > > The bottom line is that reusing the bootstrap.servers configuration here > is not a good idea. > > best, > Colin > > > Thanks, > > P > > > > On Wed, Apr 19, 2023 at 2:18 PM Colin McCabe <cmcc...@apache.org> wrote: > > > >> Hi all, > >> > >> I wrote a short KIP about allowing AdminClient to talk directly with the > >> KRaft controller quorum. Check it out here: > >> > >> https://cwiki.apache.org/confluence/x/Owo0Dw > >> > >> best, > >> Colin > >> >
