Hi Arjun, >From a high level, I feel like we are making light of the JMX api because it's convenient and the broker already has it. Personally I would take the broker out of the picture. The JMX endpoint is not something we were happy with, hence KIP-412. Ultimately I think we will deprecate and remove it and there's no point trying to standardize on a deprecated mechanism. Thinking just about connect, we already have an HTTP endpoint. The default position should be to add new APIs to it rather than introducing other mechanisms. The fewer ways you have to interact with a system, the better, right?
I think the main argument against a REST endpoint is basically that adjusting log levels is an administrative operation and connect is lacking an authorization framework to enforce administrative access. The same argument applies to JMX, but it has the benefit that you can specify different credentials and it is easier to isolate since it is running on a separate port. As you suggested, I think the same benefits could be achieved by having a separate /admin endpoint which is exposed (perhaps optionally) on another listener. This is a pretty standard pattern. If memory serves, dropwizard has something like this out of the box. We should think hard whether there are additional administrative capabilities that we would ultimately need. The answer is probably yes, so unless we want to double down on JMX, it might be worth thinking through the implications of an admin endpoint now so that we're not left with odd compatibility baggage in the future. Thanks, Jason On Fri, Aug 23, 2019 at 5:38 PM Arjun Satish <arjun.sat...@gmail.com> wrote: > Jason, > > Thanks for your comments! > > I understand the usability issues with JMX that you mention. But it was > chosen for the following reasons: > > 1. Cross-cutting functionality across different components (Kafka brokers, > Connect workers and even with Streams jobs). If we go down the REST route, > then brokers don't get this feature. > 2. Adding this to existing REST servers adds the whole-or-nothing problem. > It's hard to disable an endpoint if the functionality is not desired or > needs to be protected from users (Connect doesn't have ACLs which makes > this even harder to manage). Adding endpoints to different listeners makes > configuring Connect harder (and it's already a hard problem as it is). A > lot of the existing functionality there is driven around the connector data > model (connectors, plugins, their statuses and so on). Adding an '/admin' > endpoint may be a way to go, but that has tremendous implications (we are > effectively adding an administration endpoint similar to the admin one in > brokers), and probably requires a KIP of its own with discussions catered > around just that. > 3. JMX is currently AK's default way to report metrics and perform other > operations. Changing log levels is typically a system level/admin > operation, and fits better there, instead of REST APIs (which is more user > facing). > > Having said that, I'm happy to consider alternatives. JMX seemed to be the > lowest hanging fruit. But if there are better ideas, we can consider them. > At the end of the day, when we download and run Kafka, there should be one > way to achieve the same functionality among its components. > > Finally, I hope I didn't convey that we are reverting/changing the changes > made in KIP-412. The proposed changes would be an addition to it. It will > give brokers multiple ways of changing log levels. and there is still a > consistent way of achieving cross component goals of the KIP. > > Best, > > > On Fri, Aug 23, 2019 at 4:12 PM Jason Gustafson <ja...@confluent.io> > wrote: > > > Let me elaborate a little bit. We made the decision early on for Connect > to > > use HTTP instead of Kafka's custom RPC protocol. In exchange for losing > > some hygienic consistency with Kafka, we took easier integration with > > management tools. The scope of the connect REST APIs is really managing > the > > connect cluster. It has endpoints for creating connectors, changing > > configs, seeing their health, etc. Doesn't debugging fit in with that? I > am > > not sure I see why we would treat this as an exceptional case. > > > > I personally see JMX as a necessary evil in Kafka because most metrics > > agents have native support. But it is particularly painful when it comes > to > > use as an RPC mechanism. This was the central motivation behind KIP-412, > > which makes it very odd to see a new proposal which suggests > standardizing > > on JMX for log level adjustment. I actually see this as something we'd > want > > to eventually turn off in Kafka. Now that we have a proper API with > support > > in the AdminClient, we can deprecate and eventually remove the JMX > > endpoint. > > > > Thanks, > > Jason > > > > On Fri, Aug 23, 2019 at 10:49 AM Jason Gustafson <ja...@confluent.io> > > wrote: > > > > > Hi Arjun, > > > > > > Thanks for the KIP. Do we really need a JMX-based API? Is there > literally > > > anyone in the world that wants to use JMX if they don't have to? I > > thought > > > one of the major motivations of KIP-412 was how much of a pain JMX is. > > > > > > Thanks, > > > Jason > > > > > > On Mon, Aug 19, 2019 at 5:28 PM Arjun Satish <arjun.sat...@gmail.com> > > > wrote: > > > > > >> Thanks, Konstantine. > > >> > > >> Updated the KIP with the restrictions around log4j and added > references > > to > > >> similar KIPs. > > >> > > >> Best, > > >> > > >> On Mon, Aug 19, 2019 at 3:20 PM Konstantine Karantasis < > > >> konstant...@confluent.io> wrote: > > >> > > >> > Thanks Arjun, the example is useful! > > >> > > > >> > My point when I mentioned the restrictions around log4j is that this > > is > > >> > information is significant and IMO needs to be included in the KIP. > > >> > > > >> > Speaking of its relevance to KIP-412, I think a reference would be > > nice > > >> > too. > > >> > > > >> > Konstantine > > >> > > > >> > > > >> > > > >> > On Thu, Aug 15, 2019 at 4:00 PM Arjun Satish < > arjun.sat...@gmail.com> > > >> > wrote: > > >> > > > >> > > Hey Konstantine, > > >> > > > > >> > > Thanks for the feedback. > > >> > > > > >> > > re: the use of log4j, yes, the proposed changes will only work if > > >> log4j > > >> > is > > >> > > available in runtime. We will not add the mBean if log4j is not > > >> available > > >> > > in classpath. If we change from log4j 1 to 2, that would involve > > >> another > > >> > > KIP, and it would need to update the changes proposed in this KIP > > and > > >> > > others (KIP-412, for instance). > > >> > > > > >> > > re: use of Object types, I've changed it from Boolean to the > > primitive > > >> > type > > >> > > for setLogLevel. We are changing the signature of the old method > > this > > >> > way, > > >> > > but since it never returned null, this should be fine. > > >> > > > > >> > > re: example usage, I've added some screenshot on how this feature > > >> would > > >> > be > > >> > > used with jconsole. > > >> > > > > >> > > Hope this works! > > >> > > > > >> > > Thanks very much, > > >> > > Arjun > > >> > > > > >> > > On Wed, Aug 14, 2019 at 6:42 AM Konstantine Karantasis < > > >> > > konstant...@confluent.io> wrote: > > >> > > > > >> > > > And one thing I forgot is also related to Chris's comment > above. I > > >> > agree > > >> > > > that an example on how a user is expected to set the log level > > (for > > >> > > > instance to DEBUG) would be nice, even if it's showing only one > > out > > >> of > > >> > > the > > >> > > > many possible ways to achieve that. > > >> > > > > > >> > > > - Konstantine > > >> > > > > > >> > > > On Wed, Aug 14, 2019 at 4:38 PM Konstantine Karantasis < > > >> > > > konstant...@confluent.io> wrote: > > >> > > > > > >> > > > > > > >> > > > > Thanks Arjun for tackling the need to support this very useful > > >> > feature. > > >> > > > > > > >> > > > > One thing I noticed while reading the KIP is that I would have > > >> loved > > >> > to > > >> > > > > see more info regarding how this proposal depends on the > > >> underlying > > >> > > > logging > > >> > > > > APIs and implementations. For instance, my understanding is > that > > >> > slf4j > > >> > > > can > > >> > > > > not be leveraged and that the logging framework needs to be > > >> pegged to > > >> > > > log4j > > >> > > > > explicitly (or another logging implementation). Correct me if > > I'm > > >> > > wrong, > > >> > > > > but if such a dependency is introduced I believe it's worth > > >> > mentioning. > > >> > > > > > > >> > > > > Additionally, if the above is correct, there are differences > in > > >> > log4j's > > >> > > > > APIs between version 1 and version 2. In version 2, > > >> Logger#setLevel > > >> > > > method > > >> > > > > has been removed from the Logger interface and in order to set > > the > > >> > log > > >> > > > > level programmatically the Configurator class needs to used, > > >> which as > > >> > > > > stated in the FAQ ( > > >> > > > > > > >> > > > https://logging.apache.org/log4j/2.x/faq.html#reconfig_level_from_code > > >> > > ) > > >> > > > > it's not part of log4j2's public API. Is this a concern? I > > believe > > >> > that > > >> > > > > even if these are implementation specific details for the > > wrappers > > >> > > > > introduced by this KIP (which to a certain extent they are), a > > >> > mention > > >> > > in > > >> > > > > the KIP text and a few references would be useful to > understand > > >> the > > >> > > > changes > > >> > > > > and the dependencies introduced by this proposal. > > >> > > > > > > >> > > > > And a few minor comments: > > >> > > > > - Is there any specific reason that object types were > preferred > > in > > >> > the > > >> > > > > proposed interface compared to primitive types? My > understanding > > >> is > > >> > > that > > >> > > > > `null` is not expected as a return value. > > >> > > > > - Related to the above, I think it'd be nice for the javadoc > to > > >> > mention > > >> > > > > when a parameter is not expected to be `null` with an > > appropriate > > >> > > comment > > >> > > > > (e.g. foo bar etc; may not be null) > > >> > > > > > > >> > > > > Cheers, > > >> > > > > Konstantine > > >> > > > > > > >> > > > > On Tue, Aug 6, 2019 at 9:34 AM Cyrus Vafadari < > > cy...@confluent.io > > >> > > > >> > > > wrote: > > >> > > > > > > >> > > > >> This looks like a useful feature, the strategy makes sense, > and > > >> the > > >> > > KIP > > >> > > > is > > >> > > > >> thorough and nicely written. Thanks! > > >> > > > >> > > >> > > > >> Cyrus > > >> > > > >> > > >> > > > >> On Thu, Aug 1, 2019, 12:40 PM Chris Egerton < > > chr...@confluent.io > > >> > > > >> > > > wrote: > > >> > > > >> > > >> > > > >> > Thanks Arjun! Looks good to me. > > >> > > > >> > > > >> > > > >> > On Thu, Aug 1, 2019 at 12:33 PM Arjun Satish < > > >> > > arjun.sat...@gmail.com> > > >> > > > >> > wrote: > > >> > > > >> > > > >> > > > >> > > Thanks for the feedback, Chris! > > >> > > > >> > > > > >> > > > >> > > Yes, the example is pretty much how Connect will use the > > new > > >> > > > feature. > > >> > > > >> > > Tweaked the section to make this more clear. > > >> > > > >> > > > > >> > > > >> > > Best, > > >> > > > >> > > > > >> > > > >> > > On Fri, Jul 26, 2019 at 11:52 AM Chris Egerton < > > >> > > chr...@confluent.io > > >> > > > > > > >> > > > >> > > wrote: > > >> > > > >> > > > > >> > > > >> > > > Hi Arjun, > > >> > > > >> > > > > > >> > > > >> > > > This looks great. The changes to public interface are > > >> pretty > > >> > > small > > >> > > > >> and > > >> > > > >> > > > moving the Log4jController class into the clients > package > > >> > seems > > >> > > > like > > >> > > > >> > the > > >> > > > >> > > > right way to go. One question I have--it looks like the > > >> > purpose > > >> > > of > > >> > > > >> this > > >> > > > >> > > KIP > > >> > > > >> > > > is to enable dynamic setting of log levels in the > Connect > > >> > > > framework, > > >> > > > >> > but > > >> > > > >> > > > it's not clear how the Connect framework will use that > > new > > >> > > > utility. > > >> > > > >> Is > > >> > > > >> > > the > > >> > > > >> > > > "Example Usage" section (which involves invoking the > > >> utility > > >> > > with > > >> > > > a > > >> > > > >> > > > namespace of "kafka.connect") actually meant to be part > > of > > >> the > > >> > > > >> proposed > > >> > > > >> > > > changes to public interface? > > >> > > > >> > > > > > >> > > > >> > > > Cheers, > > >> > > > >> > > > > > >> > > > >> > > > Chris > > >> > > > >> > > > > > >> > > > >> > > > On Mon, Jul 22, 2019 at 11:03 PM Arjun Satish < > > >> > > > >> arjun.sat...@gmail.com> > > >> > > > >> > > > wrote: > > >> > > > >> > > > > > >> > > > >> > > > > Hi everyone. > > >> > > > >> > > > > > > >> > > > >> > > > > I'd like to propose the following KIP to implement > > >> changing > > >> > > log > > >> > > > >> > levels > > >> > > > >> > > on > > >> > > > >> > > > > the fly in Connect workers: > > >> > > > >> > > > > > > >> > > > >> > > > > > > >> > > > >> > > > > > > >> > > > >> > > > > > >> > > > >> > > > > >> > > > >> > > > >> > > > >> > > >> > > > > > >> > > > > >> > > > >> > > > https://cwiki.apache.org/confluence/display/KAFKA/KIP-495%3A+Dynamically+Adjust+Log+Levels+in+Connect > > >> > > > >> > > > > > > >> > > > >> > > > > Would like to hear your thoughts on this. > > >> > > > >> > > > > > > >> > > > >> > > > > Thanks very much, > > >> > > > >> > > > > Arjun > > >> > > > >> > > > > > > >> > > > >> > > > > > >> > > > >> > > > > >> > > > >> > > > >> > > > >> > > >> > > > > > > >> > > > > > >> > > > > >> > > > >> > > > > > >
