Hi Mickael, Thanks for the updated KIP, this is looking really good! I had a couple more questions -
1) Sensor names need to be unique across all groups for a `Metrics` instance. How are we planning to avoid naming clashes (both between different plugins as well as with pre-defined sensors)? 2) Connect has a `ConnectMetrics` wrapper around `Metrics` via which rebalance / worker / connector / task metrics are recorded. Could you please elaborate in the KIP how the plugin metrics for connectors / tasks will inter-operate with this? Another minor point is that the third rejected alternative appears to be an incomplete sentence? Thanks, Yash On Fri, Jan 13, 2023 at 10:56 PM Mickael Maison <mickael.mai...@gmail.com> wrote: > Hi, > > I've updated the KIP based on the feedback. > > Now instead of receiving a Metrics instance, plugins get access to > PluginMetrics that exposes a much smaller API. I've removed the > special handling for connectors and tasks and they must now implement > the Monitorable interface as well to use this feature. Finally the > goal is to allow all plugins (apart from metrics reporters) to use > this feature. I've listed them all (there are over 30 pluggable APIs) > but I've not added the list in the KIP. The reason is that new plugins > could be added in the future and instead I'll focus on adding support > in all the place that instantiate classes. > > Thanks, > Mickael > > On Tue, Jan 10, 2023 at 7:00 PM Mickael Maison <mickael.mai...@gmail.com> > wrote: > > > > Hi Chris/Yash, > > > > Thanks for taking a look and providing feedback. > > > > 1) Yes you're right, when using incompatible version, metrics() would > > trigger NoSuchMethodError. I thought using the context to pass the > > Metrics object would be more idiomatic for Connect but maybe > > implementing Monitorable would be simpler. It would also allow other > > Connect plugins (transformations, converters, etc) to register > > metrics. So I'll make that change. > > > > 2) As mentioned in the rejected alternatives, I considered having a > > PluginMetrics class/interface with a limited API. But since Metrics is > > part of the public API, I thought it would be simpler to reuse it. > > That said you bring interesting points so I took another look today. > > It's true that the Metrics API is pretty complex and most methods are > > useless for plugin authors. I'd expect most use cases only need one > > addMetric and one sensor methods. Rather than subclassing Metrics, I > > think a delegate/forwarding pattern might work well here. A > > PluginMetric class would forward its method to the Metrics instance > > and could perform some basic validations such as only letting plugins > > delete metrics they created, or automatically injecting tags with the > > class name for example. > > > > 3) Between the clients, brokers, streams and connect, Kafka has quite > > a lot! In practice I think registering metrics should be beneficial > > for all plugins, I think the only exception would be metrics reporters > > (which are instantiated before the Metrics object). I'll try to build > > a list of all plugin types and add that to the KIP. > > > > Thanks, > > Mickael > > > > > > > > On Tue, Dec 27, 2022 at 4:54 PM Chris Egerton <chr...@aiven.io.invalid> > wrote: > > > > > > Hi Yash, > > > > > > Yes, a default no-op is exactly what I had in mind should the > Connector and > > > Task classes implement the Monitorable interface. > > > > > > Cheers, > > > > > > Chris > > > > > > On Tue, Dec 20, 2022 at 2:46 AM Yash Mayya <yash.ma...@gmail.com> > wrote: > > > > > > > Hi Mickael, > > > > > > > > Thanks for creating this KIP, this will be a super useful feature to > > > > enhance existing connectors in the Kafka Connect ecosystem. > > > > > > > > I have some similar concerns to the ones that Chris has outlined > above, > > > > especially with regard to directly exposing Connect's Metrics object > to > > > > plugins. I believe it would be a lot friendlier to developers if we > instead > > > > exposed wrapper methods in the context classes - such as one for > > > > registering a new metric, one for recording metric values and so on. > This > > > > would also have the added benefit of minimizing the surface area for > > > > potential misuse by custom plugins. > > > > > > > > > for connectors and tasks they should handle the > > > > > metrics() method returning null when deployed on > > > > > an older runtime. > > > > > > > > I believe this won't be the case, and instead they'll need to handle > a > > > > `NoSuchMethodError` right? This is similar to previous KIPs that > added > > > > methods to connector context classes and will arise due to an > > > > incompatibility between the `connect-api` dependency that a plugin > will be > > > > compiled against versus what it will actually get at runtime. > > > > > > > > Hi Chris, > > > > > > > > > WDYT about having the Connector and Task classes > > > > > implement the Monitorable interface, both for > > > > > consistency's sake, and to prevent classloading > > > > > headaches? > > > > > > > > Are you suggesting that the framework should configure connectors / > tasks > > > > with a Metrics instance during their startup rather than the > connector / > > > > task asking the framework to provide one? In this case, I'm guessing > you're > > > > envisioning a default no-op implementation for the metrics > configuration > > > > method rather than the framework having to handle the case where the > > > > connector was compiled against an older version of Connect right? > > > > > > > > Thanks, > > > > Yash > > > > > > > > On Wed, Nov 30, 2022 at 1:38 AM Chris Egerton > <chr...@aiven.io.invalid> > > > > wrote: > > > > > > > > > Hi Mickael, > > > > > > > > > > Thanks for the KIP! This seems especially useful to reduce the > > > > > implementation cost and divergence in behavior for connectors that > choose > > > > > to publish their own metrics. > > > > > > > > > > My initial thoughts: > > > > > > > > > > 1. Are you certain that the default implementation of the "metrics" > > > > method > > > > > for the various connector/task context classes will be used on > older > > > > > versions of the Connect runtime? My understanding was that a > > > > > NoSuchMethodError (or some similar classloading exception) would be > > > > thrown > > > > > in that case. If that turns out to be true, WDYT about having the > > > > Connector > > > > > and Task classes implement the Monitorable interface, both for > > > > > consistency's sake, and to prevent classloading headaches? > > > > > > > > > > 2. Although I agree that administrators should be careful about > which > > > > > plugins they run on their clients, Connect clusters, etc., I > wonder if > > > > > there might still be value in wrapping the Metrics class behind a > new > > > > > interface, for a few reasons: > > > > > > > > > > a. Developers and administrators may still make mistakes, and if > we can > > > > > reduce the blast radius by preventing plugins from, e.g., closing > the > > > > > Metrics instance we give them, it may be worth it. This could also > be > > > > > accomplished by forbidding plugins from invoking these methods, and > > > > giving > > > > > them a subclass of Metrics that throws > UnsupportedOperationException from > > > > > these methods. > > > > > > > > > > b. If we don't know of any reasonable use cases for closing the > > > > instance, > > > > > adding new reporters, removing metrics, etc., it can make the API > cleaner > > > > > and easier for developers to grok if they don't even have the > option to > > > > do > > > > > any of those things. > > > > > > > > > > c. Interoperability between plugins that implement Monitorable > and > > > > their > > > > > runtime becomes complicated. For example, a connector may be built > > > > against > > > > > a version of Kafka that introduces new methods for the Metrics > class, > > > > which > > > > > introduces risks of incompatibility if its developer chooses to > take > > > > > advantage of these methods without realizing that they will not be > > > > > available on Connect runtimes built against an older version of > Kafka. > > > > With > > > > > a wrapper interface, we at least have a chance to isolate these > issues so > > > > > that the Metrics class can be expanded without adding footguns for > > > > plugins > > > > > that implement Monitorable, and to call out potential compatibility > > > > > problems in documentation more clearly if/when we do expand the > wrapper > > > > > interface. > > > > > > > > > > 3. It'd be nice to see a list of exactly which plugins will be > able to > > > > take > > > > > advantage of the new Monitorable interface. > > > > > > > > > > Looking forward to your thoughts! > > > > > > > > > > Cheers, > > > > > > > > > > Chris > > > > > > > > > > On Mon, Nov 7, 2022 at 11:42 AM Mickael Maison < > mickael.mai...@gmail.com > > > > > > > > > > wrote: > > > > > > > > > > > Hi, > > > > > > > > > > > > I have opened KIP-877 to make it easy for plugins and connectors > to > > > > > > register their own metrics: > > > > > > > > > > > > > > > > > > > > > > https://eu01.z.antigena.com/l/9lWv8kbU9CKs2LajwgfKF~yMNQVM7rWRxYmYVNrHU_2nQbisTiXYZdowNfQ-NcgF1uai2lk-sv6hJASnbdr_gqVwyVae_~y-~oq5yQFgO_-IHD3UGDn3lsIyauAG2tG6giPJH-9yCYg3Hwe26sm7nep258qB6SNXRwpaVxbU3SaVTybfLQVvTn_uUlHKMhmVnpnc1dUnusK6x4j8JPPQQ1Ce~rrg-nsSLouHHMf0ewmpsFNy4BcbMaqHd4Y > > > > > > > > > > > > Let me know if you have any feedback or suggestions. > > > > > > > > > > > > Thanks, > > > > > > Mickael > > > > > > > > > > > > > > > > > > > > > >
