Hi, Paul. I concur with the others, and I like the new approach that avoids a new configuration, especially because it does not change the behavior for anyone already using `producer.client.id` and/or `consumer.client.id`. I did leave a few comments on the PR. Perhaps the biggest one is whether the producer used for the sink task error reporter (for dead letter queue) should be `connector-producer-<sink-task-id>`, and whether that is distinct enough from source tasks, which will be of the form `connector-producer-<source-task-id>`. Maybe it is fine. (The other comments were minor.)
Best regards, Randall On Mon, Jan 7, 2019 at 1:19 PM Paul Davidson <pdavid...@salesforce.com> wrote: > Thanks all. I've submitted a new PR with a possible implementation: > https://github.com/apache/kafka/pull/6097. Note I did not include the > group > ID as part of the default client ID, mainly to avoid the connector name > appearing twice by default. As noted in the original Jira ( > https://issues.apache.org/jira/browse/KAFKA-5061), leaving out the group > ID > could lead to naming conflicts if multiple clusters run the same Kafka > cluster. This would probably not be a problem for many (including us) as > metrics exporters can usually be configured to include a cluster ID and > guarantee uniqueness. Will be interested to hear your thoughts on this. > > Paul > > > > On Mon, Jan 7, 2019 at 10:27 AM Ryanne Dolan <ryannedo...@gmail.com> > wrote: > > > I'd also prefer to avoid the new configuration property if possible. > Seems > > like a lighter touch without it. > > > > Ryanne > > > > On Sun, Jan 6, 2019 at 7:25 PM Paul Davidson <pdavid...@salesforce.com> > > wrote: > > > > > Hi Konstantine, > > > > > > Thanks for your feedback! I think my reply to Ewen covers most of your > > > points, and I mostly agree. If there is general agreement that > changing > > > the default behavior is preferable to a config change I will update my > PR > > > to use that approach. > > > > > > Paul > > > > > > On Fri, Jan 4, 2019 at 5:55 PM Konstantine Karantasis < > > > konstant...@confluent.io> wrote: > > > > > > > Hi Paul. > > > > > > > > I second Ewen and I intended to give similar feedback: > > > > > > > > 1) Can we avoid a config altogether? > > > > 2) If we prefer to add a config anyways, can we use a set of allowed > > > values > > > > instead of a boolean, even if initially these values are only two? As > > the > > > > discussion on Jira highlights, there is a potential for more naming > > > > conventions in the future, even if now the extra functionality > doesn't > > > seem > > > > essential. It's not optimal to have to deprecate a config instead of > > just > > > > extending its set of values. > > > > 3) I agree, the config name sounds too general. How about > > > > "client.ids.naming.policy" or "client.ids.naming" if you want two > more > > > > options? > > > > > > > > Konstantine > > > > > > > > On Fri, Jan 4, 2019 at 7:38 AM Ewen Cheslack-Postava < > > e...@confluent.io> > > > > wrote: > > > > > > > > > Hi Paul, > > > > > > > > > > Thanks for the KIP. A few comments. > > > > > > > > > > To me, biggest question here is if we can fix this behavior without > > > > adding > > > > > a config. In particular, today, we don't even set the client.id > for > > > the > > > > > producer and consumer at all, right? The *only* way it is set is if > > you > > > > > include an override in the worker config, but in that case you need > > to > > > be > > > > > explicitly opting in with a `producer.` or `consumer.` prefix, i.e. > > the > > > > > settings are `producer.client.id` and `consumer.client.id`. > > > Otherwise, I > > > > > think we're getting the default behavior where we generate unique, > > > > > per-process IDs, i.e. via this logic > > > > > > > > > > > > > > > > > > > > https://github.com/apache/kafka/blob/trunk/clients/src/main/java/org/apache/kafka/clients/consumer/KafkaConsumer.java#L662-L664 > > > > > > > > > > If that's the case, would it maybe be possible to compatibly change > > the > > > > > default to use task IDs in the client ID, but only if we don't see > an > > > > > existing override from the worker config? This would only change > the > > > > > behavior when someone is using the default, but since the default > > would > > > > > just use what is effectively a random ID that is useless for > > monitoring > > > > > metrics, presumably this wouldn't affect any existing users. I > think > > > that > > > > > would avoid having to introduce the config, give better out of the > > box > > > > > behavior, and still be a safe, compatible change to make. > > > > > > > > > > > > > > > Other than that, just two minor comments. On the config naming, not > > > sure > > > > > about a better name, but I think the config name could be a bit > > clearer > > > > if > > > > > we need to have it. Maybe something including "task" like > > > > > "task.based.client.ids" or something like that (or change the type > to > > > be > > > > an > > > > > enum and make it something like task.client.ids=[default|task] and > > > leave > > > > it > > > > > open for extension in the future if needed). > > > > > > > > > > Finally, you have this: > > > > > > > > > > *"Allow overriding client.id <http://client.id/> on a > per-connector > > > > > basis"* > > > > > > > > > > > > This is a much more complex change, and would require individual > > > > > > connectors to be updated to support the change. In contrast, the > > > > proposed > > > > > > approach would immediately allow detailed consumer/producer > > > monitoring > > > > > for > > > > > > all existing connectors. > > > > > > > > > > > > > > > > I don't think this is quite accurate. I think the reason to reject > is > > > > that > > > > > for your particular requirement for metrics, it simply doesn't give > > > > enough > > > > > granularity (there's only one value per entire connector), but it > > > doesn't > > > > > require any changes to connectors. The framework allocates all of > > these > > > > and > > > > > there are already framework-defined config values that all > connectors > > > > share > > > > > (some for only sinks or sources), so the framework can handle all > of > > > this > > > > > without changes to connectors. Further, with connector-specific > > > > overrides, > > > > > you could get task-specific values if interpolation were supported > in > > > the > > > > > config value (as we now do with managed secrets). For example, it > > could > > > > > support something like client.id=connector-${taskId} and the task > ID > > > > would > > > > > be substituted automatically into the string. > > > > > > > > > > I don't necessarily like that solution (seems complicated and not a > > > great > > > > > user experience), but it could work. > > > > > > > > > > -Ewen > > > > > > > > > > > > > > > > > > > > > > > > > On Thu, Dec 20, 2018 at 5:05 PM Paul Davidson < > > > pdavid...@salesforce.com> > > > > > wrote: > > > > > > > > > > > Hi everyone, > > > > > > > > > > > > I would like to start a discussion around the following KIP: > > > > > > * > > > > > > > > > > > > > > > > > > > > > https://cwiki.apache.org/confluence/display/KAFKA/KIP-411%3A+Add+option+to+make+Kafka+Connect+task+client+ID+values+unique > > > > > > < > > > > > > > > > > > > > > > > > > > > > https://cwiki.apache.org/confluence/display/KAFKA/KIP-411%3A+Add+option+to+make+Kafka+Connect+task+client+ID+values+unique > > > > > > >* > > > > > > > > > > > > This proposes a small change to allow Kafka Connect the option to > > > > > > auto-generate unique client IDs for each task. This enables > > granular > > > > > > monitoring of the producer / consumer client in each task. > > > > > > > > > > > > Feedback is appreciated, thanks in advance! > > > > > > > > > > > > Paul > > > > > > > > > > > > > > > > > > > > >
