> Thanks for the clarification. The proposed behavior sounds reasonable. Can you add a note about the implementation on the client? The client needs to be prepared to handle > a response that doesn't include the versions, as well, since v1 did not.
I have added a note about the implementation in the KIP. In short, when the client receives an unsupported version, it defaults to version 0. As version 0 contains both the ErrorCode and the ApiKeys fields, the client can check the error and in case of UNSUPPORTED_VERSION, it can check the ApiKeys to get the supported versions. If not present, it default to version 0. > Hmm. Like we discussed above, there is a very important difference in the v3 response, which is that the versions will be included even if the client's version was higher than > what the broker supports. We should add a comment about that. Yeah. I think the change that we propose to enhance the handling of unsupported versions of ApiVersionsRequest/Response is orthogonal to the version bump. Concretely, the versions will be included in the ApiVersionsResponse v0 - the request/response used by the broker when failing back - so it is a bit misleading to say that starting from version 3, the broker populate the ApiKeys field with the information about the supported versions of the AVR. I would rather put a note saying: Starting from Apache Kafka 2.4, ApiKeys field is populated with the supported versions of the ApiVersionsRequest when an UNSUPPORTED_VERSION error is returned. Would this work for you? > Agreed. This is a good use-case for INVALID_REQUEST. We should add a comment that this is now a valid error. I have documented the error in the Public Interfaces section. Best, David On Thu, Sep 19, 2019 at 10:52 PM Colin McCabe <cmcc...@apache.org> wrote: > On Wed, Sep 18, 2019, at 23:44, David Jacot wrote: > > Hi Colin, > > > > Thank you for your feedback! Please find my comments/answers below: > > > > *> Nitpick: in the intro paragraph, "Operators of Apache Kafka clusters > > have literally no information about the clients connected to their > > clusters" seems a bit too strong. We have some information, right? For > > example, the client ID, where clients are connecting from, etc. Maybe it > > would be clearer to say "very little information about the type of client > > software..."* > > > > That's fair. I will update it. > > > > *> Instead of ClientName and ClientVersion, how about ClientSoftwareName > > and ClientSoftwareVersion? This might make it clearer what the fields > are > > for. I can see people getting confused about the difference between > > ClientName and ClientId, which sound pretty similar. Adding "Software" > to > > the field name makes it much clearer what the difference is between these > > fields.* > > > > Good point. I like your suggestion. I will update it. > > > > *> In the "ApiVersions Request/Response Handling" section, it seems like > > there is some out-of-date text. Specifically, it says "we propose to add > > the supported version of the ApiVersionsRequest in the response sent back > > to the client alongside the error...". But on the other hand, elsewhere > in > > the KIP, we say "ApiVersionsResponse is bumped to version 3 but does not > > have any changes in the schema" Based on the offline discussion we had, > > the correct text is the latter (we're not changing ApiVersionsRerponse). > > So we should remove the text about adding stuff to ApiVersionsResponse.* > > > > Sorry for the confusion. I think my usage of the word "add" is not > > appropriate here. The ApiVersionsResponse won't change as you said. My > > proposal is to provide the supported versions of the ApiVersionsRequest > in > > the response by populating the existing `api_versions` field. In the > > current version, when an error occurs, the `api_versions` is empty in the > > response. Providing it enables the client to re-send the latest version > > supported by the broker instead of defaulting to zero. I will update the > > KIP to make this clearer. > > Thanks for the clarification. The proposed behavior sounds reasonable. > Can you add a note about the implementation on the client? The client > needs to be prepared to handle a response that doesn't include the > versions, as well, since v1 did not. > > > > > *> In a similar vein, the comment " // Version 3 is similar to version > 2" > > should be " // Version 3 is identical to version 2" or something like > > that. Although I guess technically things which are identical are also > > similar, the current phrasing could be misleading.* > > > > Good point. I will use `Version 3 is the same as version 2.` which is the > > statement already used in other requests/responses. > > Hmm. Like we discussed above, there is a very important difference in the > v3 response, which is that the versions will be included even if the > client's version was higher than what the broker supports. We should add a > comment about that. > > > > > > > *> Now that KIP-482 has been accepted, I think there are a few things > that > > are worth clarifying in the KIP. Firstly, ApiVersionsRequest v3 should > be > > a "flexible version". Mainly, that means its request header will support > > optional tagged fields. However, ApiVersionsResponse v3 will *not* > support > > optional tagged fields in its response header. This is necessary > because-- > > as you said-- the broker must look at a fixed offset to find the error > > code, regardless of the response version.* > > Right. I have put it because I thought your PR would do it. I will update > > this. By the way, it means that the request/response must be updated to > the > > generated ones, isn't it? AVR is still using the old mechanism. > > Yeah, I think we should move to the new mechanism. It should be very easy > for the request. The response may be slightly more difficult, but probably > not that much more. > > > > > > > > > *> I think we should force client software names and versions to follow a > > regular expression and disconnect if they do not. This will prevent > issues > > when using these strings in metrics names. Probably we want something > > like:> [\.\-A-Za-z0-9]?[\.\-A-Za-z0-9 ]*[\.\-A-Za-z0-9]?> Notice this > does > > _not* include underscores, since they get converted to dots in JMX, > causing > > ambiguity. It also doesn't allow the first or last character to be a > > space.* > > > > I do agree and I have already put something along those lines in the > > proposal. See the "Validation" chapter. I have proposed to use a more > > restrictive validation which does not allow white spaces. I think spaces > > wouldn't be used in software name nor version. Is it OK for you if we > stick > > to the more restrictive one? Thank your letting me know about the > > underscores. I have missed this. > > Yeah, the one you proposed sounds fine. > > > > > Regarding disconnecting when the validation fails, this is what I have > > proposed as well. Magnus has brought a good point though. Using an > explicit > > error like `INVALID_REQUEST` may be better. In this case, the client > would > > have to disconnect when it happens. I will update the KIP to reflect > this. > > Agreed. This is a good use-case for INVALID_REQUEST. We should add a > comment that this is now a valid error. > > best, > Colin > > > > > > Best, > > David > > > > > > On Wed, Sep 18, 2019 at 9:46 PM Colin McCabe <cmcc...@apache.org> wrote: > > > > > That's fair. We could use the existing error code in the response and > > > pass back something like INVALID_REQUEST. > > > > > > I'm not sure if we want to add an error string field just for this > > > (although they're a good idea in general...) > > > > > > best, > > > Colin > > > > > > On Wed, Sep 18, 2019, at 12:31, Magnus Edenhill wrote: > > > > > I think we should force client software names and versions to > follow a > > > > regular expression and disconnect if they do not. > > > > > > > > Disconnecting is not really a great error propagation method since it > > > > leaves the client oblivious to what went wrong. > > > > Instead suggest we return an ApiVersionResponse with an error code > and a > > > > human-readable error message field. > > > > > > > > > > > > > > > > Den ons 18 sep. 2019 kl 20:05 skrev Colin McCabe <cmcc...@apache.org > >: > > > > > > > > > Hi David, > > > > > > > > > > Thanks for the KIP! > > > > > > > > > > Nitpick: in the intro paragraph, "Operators of Apache Kafka > clusters > > > have > > > > > literally no information about the clients connected to their > clusters" > > > > > seems a bit too strong. We have some information, right? For > > > example, the > > > > > client ID, where clients are connecting from, etc. Maybe it would > be > > > > > clearer to say "very little information about the type of client > > > > > software..." > > > > > > > > > > Instead of ClientName and ClientVersion, how about > ClientSoftwareName > > > and > > > > > ClientSoftwareVersion? This might make it clearer what the fields > are > > > > > for. I can see people getting confused about the difference > between > > > > > ClientName and ClientId, which sound pretty similar. Adding > > > "Software" to > > > > > the field name makes it much clearer what the difference is between > > > these > > > > > fields. > > > > > > > > > > In the "ApiVersions Request/Response Handling" section, it seems > like > > > > > there is some out-of-date text. Specifically, it says "we propose > to > > > add > > > > > the supported version of the ApiVersionsRequest in the response > sent > > > back > > > > > to the client alongside the error...". But on the other hand, > > > elsewhere in > > > > > the KIP, we say "ApiVersionsResponse is bumped to version 3 but > does > > > not > > > > > have any changes in the schema" Based on the offline discussion we > > > had, > > > > > the correct text is the latter (we're not changing > > > ApiVersionsRerponse). > > > > > So we should remove the text about adding stuff to > ApiVersionsResponse. > > > > > > > > > > In a similar vein, the comment " // Version 3 is similar to > version 2" > > > > > should be " // Version 3 is identical to version 2" or something > like > > > > > that. Although I guess technically things which are identical are > also > > > > > similar, the current phrasing could be misleading. > > > > > > > > > > Now that KIP-482 has been accepted, I think there are a few things > that > > > > > are worth clarifying in the KIP. Firstly, ApiVersionsRequest v3 > > > should be > > > > > a "flexible version". Mainly, that means its request header will > > > support > > > > > optional tagged fields. However, ApiVersionsResponse v3 will *not* > > > support > > > > > optional tagged fields in its response header. This is necessary > > > because-- > > > > > as you said-- the broker must look at a fixed offset to find the > error > > > > > code, regardless of the response version. > > > > > > > > > > I think we should force client software names and versions to > follow a > > > > > regular expression and disconnect if they do not. This will > prevent > > > issues > > > > > when using these strings in metrics names. Probably we want > something > > > like: > > > > > > > > > > [\.\-A-Za-z0-9]?[\.\-A-Za-z0-9 ]*[\.\-A-Za-z0-9]? > > > > > > > > > > Notice this does _not* include underscores, since they get > converted to > > > > > dots in JMX, causing ambiguity. It also doesn't allow the first or > > > last > > > > > character to be a space. > > > > > > > > > > best, > > > > > Colin > > > > > > > > > > > > > > > On Mon, Sep 16, 2019, at 04:39, Mickael Maison wrote: > > > > > > +1 (non binding) > > > > > > Thanks for the KIP! > > > > > > > > > > > > On Mon, Sep 16, 2019 at 12:07 PM David Jacot < > dja...@confluent.io> > > > > > wrote: > > > > > > > > > > > > > > Hi all, > > > > > > > > > > > > > > I would like to start a vote on KIP-511: > > > > > > > > > > > > > > > > https://cwiki.apache.org/confluence/display/KAFKA/KIP-511%3A+Collect+and+Expose+Client%27s+Name+and+Version+in+the+Brokers > > > > > > > > > > > > > > Best, > > > > > > > David > > > > > > > > > > > > > > > > > > > > >
