Hi Andrew,
Thanks for the KIP. I have one question.

AM1: Since this KIP adds --by-duration to kafka-share-groups.sh while share
groups also accept share.auto.offset.reset=by_duration:… as a group config. I
can find share.auto.offset.reset=by_duration:PT3H (config,
kafka-configs.sh): an init-time policy the broker applies only when no SPSO
exists. I can set the config while there are active members but the config
update is not honoured. It's only honoured when a new group is specified.
Can you please clarify the expected behaviour when using
`kafka-share-groups.sh --by-duration` reset option?

Regards,
Apoorv Mittal


On Thu, Aug 13, 2026 at 1:21 PM Andrew Schofield <[email protected]>
wrote:

> Hi Aditya,
> Thanks for your reply. I'm not dismissing your concern, but I'm not
> entirely happy with a way forward yet.
>
> AK1: I wonder whether we could use a pipe here to avoid the intermediate
> file. After all, export uses stdout as the destination, so if we could use
> stdin as the source. For example, I've used '-' with --from-file below to
> indicate stdin. It could be --from-stdin or similar too.
>
> $ bin/kafka-consumer-groups.sh --bootstrap-server localhost:9092
> --export-offsets --group CG1 --topic T1 --to-current |
> bin/kafka-consumer-groups.sh --bootstrap-server localhost:9092
> --reset-offsets --group CG2 --from-file - --topic T1 --execute
>
> This would still allow resetting between different group types, which is
> actually the use case that got me into this, taking offsets from a consumer
> group to initialise a share group.
>
> Thanks,
> Andrew
>
> On 2026/08/12 18:12:08 Aditya Kousik wrote:
> > Hi Andrew.
> >
> > Thanks for addressing AK2-4: lgtm
> >
> > On AK1: I’m thinking from a user experience standpoint. The teams
> managing the kafka clusters and the ones using the clients diverge - but
> the Admin role may be exposed via self-service typically via API requests
> and a K8S CRD. I’m thinking of the latter case where Kafka SRE allow a thin
> wrapper for client teams to manage their own groups and topics and also
> isolating the environment where the scripts can be run. In such a case, it
> becomes cumbersome to fetch/write the csv file from a Pod that runs the
> migration ; it may not have any volume to persist it or running kubectl to
> scp the file is usually denied. So in such a case, the migrate offsets
> script could just be a single command that fetches offsets from one group
> and applies it to another (or several) without the middleman CSV.
> >
> > Thanks,
> > Aditya
> >
> > > On Aug 12, 2026, at 06:39, Andrew Schofield <[email protected]>
> wrote:
> > >
> > > Hi Aditya,
> > > Thanks for your reply. Sorry for the delay.
> > >
> > > AK1: One of the advantages of the file is that it's possible to export
> from one group type and import into another. You can also export offsets
> for multiple groups, and this changes the file format (4 columns instead of
> 3). Sadly, this is all existing behaviour that I only learnt by trial and
> error. Personally, I think that export-import is a relatively infrequent
> operation, and the KIP makes enough of a step forward. I wonder if you
> could elaborate a bit more on your idea. Feel free to push back if you
> disagree.
> > >
> > > AK2: The KIP does not mention this and I wasn't quite sure where to
> add it. You are correct. Should we improve the javadoc for the admin client?
> > >
> > > AK3: Yes, Admin is better. KIP updated.
> > >
> > > AK4: The DescribeGroupsGeneric RPC is intended as a way of getting a
> minimal set of group information for groups of any group types. It is used
> in Admin.describeGroups and this would be the initial Admin call from
> bin/kafka-XXXX-groups.sh when exporting or resetting offsets. It gives a
> way to check the group type and existence unambiguously. By contrast, if
> you use Admin.describeXXXGroups() for an existing group ID of the wrong
> type, the error code is GROUP_ID_NOT_FOUND.
> > >
> > > AK4.1: I didn't include the members intentionally. This RPC is
> agnostic to group type. I think the common subset of group attributes is
> group ID, group type, group state and coordinator. If you dig deeper, you
> find that the group types differ. For example, assignments in consumer
> groups and share groups are partitions, while in streams groups they are
> tasks. I expect we'll add connect groups one day, and until we have the
> KIP, I wouldn't want to assume the details of membership.
> > >
> > > AK4.2: This is following existing convention I'm afraid and I want to
> be consistent. If the broker does not send back a list of ACL operations,
> then this will be null. This is different than an empty set which means
> "you are not authorized", as opposed to "you don't know anything about
> authorized operations".
> > >
> > > Thanks,
> > > Andrew
> > >
> > >> On 2026/07/27 05:45:27 Aditya Kousik wrote:
> > >> A small correction.
> > >>
> > >> I meant to add a statement before the numbered comments.
> > >>
> > >> “Since the KIP motivation revolves around usability of the script,
> would you be willing to consider the following:”
> > >>
> > >> Also, typo on AK2. “…As long as there aren’t any active members in
> the group subscription”.
> > >>
> > >> -Aditya
> > >>
> > >>>> On Jul 25, 2026, at 23:53, Aditya Kousik <[email protected]>
> wrote:
> > >>>
> > >>> Hi Andrew,
> > >>>
> > >>> Thanks for this KIP, I’ve personally faced more than one scenario
> where I’ve had to use the admin client to update the offsets.
> > >>>
> > >>> The main uses cases teams face ( and TIL about the various options
> the script offers):
> > >>> Teams migrating to a new consumer group from an existing one but
> don’t want to commit to either (or duration for newer clients) auto offset
> reset policy on the new group.
> > >>> Teams want to seek past certain offsets in partitions because they
> have poison pill messages/corrupted deserialization story and they have no
> control over the data. In this case, they actually want latest offset
> strategy but most don’t even realise that it applies to streams out of
> range and are surprised/frustrated that the property doesn’t move their
> position in the queues. (I see we can do this already in the script today)
> > >>> When I migrated our clients from 2.x directly to 4.x, out of
> abundance of caution, I wrote an hourly cron job that dumps the snapshot of
> group_id,topic,partition,offset into csv files so that if I had to revert a
> few affected partitions I can trim the csv down to those partitions and
> alter them with the admin client.
> > >>>
> > >>> AK1. A single command line with args —source-group-id
> —target-group-id —export so that the intermediate CSV file can be skipped.
> > >>> AK2. Minor: The KIP does not mention the alter/list consumer group
> offsets Admin call. alter only works as long as there are active members in
> the group subscription. I was bitten by this in the past; the documentation
> was sparse around that expectation but it does make sense.
> > >>> AK3. KIP references AdminClient but I believe this is deprecated in
> favour of just Admin which is the interface that has the methods?
> > >>> AK4. How do we leverage the describe groups RPC when using the
> script/admin client? I’m trying to imagine flows other than the export
> group offsets.
> > >>> AK4.1 Can we list active members/offset state in a given group in
> addition to the GroupState? This has helped identify zombie processes to
> hunt down. The broker’s view would be useful here. (related to AK2)
> > >>> AK4.2 Could we return empty set of acl operations instead of null?
> > >>>
> > >>> Best,
> > >>> Aditya
> > >>>
> > >>>
> > >>>> On 2026/06/17 06:34:11 Andrew Schofield wrote:
> > >>>> Hi,
> > >>>> I’d like to start discussion on a small KIP to improve the
> usability of the command-line tools for resetting group offsets. When
> implementing KIP-1323 and reviewing the PRs, it became clear that the
> usability of this area of Kafka isn’t great. This KIP improves the
> usability of these tools.
> > >>>>
> > >>>>
> https://cwiki.apache.org/confluence/display/KAFKA/KIP-1359%3A+Improve+usability+of+resetting+group+offsets
> > >>>>
> > >>>> Thanks,
> > >>>> Andrew
> > >>>>
> > >>
> >
>

Reply via email to