Hi Andrew,
Thanks for the feedback. I've updated the KIP.

AS1) Fixed.

AS2) Yes, setting group.share.max.release.backoff.ms to 0 is meant to turn
off redelivery delays for all share groups, so I've made it a dynamic
broker config with a cluster-wide update mode.

AS3) Done.

AS4) Agreed, that's simpler to use. I've replaced
share.release.backoff.exponential.enable
with share.release.backoff.mode, which takes the suggested values.

Thanks,
Abhinav

On Fri, Nov 13, 2026 at 8:10 PM Andrew Schofield via dev <
[email protected]> wrote:

> Hi Abhinav,
> Thanks for the KIP. This is a nice improvement which fits in well. I
> see that the way you have done the broker configurations doesn't match
> what we've done before (there's no max-max), but I think that what
> you've done is fine and less complicated.
>
> A few comments from an initial review.
>
> AS1) Typo "ype" instead of "Type" in the group config table.
>
> AS2) If the broker configuration group.share.max.release.backoff.ms
> is intended as a way to disable redelivery delays, please call it out as
> a dynamic broker config so the code is written accordingly.
>
> AS3) In the section on lag, you say "so they count as lag". I suggest
> that "remain in-flight" would be more elegant.
>
> AS4) I wonder whether another way to do the configuration might be
> to replace share.release.backoff.exponential.enable with
> share.release.backoff.mode with three values:
>
>   off (default)
>   fixed - Every wait is share.release.backoff.ms without jitter
>   exponential - what you've written for the exponential case
>
> The reason I suggest this is that we could have a non-zero
> share.release.backoff.ms value (say 2000ms) which acts as a suggested
> starting point for a backoff delay. The default exponential backoff
> delay would then be 2000+4000+8000+16000 = 30000ms, and all a user
> would have to set is `share.release.backoff.mode=exponential`. Wdyt?
>
> Thanks,
> Andrew
>
>
> On 2026-11-12 06:57 UTC Abhinav Dixit via dev wrote:
> > Hi all,
> > I'd like to start a discussion on KIP-1395, which allows a share group to
> > introduce a delay and optional exponential backoff capability within its
> > redelivery model.
> >
> > Today, a released or expired record becomes available for redelivery as
> > soon as its lock clears, there is no way to slow down or space out
> retries
> > per share group. This KIP adds a per-group config for how long a record
> is
> > unavailable for fetching after an explicit release or record expiry. It
> > also adds an optional exponential-backoff mode that increases this delay
> as
> > a record's delivery count grows, up to a bounded maximum.
> >
> >
> > Looking forward to your feedback.
> >
> > KIP link -
> >
> https://urldefense.com/v3/__https://cwiki.apache.org/confluence/spaces/KAFKA/pages/462817066/KIP-1395*Redelivery*delay*and*backoff*for*share*groups__;KysrKysrKw!!Ayb5sqE7!uwJ22IRxAASfybHWupjoyXq7dWAQE9YNGTBs0W7z6t4sdCBM0JoYpN3OcMha8WD8U0he0RrRkumNJAQ6a3w$
> >
> >
> > JIRA link -
> https://urldefense.com/v3/__https://issues.apache.org/jira/browse/KAFKA-21482__;!!Ayb5sqE7!uwJ22IRxAASfybHWupjoyXq7dWAQE9YNGTBs0W7z6t4sdCBM0JoYpN3OcMha8WD8U0he0RrRkumNSpKK6gk$
> >
> >
> > Thanks,
> > Abhinav
> >
>

Reply via email to