Hi Everyone,

I have been reviewing the *KIP-1150* proposal and the reference
implementation provided in the aiven/inkless repository
<https://github.com/aiven/inkless/issues>. While the cost-saving goals are
compelling, I have identified three critical architectural flaws in the
current design that directly undermine the KIP's feasibility and
correctness.

1.* Unbounded Storage Leaks (Orphaned Segments) *The current
"fire-and-forget" upload logic creates a race condition where a broker
crash after S3 upload (but before metadata commit) results in orphaned
objects that are never garbage collected. In high-churn environments, this
negates the proposed cost benefits and creates an unbounded storage leak.
Issue1 <https://github.com/aiven/inkless/issues/494>:

2. EOS Violations in Leaderless Architecture The EndTxnRequest handling
does not strictly synchronize with the Batch Coordinator before advancing
the local LSO. This creates a split-brain scenario where consumers can see
an inconsistent transaction state, effectively breaking Exactly-Once
Semantics (EOS). Issue2 <https://github.com/aiven/inkless/issues/495>:

3. Critical Testing Gaps The integration tests rely exclusively on MinIO
(strong consistency) and do not simulate S3's eventual consistency or 503
throttling. The system is currently unverified against real-world cloud
failure modes, posing a significant risk to production stability. Issue3
<https://github.com/aiven/inkless/issues/496>:

Given that these issues affect data integrity and the core economic promise
of the KIP, I strongly suggest we pause the vote until a concrete design
for "Distributed Garbage Collection" and "Coordinator-Synchronized
Transactions" is explicitly added to the proposal.

Regards,
Viquar Khan
Sr Data Architect
https://www.linkedin.com/in/vaquar-khan-b695577/

On Wed, 14 Jan 2026 at 17:04, Ghai, Varun via dev <[email protected]>
wrote:

> +1 (non-binding)
>
> We (NetApp) have been closely following the changes in this KIP, its
> sub-KIPs and the discussions around diskless in the community. We're happy
> to see support coalescing around a single path forward to bring diskless to
> Kafka. In its current state, we support this KIP.
>
> On 2026/01/13 20:29:14 Stanislav Kozlovski wrote:
> > +1 (binding)
> >
> > Thank you so much for this effort. Seeing you reach consensus with the
> competing KIP-1176 is very promising as well. I'm excited to work with you
> on closing design for the rest of the KIPs and ultimately merging this. I
> agree wholeheartedly on the motivational aspect - this is a necessary
> feature if we wish for Kafka to remain competitive in the industry.
> >
> > On 2026/01/12 12:33:37 Andrew Mills wrote:
> > > +1 (non-binding)
> > >
> > > As a long time Kafka use and member of a team who provides management
> for
> > > Apache Kafka, I'm excited to see this KIP move forward. It will help to
> > > keep Kafka cost competitive in the face of many cloud-first
> alternatives.
> > >
> > > -- Andrew
> > >
> > > On Fri, Jan 9, 2026, 7:49 PM Henry Haiying Cai via dev <
> [email protected]>
> > > wrote:
> > >
> > > > +1 (non-binding)
> > > >
> > > > As the author of competing KIP (KIP-1176: Tiered Storage for Active
> Log
> > > > Segment), I am supporting KIP-1150 as KIP-1150 represents the right
> > > > long-term architecture for Kafka in the cloud.
> > > >
> > > > Henry Cai
> > > >
> > > >
> > > > On Friday, January 9, 2026 at 04:08:03 AM PST, Anatolii Popov via
> dev <
> > > > [email protected]> wrote:
> > > >
> > > >
> > > >
> > > >
> > > >
> > > > Hi all,
> > > >
> > > > We are restarting the vote for KIP-1150: Diskless Topics.
> > > >
> > > > Since the previous vote, the KIP has been substantially updated to
> address
> > > > community feedback.
> > > >
> > > > As a reminder, KIP-1150 is a motivational KIP intended to establish
> the
> > > > architectural direction for diskless topics. The specific
> implementation
> > > > details are delegated to separate sub-KIPs, which will be discussed
> and
> > > > voted on individually.
> > > >
> > > > To give you a concrete idea of the path forward, KIP-1163 is already
> in an
> > > > almost-ready-for-review state, so you can see where we are heading
> after
> > > > this.
> > > >
> > > > KIP-1150:
> > > >
> > > >
> https://cwiki.apache.org/confluence/display/KAFKA/KIP-1150%3A+Diskless+Topics
> > > > KIP-1163
> > > > <
> https://cwiki.apache.org/confluence/display/KAFKA/KIP-1150%3A+Diskless+TopicsKIP-1163
> >
> > > > :
> > > >
> https://cwiki.apache.org/confluence/display/KAFKA/KIP-1163%3A+Diskless+Core
> > > >
> > > > Best,
> > > > Anatolii
> > > > --
> > > > Anatolii Popov
> > > > Senior Software Developer, *Aiven OY*
> > > > m: +358505126242
> > > > w: aiven.io  e: [email protected]
> > > > <https://www.facebook.com/aivencloud>
> > > > <https://www.linkedin.com/company/aiven/>  <
> https://twitter.com/aiven_io>
> > > >
> > >
> >



-- 
Regards,
Vaquar Khan

Reply via email to