Hi all,

Thank you for your comments.

chia_00:
I wanted to keep the changes to the scope of the KIP, but sure we can close 
this gap.

chia_01:
We have the total bytes bytes-consumed-rate and are adding 
bytes-consumed-in-rack-rate. So technically one can do `bytes-consumed-rate - 
bytes-consumed-in-rack-rate` to get what goes outside of the rack. Or do you 
mean to explicitly mark the traffic that the operator wanted to be in rack, but 
the client couldn't figure out the rack?

svv1:
It can be done, but we should be aware that it's asymmetrical: Produce request 
don't carry the rack information now.

svv2:
So basically replace "byte-in-rack-rate" with "byte-cross-rack-rate" and so on? 
If we scope it to "traffic that goes to a _known_ rack, different from ours", 
IMO the usefulness will be quite limited, because racks may be just not 
configured. To make it work, we need to give it the meaning of "traffic that 
goes to a _known_ rack, different from ours, or to an unknown rack".

Best,
Ivan


On Wed, Jul 15, 2026, at 17:01, Viktor Somogyi-Vass via dev wrote:
> Hi Ivan,
> 
> My comments:
> 
> svv1. Should we add a BytesOutRackAlignedPerSec that is similar to
> BytesOutPerSec but counts only the rack-aligned traffic? Since we have the
> rack IDs in fetch requests it may be easy to calculate it and might be
> higher value if client metrics aren't available.
> 
> svv2. Operators I guess would mostly care about cross-rack traffic for cost
> reasons. Have you considered computing directly that instead or
> additionally? Also "CrossRack" would also resolve the naming complication
> that was proposed in AS1.
> 
> Best,
> Viktor
> 
> On Fri, Jul 3, 2026 at 11:19 PM Chia-Ping Tsai <[email protected]> wrote:
> 
> > hi Ivan
> >
> > chia_00: Should we also add
> > "kafka.producer:type=producer-metrics,client-id=<id>" for the producer? The
> > proposed consumer metrics have both levels (client vs. client+topic), but
> > the producer currently only has the topic level.
> >
> > chia_01: Should we provide "unknown-rack-bytes" for the misconfiguration?
> >
> > Best,
> > Chia-Ping
> >
> > On 2026/06/22 11:13:50 Ivan Yurchenko wrote:
> > > Hi team,
> > >
> > > I'd like to start the discussion for a new KIP-1355: In-rack traffic
> > observability for clients and brokers [1] The KIP is small, just to add
> > some additional metrics. Thanks to Chia-Ping Tsai for the idea, proposed as
> > a follow-up to KIP-1123: Rack-aware partitioning for Kafka Producer [2].
> > >
> > > Best regard,
> > > Ivan Yurchenko
> > >
> > > [1]
> > https://cwiki.apache.org/confluence/display/KAFKA/KIP-1355%3A+In-rack+traffic+observability+for+clients+and+brokers
> > > [2]
> > https://cwiki.apache.org/confluence/display/KAFKA/KIP-1123:+Rack-aware+partitioning+for+Kafka+Producer
> >
> 

Reply via email to