Throttle replicas based on client-id

[Archie <an...@gmail.com>]

Hi All,
            I am pretty new to kafka development and I am trying to limit
the amount of traffic that a client can send/produce to the kafka cluster.

             Using quotas on client-id, I was able to limit the amount of
traffic that can be produced by a client to a value very close to the
quota. Now I also want to make sure that the replicated traffic also
follows the same limit.

             For example, I have 2 clients: client-1 and client-2, and 3
machines: mac-1, mac-2 and mac-3. Both the clients have a master replica at
mac-1 and both have one slave replica each in mac-2 and mac-3 (Replication
factor = 3). I have set quota of client-1 as 50 MBps and client-2 as 40
MBps. The client-quotas were able to ensure that data is written into
master at the specified rate ( approximately 50 and 40 MBps respectively).
I tested it by running producers of both clients simultaneously, setting
ack as 1 (means ack when master receives) and using the script
bin/kafka-producer-perf-test.sh.

              But when I ran the same script with ack as -1 (ack from all
replicas), both my client's performance drops down to 30 MBps. My first
question is shouldn't kafka still allow producers to get throughput of 50
and 40 MBps respectively even in presence of replicas?

             I then thought may be this is happening because the replica
traffic does not have the same quota. Now I've been exploring the kafka
source code and found out that there is already something called
replication quota but I believe it does throttling only during partition
movement.

             I've looked at ReplicationQuotaManager.scala ,
ClientQuotaManager.scala files but I am not sure how to implement this
(throttle replica traffic traffic based on client-id). Does anyone have any
ideas/suggestions on how this can be done in the existing kafka code?




Thanks,
Archie