Hi Colt,
Thanks for the feedback.
> most deployments have three racks, RF=3, and one replica for
each partition in each rack
This KIP is mainly targeting the case where the client won't always be in
the same rack as any replica. There's some proposal to make RF=2 and use
other tiered storage to
Hi all,
I've got a rather naive question here. The spiritual goal of this KIP is to
reduce cross-rack traffic (normally, this manifests itself in terms of a
higher AWS/Azure bill as cloud providers charge for cross-AZ traffic).
To generalize, most deployments have three racks, RF=3, and one repli
Hi Bruno,
dropping config rack.aware.assignment.enabled
and add value NONE to the enum for the possible values of config
rack.aware.assignment.strategy sounds good to me.
On Thu, Jun 1, 2023 at 12:39 AM Bruno Cadonna wrote:
> Hi Hao,
>
> Thanks for the updates!
>
> What do you think about dropp
Hi Hao,
Thanks for the updates!
What do you think about dropping config rack.aware.assignment.enabled
and add value NONE to the enum for the possible values of config
rack.aware.assignment.strategy?
Best,
Bruno
On 31.05.23 23:31, Hao Li wrote:
Hi all,
I've updated the KIP based on the fee
Thanks Sophie! I've just made the updates :)
On Wed, May 31, 2023 at 2:29 PM Sophie Blee-Goldman
wrote:
> Thanks Hao -- I really like Walker's idea to separate the assignment
> strategy from the rack awareness option.
> Having two different configs for these makes a lot of sense to me and seems
Hi all,
I've updated the KIP based on the feedback. Major changes I made:
1. Add rack aware assignment to `StickyTaskAssignor`
2. Reject `Prefer reliability and then find optimal cost` option in standby
task assignment.
On Wed, May 31, 2023 at 12:09 PM Hao Li wrote:
> Hi all,
>
> Thanks for th
Thanks Hao -- I really like Walker's idea to separate the assignment
strategy from the rack awareness option.
Having two different configs for these makes a lot of sense to me and seems
like it will benefit the most users.
Given your KIP already has a lot going on with it, I would be happy to
writ
Hi all,
Thanks for the feedback! I will update the KIP accordingly.
*For Sophie's comments:*
1 and 2. Good catch. Fixed these.
3 and 4 Yes. We can make this public config and call out the clientConsumer
config users need to set.
5. It's ideal to take the previous assignment in HAAssignor into
Hi Hao,
Most of the comments I had on this kip are already mentioned, but I did
want to share my two major concerns.
1. Stability. I worry about stability. If we only have the HA assignor work
with rack awareness we will have a lot of state movement in many cases.
Sophie and Bruno have this conce
Hi Hao,
Thank you for the KIP! Really interesting!
In general, I think the KIP is a bit too vague. You explain the main
algorithm and different options. It is not clear to me on what option we
will start voting. One way out of this situation would be to cut the KIP
down to the simplest optio
Hey Hao, thanks for the KIP!
1. There's a typo in the "internal.rack.aware.assignment.strategry" config,
this
should be internal.rack.aware.assignment.strategy.
2.
> For O(E^2 * (CU)) complexity, C and U can be viewed as constant. Number of
> edges E is T * N where T is the number of clients an
Thanks John! Yeah. The ConvergenceTest looks very helpful. I will add it to
the test plan. I will also add tests to verify the new optimizer will
produce a balanced assignment which has no worse cross AZ cost than the
existing assignor.
Hao
On Mon, May 22, 2023 at 3:39 PM John Roesler wrote:
>
Hi Hao,
Thanks for the KIP!
Overall, I think this is a great idea. I always wanted to circle back after the
Smooth Scaling KIP to put a proper optimization algorithm into place. I think
this has the promise to really improve the quality of the balanced assignments
we produce.
Thanks for provi
Hi Colt,
Thanks for the comments.
> and I struggle to see how the algorithm isn't at least O(N) where N is
the number of Tasks...?
For O(E^2 * (CU)) complexity, C and U can be viewed as constant. Number of
edges E is T * N where T is the number of clients and N is the number of
Tasks. This is be
Hello Hao,
First of all, THANK YOU for putting this together. I had been hoping
someone might bring something like this forward. A few comments:
**1: Runtime Complexity
> Klein’s cycle canceling algorithm can solve the min-cost flow problem in
O(E^2CU) time where C is max cost and U is max capaci
Hi all,
I have submitted KIP-925 to add rack awareness logic in task assignment in
Kafka Streams and would like to start a discussion:
https://cwiki.apache.org/confluence/display/KAFKA/KIP-925%3A+Rack+aware+task+assignment+in+Kafka+Streams
--
Thanks,
Hao
16 matches
Mail list logo