Hello Rohan, Thanks for the KIP! Overall it looks very nice. Just some quick thoughts :
1. All the API logic is granular at the Task level, except the previousOwnerForPartition func. I’m not clear what’s the motivation behind it, does our controller also want to change how the partitions->tasks mapping is formed? 2. Just on the API layering itself: it feels a bit weird to have the three built-in functions (defaultStandbyTaskAssignment etc) sitting in the ApplicationMetadata class. If we consider them as some default util functions, how about introducing moving those into their own static util methods to separate from the ApplicationMetadata “fact objects” ? 3. I personally prefer `NodeAssignment` to be a read-only object containing the decisions made by the assignor, including the requestFollowupRebalance flag. For manipulating the half-baked results inside the assignor itself, maybe we can just be flexible to let users use whatever struts / their own classes even, if they like. WDYT? Thanks, Guozhang On Tue, Nov 7, 2023 at 12:04 PM Rohan Desai <desai.p.ro...@gmail.com> wrote: > > Hello All, > > I'd like to start a discussion on KIP-924 ( > https://cwiki.apache.org/confluence/display/KAFKA/KIP-924%3A+customizable+task+assignment+for+Streams) > which proposes an interface to allow users to plug into the streams > partition assignor. The motivation section in the KIP goes into some more > detail on why we think this is a useful addition. Thanks in advance for > your feedback! > > Best Regards, > > Rohan
