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 concern as well. 2. It seems the rack awareness assignment operation can be run after any assignment algorithm. I would think that maybe we can leave it agnostic if it is using the sicky assignor or the HA assignor and let the users choose the strategy. Maybe just have the rack awareness be off or on, independent of the assignment strategy. Walker On Wed, May 31, 2023 at 7:46 AM Bruno Cadonna <cado...@apache.org> wrote: > 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 options and evaluate those. Then, we would have a > starting point from which we can move on. > > > 1. > Mention that the KIP optimizes only the read path and does nothing about > the write path. > > > 2. > "U is 1 and C is at most 3 (A task can have at most 3 topics including > changelog topic?)" > where C is max cost and U is max capacity > > C is not at most 3 to answer the question in the KIP. A task can have > any number of topics it reads from. Some examples: > - a processor API operator with multiple state stores reads from > multiple changelog topics > - a cascade of merge operators would result in a task with multiple > input topics > - a cascade of joins would result in a task with multiple input topics > and multiple changelog topics. > I am pretty sure there are also other examples. > > Assuming cost C is 1 for each topic partition is a simplification. > Traffic for tasks can vary significantly. I saw joins that had 100s of > bytes/s on one side and 10s of MB/s on the other side. I guess the > cross-rack traffic depends on the data rate. Please correct me if I am > wrong. I am fine with simplifying but then we also need to explicitly > state the simplification and its limitation in the KIP to manage > expectations. > > C is just a factor in the complexity but we should be clear about the > simplification we made. How much C influences the actual performance we > do not know and we should evaluate this as part of the implementation of > the KIP. Maybe add this aspect to the performance experiments in the > test plan section. > > > 3. > I second Sophie's question about the complexity being O(T^2 * N^2) > instead of O(T*N). > > > 4. > The improved algorithm in "Min cost with balanced sub-topology" contains > a bunch more edges and the complexity of the algorithm depends on the > square of the number of edges. Can you say something about the trade-off > or even quantify it? How does does the complexity change from O(T^2 * N^2)? > > > 5. > If you propose to implement multiple algorithms, the KIP should add > public configs as Sophie proposed. > > > 6. > Does any of the algorithm change the subscription protocol? Usually we > describe those changes in a KIP. > > > 7. > I have a couple of minor comment about notation: > > 7.1 For the complexity, I think it would be better to either use |T| and > |C| or define new variables like for example N_task = |T| and N_client = > |C| for the formula to be consistent with the mapping function you > define in the previous section. > > 7.2 Using C for the set of clients and the cost is confusing. Maybe use > Cost or $ for cost. > > > 8. > In the "Graph construction" section in the "Min cost with balanced > sub-topology" section, you write > "Create new set of nodes which has same number as clients" > Shouldn't this be number of tasks. > > > 9. > Regarding standby assignment, have you considered to simplify the setup > by defining if rack-aware configs are set, the standby assignment is > optimized for reliability and if they are not set costs are optimized. I > think that would be a good starting point on which we can iterate in > future. > > > 10. > Just a clarification and something that you should corrected in the KIP. > In "Assignment of stateless tasks" you contrast stateless tasks with > active task. However, active tasks can be stateless or stateful. So > "Rack awareness assignment algorithm for active tasks" should actually > be "Rack awareness assignment algorithm for active stateful tasks". > Please use the terms accordingly otherwise, it gets confusing. > > > 11. > In the testing plan, I think it would be useful to also have performance > experiments along other dimension like number of topic partitions a task > reads from, i.e., basically the varying costs per task. > > > Best, > Bruno > > On 30.05.23 23:28, Sophie Blee-Goldman wrote: > > 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 and N is the number of > >> Tasks. This is because a task can be assigned to any client so there > will > >> be an edge between every task and every client. The total complexity > would > >> be O(T * N) if we want to be more specific. > > > > I feel like I'm missing something here, but if E = T * N and the > complexity > > is ~O(E^2), doesn't > > this make the total complexity order of O(T^2 * N^2)? > > > > 3. > > > >> Since 3.C.I and 3.C.II have different tradeoffs and work better in > >> different workloads etc, we > > > > could add an internal configuration to choose one of them at runtime. > >> > > Why only an internal configuration? Same goes for > > internal.rack.aware.assignment.standby.strategry (which also has the > typo) > > > > 4. > > > >> There are no changes in public interfaces. > > > > I think it would be good to explicitly call out that users can utilize > this > > new feature by setting the > > ConsumerConfig's CLIENT_RACK_CONFIG, possibly with a brief example > > > > 5. > > > >> The idea is that if we always try to make it overlap as much with > >> HAAssignor’s target > > > > assignment, at least there’s a higher chance that tasks won’t be > shuffled a > >> lot if the clients > > > > remain the same across rebalances. > >> > > This line definitely gave me some pause -- if there was one major > takeaway > > I had after KIP-441, > > one thing that most limited the feature's success, it was our assumption > > that clients are relatively > > stable across rebalances. This was mostly true at limited scale or for > > on-prem setups, but > > unsurprisingly broke down in cloud environments or larger clusters. Not > > only do clients naturally > > fall in and out of the group, autoscaling is becoming more and more of a > > thing. > > > > Lastly, and this is more easily solved but still worth calling out, an > > assignment is only deterministic > > as long as the client.id is persisted. Currently in Streams, we only > write > > the process UUID to the > > state directory if there is one, ie if at least one persistent stateful > > task exists in the topology. This > > made sense in the context of KIP-441, which targeted heavily stateful > > deployments, but this KIP > > presumably intends to target more than just the persistent & stateful > > subset of applications. To > > make matters even worse, "persistent" is defined in a semantically > > inconsistent way throughout > > Streams. > > > > All this is to say, it may sound more complicated to remember the > previous > > assignment, but (a) > > imo it only introduces a lot more complexity and shaky assumptions to > > continue down this > > path, and (b) we actually already do persist some amount of state, like > the > > process UUID, and > > (c) it seems like this is the perfect opportunity to finally rid > ourselves > > of the determinism constraint > > which has frankly caused more trouble and time lost in sum than it would > > have taken us to just > > write the HighAvailabilityTaskAssignor to consider the previous > assignment > > from the start in KIP-441 > > > > 6. > > > >> StickyTaskAssignor users who would like to use rack aware assignment > >> should upgrade their > > > > Kafka Streams version to the version in which > HighAvailabilityTaskAssignor > >> and rack awareness > > > > assignment are available. > > > > Building off of the above, the HAAssignor hasn't worked out perfectly for > > everybody up until now, > > given that we are only adding complexity to it now, on the flipside I > would > > hesitate to try and force > > everyone to use it if they want to upgrade. We added a "secret" backdoor > > internal config to allow > > users to set the task assignor back in KIP-441 for this reason. WDYT > about > > bumping this to a public > > config on the side in this KIP? > > > > > > On Tue, May 23, 2023 at 11:46 AM Hao Li <h...@confluent.io.invalid> > wrote: > > > >> 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 <vvcep...@apache.org> > 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 providing the details about the MaxCut/MinFlow algorithm. It > >>> seems like a good choice for me, assuming we choose the right scaling > >>> factors for the weights we add to the graph. Unfortunately, I don't > think > >>> that there's a good way to see how easy or hard this is going to be > until > >>> we actually implement it and test it. > >>> > >>> That leads to the only real piece of feedback I had on the KIP, which > is > >>> the testing portion. You mentioned system/integration/unit tests, but > >>> there's not too much information about what those tests will do. I'd > like > >>> to suggest that we invest in more simulation testing specifically, > >> similar > >>> to what we did in > >>> > >> > https://github.com/apache/kafka/blob/trunk/streams/src/test/java/org/apache/kafka/streams/processor/internals/assignment/TaskAssignorConvergenceTest.java > >>> . > >>> > >>> In fact, it seems like we _could_ write the simulation up front, and > then > >>> implement the algorithm in a dummy way and just see whether it passes > the > >>> simulations or not, before actually integrating it with Kafka Streams. > >>> > >>> Basically, I'd be +1 on this KIP today, but I'd feel confident about it > >> if > >>> we had a little more detail regarding how we are going to verify that > the > >>> new optimizer is actually going to produce more optimal plans than the > >>> existing assigner we have today. > >>> > >>> Thanks again! > >>> -John > >>> > >>> On 2023/05/22 16:49:22 Hao Li wrote: > >>>> 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 because a task can be assigned to any client so there > >> will > >>>> be an edge between every task and every client. The total complexity > >>> would > >>>> be O(T * N) if we want to be more specific. > >>>> > >>>>> But if the leaders for each partition are spread across multiple > >> zones, > >>>> how will you handle that? > >>>> > >>>> This is what the min-cost flow solution is trying to solve? i.e. Find > >> an > >>>> assignment of tasks to clients where across AZ traffic can be > >> minimized. > >>>> But there are some constraints to the solution and one of them is we > >> need > >>>> to balance task assignment first ( > >>>> > >>> > >> > https://cwiki.apache.org/confluence/display/KAFKA/KIP-925%3A+Rack+aware+task+assignment+in+Kafka+Streams#KIP925:RackawaretaskassignmentinKafkaStreams-Designforrackawareassignment > >>> ). > >>>> So in your example of three tasks' partitions being in the same AZ of > a > >>>> client, if there are other clients, we still want to balance the tasks > >> to > >>>> other clients even if putting all tasks to a single client can result > >> in > >>> 0 > >>>> cross AZ traffic. In > >>>> > >>> > >> > https://cwiki.apache.org/confluence/display/KAFKA/KIP-925%3A+Rack+aware+task+assignment+in+Kafka+Streams#KIP925:RackawaretaskassignmentinKafkaStreams-Algorithm > >>>> section, the algorithm will try to find a min-cost solution based on > >>>> balanced assignment instead of pure min-cost. > >>>> > >>>> Thanks, > >>>> Hao > >>>> > >>>> On Tue, May 9, 2023 at 5:55 PM Colt McNealy <c...@littlehorse.io> > >> wrote: > >>>> > >>>>> 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 capacity. In our > >>> particular > >>>>> case, C is 1 and U is at most 3 (A task can have at most 3 topics > >>> including > >>>>> changelog topic?). So the algorithm runs in O(E^2) time for our case. > >>>>> > >>>>> A Task can have multiple input topics, and also multiple state > >> stores, > >>> and > >>>>> multiple output topics. The most common case is three topics as you > >>>>> described, but this is not necessarily guaranteed. Also, math is one > >>> of my > >>>>> weak points, but to me O(E^2) is equivalent to O(1), and I struggle > >> to > >>> see > >>>>> how the algorithm isn't at least O(N) where N is the number of > >>> Tasks...? > >>>>> > >>>>> **2: Broker-Side Partition Assignments > >>>>> Consider the case with just three topics in a Task (one input, one > >>> output, > >>>>> one changelog). If all three partition leaders are in the same Rack > >> (or > >>>>> better yet, the same broker), then we could get massive savings by > >>>>> assigning the Task to that Rack/availability zone. But if the leaders > >>> for > >>>>> each partition are spread across multiple zones, how will you handle > >>> that? > >>>>> Is that outside the scope of this KIP, or is it worth introducing a > >>>>> kafka-streams-generate-rebalance-proposal.sh tool? > >>>>> > >>>>> Colt McNealy > >>>>> *Founder, LittleHorse.io* > >>>>> > >>>>> > >>>>> On Tue, May 9, 2023 at 4:03 PM Hao Li <h...@confluent.io.invalid> > >>> wrote: > >>>>> > >>>>>> 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 > >>>>>> > >>>>> > >>>> > >>>> > >>>> -- > >>>> Thanks, > >>>> Hao > >>>> > >>> > >> > >> > >> -- > >> Thanks, > >> Hao > >> > > >
