Hi Karthick, We’ve experienced the similar issue before. What we were doing at that time was to define multiple topics and each has a different # of partitions which means some of the topics with more partitions will have the high parallelisms for processing. And you can further divide the topics into several groups and each group should have the similar # of partitions. For each group, you can define as the source of flink data stream to run them in parallel with different parallelism.
------------------ Original ------------------ From: Giannis Polyzos <ipolyzos...@gmail.com> Date: Sat,Sep 16,2023 11:52 PM To: Karthick <ibmkarthickma...@gmail.com> Cc: Gowtham S <gowtham.co....@gmail.com>, user <user@flink.apache.org> Subject: Re: Urgent: Mitigating Slow Consumer Impact and Seeking Open-SourceSolutions in Apache Kafka Consumers Can you provide some more context on what your Flink job will be doing? There might be some things you can do to fix the data skew on the link side, but first, you want to start with Kafka.For starters, you need to better size and estimate the required number of partitions you will need on the Kafka side in order to process 1000+ messages/second. The number of partitions should also define the maximum parallelism for the Flink job reading for Kafka. If you know your "hot devices" in advance you might wanna use a custom partitioner that spreads those devices to somewhat separate partitions. Overall this is somewhat of a trial-and-error process. You might also want to check that these partitions are evenly balanced among your brokers and don't cause too much stress on particular brokers. Best On Sat, Sep 16, 2023 at 6:03 PM Karthick <ibmkarthickma...@gmail.com> wrote: Hi Gowtham i agree with you, I'm eager to resolve the issue or gain a better understanding. Your assistance would be greatly appreciated. If there are any additional details or context needed to address my query effectively, please let me know, and I'll be happy to provide them. Thank you in advance for your time and consideration. I look forward to hearing from you and benefiting from your expertise. Thanks and Regards Karthick. On Sat, Sep 16, 2023 at 11:04 AM Gowtham S <gowtham.co....@gmail.com> wrote: Hi Karthik This appears to be a common challenge related to a slow-consuming situation. Those with relevant experience in addressing such matters should be capable of providing assistance. Thanks and regards,Gowtham S On Fri, 15 Sept 2023 at 23:06, Giannis Polyzos <ipolyzos...@gmail.com> wrote: Hi Karthick, on a high level seems like a data skew issue and some partitions have way more data than others? What is the number of your devices? how many messages are you processing? Most of the things you share above sound like you are looking for suggestions around load distribution for Kafka. i.e number of partitions, how to distribute your device data etc. It would be good to also share what your flink job is doing as I don't see anything mentioned around that.. are you observing back pressure in the Flink UI? Best On Fri, Sep 15, 2023 at 3:46 PM Karthick <ibmkarthickma...@gmail.com> wrote: Dear Apache Flink Community, I am writing to urgently address a critical challenge we've encountered in our IoT platform that relies on Apache Kafka and real-time data processing. We believe this issue is of paramount importance and may have broad implications for the community. In our IoT ecosystem, we receive data streams from numerous devices, each uniquely identified. To maintain data integrity and ordering, we've meticulously configured a Kafka topic with ten partitions, ensuring that each device's data is directed to its respective partition based on its unique identifier. This architectural choice has proven effective in maintaining data order, but it has also unveiled a significant problem: One device's data processing slowness is interfering with other devices' data, causing a detrimental ripple effect throughout our system. To put it simply, when a single device experiences processing delays, it acts as a bottleneck within the Kafka partition, leading to delays in processing data from other devices sharing the same partition. This issue undermines the efficiency and scalability of our entire data processing pipeline. Additionally, I would like to highlight that we are currently using the default partitioner for choosing the partition of each device's data. If there are alternative partitioning strategies that can help alleviate this problem, we are eager to explore them. We are in dire need of a high-scalability solution that not only ensures each device's data processing is independent but also prevents any interference or collisions between devices' data streams. Our primary objectives are: 1. Isolation and Independence: We require a strategy that guarantees one device's processing speed does not affect other devices in the same Kafka partition. In other words, we need a solution that ensures the independent processing of each device's data. 2. Open-Source Implementation: We are actively seeking pointers to open-source implementations or references to working solutions that address this specific challenge within the Apache ecosystem or any existing projects, libraries, or community-contributed solutions that align with our requirements would be immensely valuable. We recognize that many Apache Flink users face similar issues and may have already found innovative ways to tackle them. We implore you to share your knowledge and experiences on this matter. Specifically, we are interested in: - Strategies or architectural patterns that ensure independent processing of device data. - Insights into load balancing, scalability, and efficient data processing across Kafka partitions. - Any existing open-source projects or implementations that address similar challenges. We are confident that your contributions will not only help us resolve this critical issue but also assist the broader Apache Flink community facing similar obstacles. Please respond to this thread with your expertise, solutions, or any relevant resources. Your support will be invaluable to our team and the entire Apache Flink community. Thank you for your prompt attention to this matter. Thanks & Regards Karthick.
