Hello PoAn, Thanks for the feedback
poan_00: In range mode, keys are generated by `recordIndex % keyRange`, which is fully deterministic and not affected by `--random-seed`. The seed only controls the PRNG used for random payload generation in that case. The example is misleading, I will remove it. poan_01: According to the JDK documentation, `SplittableRandom` generates uniformly distributed pseudorandom values. With a sufficiently large number of records, each key in random mode appears roughly the same number of times, so the partition distribution s tatistically converges toward behavior similar to range mode. The main difference is that random mode introduces short-term burstiness, where the same key may appear consecutively for a period of time, while range mode produces a perfectly even round-robin pattern. However, neither mode inherently creates a truly skewed (hot-partition) distribution. I’ll update the motivation section to remove the hot-partition claim for random mode. Best Regards, Jiunn-Yang > PoAn Yang <[email protected]> 於 2026年5月26日 晚上7:18 寫道： > > Hi Jiunn, > > Thanks for the KIP. > > poan_00: In example usage, there is a case use --key-distribution range with > --random-seed. > In this case, does the --random-seed parameter take effect? If not, can we > remove it? > > poan_01: In motivation, one use case of random distribution is hot-partition > scenario. > However, in JDK document, the SplittableRandom is a generator of uniform > pseudorandom values [0]. > If hot-partition scenario is just because small key range, can we do it with > range key distribution directly? > > https://docs.oracle.com/en/java/javase/17/docs/api/java.base/java/util/SplittableRandom.html > > Best, > PoAn > >> On May 20, 2026, at 8:56 PM, 黃竣陽 <[email protected]> wrote: >> >> Hi chia, >> >> Thanks for the feedback, >> >> chia_00: I have added a new optional argument --random-seed <SEED> (default: >> 0) >> to let users set the seed manually. The default value of 0 ensures >> deterministic, reproducible >> benchmark runs by default. >> >> chia_01: I have updated the Motivation section in the KIP to elaborate on >> the practical >> use cases for each key distribution mode. >> >> Best Regards, >> Jiunn-Yang >> >>> Chia-Ping Tsai <[email protected]> 於 2026年5月20日 上午11:48 寫道： >>> >>> hi Jiunn >>> >>> thanks for this KIP! >>> >>> chia_00: Regarding the random seed, what are your thoughts on its >>> initialization? >>> >>> chia_01: Could you elaborate on the practical use cases for each key >>> distribution mode in the Motivation section? >>> >>> Best,Chia-Ping >>> >>> On 2026/03/30 13:06:05 黃竣陽 wrote: >>>> Hello everyone, >>>> >>>> I would like to start a discussion on KIP-1299 Use key range in >>>> ProducerPerformance >>>> <https://cwiki.apache.org/confluence/x/XpQ8G> >>>> >>>> This proposal aims to add configurable key distribution support to >>>> kafka-producer-perf-test. >>>> Currently, the tool always produces records with null keys, which does not >>>> reflect real-world >>>> keyed workloads. This KIP introduces two new arguments — >>>> --key-distribution and --message-key-range >>>> — enabling engineers to benchmark with round-robin or random key >>>> strategies over a bounded >>>> key space, providing more realistic performance measurements. >>>> >>>> Best regards, >>>> Jiunn-Yang >> >
