Thanks for the explanation LGTM Thanks, Penghui
On Tue, Aug 23, 2022 at 11:41 PM Dave Fisher <w...@apache.org> wrote: > While not a comment about this proposal I have a comment another split > bundle concept. > > Manually split out a topic into its own bundle. > > Say a bundle is 0x00000000 to 0x00000200 with 5 topics. > > t1 at 0x00000010 > t2 at 0x00000030 > t3 at 0x00000110 > t4 at 0x00000123 > t5 at 0x000001AE > > Let’s split out t3 to it’s own bundle and then have three new bundles: > > 0x00000000:0x0000010F - t1, t2 > 0x00000110:0x00000110 - t3 > 0x00000111:0x00000200 - t4, t5 > > This can be useful when t3 has most of the traffic > > > On Aug 23, 2022, at 2:13 AM, lordcheng10 <1572139...@qq.com.INVALID> > wrote: > > > > "It's a good idea to improve the bundle split for the case that the > traffic > > of the topic doesn't change drastically > > Otherwise, we should not use this policy. or can we use it for all > cases?" > > > > > > 1.It is suitable for scenarios where topic traffic is relatively stable. > In addition, we can also adjust the flowOrQpsDifferenceThresholdPercentage > configuration to adapt to traffic fluctuations; > > 2.The current strategy is to split bundles based on the configuration > loadBalancerNamespaceBundleMaxMsgRate or > loadBalancerNamespaceBundleMaxBandwidthMbytes; > > 3.If the qps or traffic of a bundle is less than > loadBalancerNamespaceBundleMaxMsgRate*(1+ > flowOrQpsDifferenceThresholdPercentage) or > loadBalancerNamespaceBundleMaxBandwidthMbytes*(1+flowOrQpsDifferenceThresholdPercentage), > the policy will no longer trigger split; > > > > > > "do we need to consider the consumer rate > > the `flow or qps` is based on the entries or messages?" > > > > > > 1. The consumer rate has been considered; > > 2. The strategy has been based on the entries and throughput; > > > > > > > > > > Thanks, > > lordcheng10 > > > > > > Reply for Penghui > >
