zhilong

@jjyao I think this PR can help to reduce the object memory store usage by dispatching the task with large args first. But I just want to check if the...

> Hi @Bye-legumes, > > You are absolutely right that Ray is currently dumb on selecting which task to run first so it can achieve global optimal outcome. When we...

@jjyao I think another way is to loop the tasks in the queue based on the rate. Here I use `P1,P2` for the Producer tasks and `C1, C2` for the...

> Another direction is that we can implement priorities for tasks and actors so that applications can set higher priority for consumer than producer. If we implemented in this way,...

> Hmm this actually looks like it may be a regression. Last I remembered, Ray's LocalTaskManager should have a fairness policy based on each task's scheduling class. So if there...

> > > Hmm this actually looks like it may be a regression. Last I remembered, Ray's LocalTaskManager should have a fairness policy based on each task's scheduling class. So...

Sure, the starvationis the issue that may be created, but it can solve the memory accumulation..As compared with memory accumulation, starvation is more acceptable to keep the node memory safe.

> Hmm I'm having trouble finding the code reference now. Maybe it did not exist before...Will try to look into it more when I get some time. > > In...

Hi, @stephanie-wang I just have a prototype of fairness policy dispatching that insert between the shcheduling class loop and worker loop. That is we skip that scheduling class when that...

> > Hi, @stephanie-wang I just have a prototype of fairness policy dispatching that insert between the shcheduling class loop and worker loop. That is we skip that scheduling class...