Thank you Peter, That is a good point. We need to add extensive documentation to the configuration with examples. YUNIKORN-1 has been changed into an epic, let's add jiras around new documentation to that jira and make sure we cover that.
Wilfred On Wed, 7 Dec 2022 at 22:53, Peter Bacsko <[email protected]> wrote: > Hi Wilfred, > > I added some comments to the document. I'll re-read it later after you > respond. > > I like the document very much, it's clear and concise. What I think we need > (maybe not in this document, but in the user guide) is realistic examples > with numbers/tables/diagrams to make it easier to comprehend. I had to > re-read some paragraphs to make sure I truly graps the concepts of fencing > and offsetting. I think I get it, but it would be much easier with a use > case and explains step-by-step how YK will sort queues/apps with a certain > setting. > > Peter > > On Mon, Dec 5, 2022 at 7:58 AM Wilfred Spiegelenburg <[email protected]> > wrote: > > > Weiwei, > > > > Thanks for the feedback. > > > > The dynamic priority is based on the fact that every pod on K8s has a > > priority. K8s will use that priority to decide on node evictions and > > scheduling preemptions. If no priority is set the default priority will > be > > set by K8s (default == 0, can be changed cluster wide). > > > > If you want all pods in the application to behave the same then they > should > > have the same priority. If you do not do that the scheduler, or the node, > > could decide that a pod of what you think belongs to a high priority > > application gets preempted or evicted as it has a real much lower > priority. > > I would consider that a far bigger issue. You cannot say a spark driver > has > > a higher priority than the executors. Preventing evictions and pushing > > drivers to the back of the list in preemptions > > > > Mixing priority adherence is dangerous. Example: an app's priority is 0, > > the pod's priority is 1000. The app gets sorted in the back of the queue. > > The high priority pod part of the low priority app will not get looked > at. > > That is more confusing then changing the app priority to 1000 until that > > high priority pod is scheduled and then lower the app priority again. > Like > > state changes we can track the dynamic priority change and expose it. > > > > I have dropped the same comment in the document. > > > > WIlfred > > > > On Mon, 5 Dec 2022 at 15:36, Weiwei Yang <[email protected]> wrote: > > > > > Hi Wilfred > > > > > > Thanks for putting things together, this looks really good. > > > I really like the concepts introduced in the proposal, such as priority > > > fencing and offset brilliant ideas! > > > I have one major doubt here, which is more related to how users can > > consume > > > this feature. I have added some comments in the doc, pls check. > > > The key idea is I want to make sure this is something users can easily > > > understand and use these concepts to achieve what they cannot do in the > > > past. And at the same time, they won't get confused by lots of > > > implications. > > > > > > Thanks > > > Weiwei > > > > > > On Thu, Dec 1, 2022 at 5:42 PM Wilfred Spiegelenburg < > > [email protected]> > > > wrote: > > > > > > > Hi, > > > > > > > > Priorities are a standard feature in K8s. YuniKorn has only had a > very > > > > limited use of priorities in Daemon Set preemption and request > sorting. > > > > Extending this into the rest of the scheduling cycle has been on the > > > cards > > > > for a long time. > > > > > > > > This design does not change the sorting of requests, i.e. > > AllocationAsk, > > > > within an application. Priority based sorting has been used since our > > > > earliest releases. Adding a different sorting for requests falls > > outside > > > of > > > > the design. > > > > > > > > Implementation will be tracked under YUNIKORN-1 > > > > > > > > Please provide feedback on the design > > > > > > > > Wilfred > > > > > > > > [1] > > > > > > > > > > > > > > https://docs.google.com/document/d/1P_xlWSj65BiwAglvDvFyrnh8wZ7M72FRtIg199fyHNo/edit?usp=sharing > > > > > > > > > >
